This is week seven of AI in Finance at NYU Stern (lecture slides here and last week here). We start with the iconic Diego Rivera Industry Murals. These depict the process of mass production and industrialization, based on Ford’s massive River Rouge complex, with workers and machines tangled together in a web of production. Rivera’s murals were painted in the depth of the Great Depression, and admit many possible interpretations: the wonder of modern industrial production, as well as Marxist notions of class conflict.

From the standpoint of this class, these murals reflect the challenges inherent in even “normal” technologies, a term we will get to. Industrial development was transformative, but did not diffuse instantaneously. Frictions and hurdles led to large lags in industrial growth, and much of the world still lags behind the dissemination of these technologies. Industry disrupted many workers, while creating new jobs and tasks.

The basic question for today is: will AI be the same? That is, will it be a “normal” technology? Or is it going to be something weirder? And where do the rents go depending on how normal it is?

The Case for Normal

Of course, this all hinges on what “normal” means in the first place. This is set out in a wonderful essay by Arvind Narayanan and Sayash Kapoor which makes the affirmative case for “AI as a Normal Technology” in one of the most influential articles about AI, ever.

By “normal” technology they mean a technology with transformative general purpose potential, but whose actual use and diffusion throughout the economy is rate-limited by a host of frictions, which both limit the downsides as well as upside potential. A normal technology, in this view, certainly has the ability to radically reshape the economy over time, as indeed did industrialization. But even AI, which appears unique in several aspects of rapid growth potential, is going to be limited (in their view) by a variety of real-world frictions which wind up making it “normal,” includinng:

1. The limits to diffusion set by human, organizational, regulatory, and institutional change

2. The limited value of conventional “benchmarks” in assessing real-world usefulness of AI tools to business tasks

3. The economic feedback loops in response to automation, which will divert human resources and talent towards newer problems

Note that this framework does not suggest risks are entirely absent: they discuss several, and indeed the regulatory frictions come exactly because human institutions respond to real risk through policy solutions. A key statistic in favor of their view is that, at least as of August 2024; 40% of adults had used AI, but people used it so infrequently it only made up 0.5-3.5% of work hours.

The basic argument against this perspective is that there are sufficiently weird or distinct aspects of AI which mean its influence is going to be decidedly non-Normal. This might happen, for instance, if we hit recursive self-improvement (RSI): if the models start building themselves and growth rates hit exponential. Alternatively, AI may look normal for some period of time, but after it is finally able to automate every last function that humans can do, we then fully automate the entire economy, and have no need for humans.

How People Use AI

For one guide into the “normality” of AI, we can look at how people tend to use it. A classic and early guide here is a McKinsey study which estimated $2.6-4.4 trillion in value. In Finance, they highlighted banking and insurance, customer service, and underwriting. McKinsey emphasizes bottlenecks of organizational frictions, risk management, and data infrastructure, not just raw model performance.

The Anthropic Economic Index also gives us one guide to adoption. One of their interesting findings is that the quality of the output you get from AI is a function of the inputs, which suggests there are real returns to skill in AI proficiency. You see that across countries, because high-income countries (Canada, Nordics, Singapore) tend to use AI in ways that are more augmenting or complementing in nature, while low-income countries use AI for coding and technical tasks. Augmented conversations in general have been rising, as product changes like skill files, project folders, and persistent memory push people towards more human-in-the-loop operations.

Another related result is that the correlation between the education level implied by the user’s prompt and that of the AI response is quite high; around a 0.93 correlation across countries or US states. What you prompt is what you get. This suggests that another bottleneck (and source of rents/profits) is likely to remain human intelligence, to the extent it continues to unlock higher quality AI output.

There are similar results with ChatGPT’s data as well. You see some signs of the bifurcation between Claude as an Enterprise workhorse, and ChatGPT becoming a dominant consumer player; with 10% of the world’s population using ChatGPT by mid-2025, and a rising share of the topics related to non-work content. For Finance applications, it seems like there is a lot of bottom-up incorporation of AI tools into workflows by analysts, PMs, and risk managers (as opposed, for example, to firms pushing top-down AI rules).

Restructuring Firms

This pattern of usage is suggestive that we are going to face really severe organizational frictions to the broader adoption of AI tools, and ultimately probably need to restructure organizations substantially to get the most out of the tools. A paper by Yotzov, Barrero, Bloom, Bunn, Davis, and coauthors gives us one firm-level perspective by surveying 6,000 executives across the world. AI usage at some level is pretty common (69% of firms actively using AI); but executives only use 1.5 hours/week. Firms don’t think AI has changed that much so far; but going forward they project +1.4% productivity, +0.8% output, and -0.7% employment. Their employees, meanwhile, think there will be a rise of +0.5% in employment, so someone is going to be wrong on the employment impacts.

Another survey from Baslandze, Edwards, Graham, and coauthors similarly documents rising interest; finance firms went from a 59% investment rate to 82% between 2025 and 2026. Some of the remaining points of friction include workforce training, privacy concerns, and insufficiently advanced AI technology. They find a productivity paradox: the perceived gains are larger than the realized ones, which they think reflects a lag in revenue realization (it could also be misperceptions of AI’s benefit for them). They also find interesting patterns of where AI is augmenting tasks compared to replacing them. AI enhances higher-order business functions (marketing, finance, and accounting) while replacing operational tasks (data entry, routine tasks, administrative), and so clerical staff are down while skilled technical roles are up.

One clue on the broader organizational shifts required comes from a paper on “Corporate Hierarchy” by Michael Ewens and Xavier Giroud. They construct organizational charts for over three thousand public firms using LinkedIn data (the network estimation technique alone is pretty interesting; hierarchical “layers” are estimated by looking at which roles people tend to come from and go to over their careers). The key finding is that firms have, on average, around ten layers of hierarchy; but AI adoption tends to lower the number of layers. It makes some intuitive sense if you think that the role of additional layers is to function as “problem solvers” to address complex problems faced by lower-layer employees, and AI can now handle some of those tasks. But suggests another wrenching economy-wide transition as firms are going to have to slowly figure out how to reconstruct their org chart assuming workers have access to AI.

We can also look to stock markets for some guide to what the market is pricing in. A paper by Andrea Eisfeldt, Gregor Schubert, and Ben Zhang generates a measure of firm workforce exposure to AI, and finds the market is basically pricing in some degree of worker substitution.

Task Chains, and Fixing the Slow Part

This brings us to another really influential paper on how frictions can hold back production with AI, and how we need to rethink our process. This is the task chaining paper by Demirer, Horton, Immorlica, Lucier, and Shahidi. They think about production as a series of steps which can be manual, augmented, or fully automated. AI is going to do great when you you can sequence a large number of connected simple steps into “chains.” Firms then bundle the steps and chains into tasks and jobs trying to balance specialization and trading costs.

The key idea is that AI is going to really fail when you have to interweave hard and easy tasks one after the other, because the coordination of involving the AI escalates the cost even though there is a comparative advantage for the AI to handle it.

By contrast, if you can cluster the easy steps into one big step, you can handle them all in one go with AI, and leave some of the harder steps (perhaps involving verification or more judgement) for humans at the end.

The implication for firms is that the benefits of AI investments are likely to be low at first, but then really escalate after you pass some threshold and you are now able to automate a much larger set of steps, perhaps after you reconfigure your production system to put all the easy steps together. This comes back to one of our key themes: figuring out and improving the “slow part” of a system can yield really large returns.

This raises a natural question though: how exactly are you supposed to find the value of AI in your production chain? Kim, Kim, and Koning call this the mapping problem and test this with a clean field experiment across 515 high-growth startups. The treatment is just information: they tell firms how other firms reorganized production around AI. This simple intervention alone results in 44% more discovered use cases, especially in product development and strategy. Treated firms actually cut back demand for external capital heavily, by 39.5%, while their labor demand stayed flat. Some follow-on work suggests that AI-native firms, those that start with AI already available, have fewer workers (especially entry-level workers), flatter hierarchies; reflecting different production technologies optimized around AI from the beginning.

The broader point here is that firms face a whole range of bottlenecks around actually deploying AI into production, and these are unlikely to be fixed (at least in the short-run) by just improving model quality. The upshot is that the effective diffusion of AI across the economy is likely to be limited by information access and production systems. Companies that struggle to gather and organize the right contextual data are going to suffer, especially when this tacit and diffuse information is really important for business success.

Implications for Productivity

The big question here is ultimately productivity: how much is AI going to actually increase this? A major study here is the METR experiment which conducted a high-quality randomized trial given some developers access to AI tools in 2025, to see the impacts on task speedup. They surveyed experts about the likely productivity gains, who dutifully said they expected to see one (disclosure: I was one of the people surveyed, and I also thought there would be some productivity benefits). The developers themselves thought AI was effective. The actual result, remarkably enough, was a 20% slowdown in productivity.

Subsequent followups were more positive, but METR itself is upfront that selection makes this harder to analyze. Developers now often refuse to work without AI, making it hard to establish a non-AI benchmark.

So that’s surprisingly bad news in arguably one of the best conducted studies in this space. Yet, as the developer revealed preference shows; AI use is everywhere. And we have a lot of aggregate indicators showing growth. The FT’s John Burns-Murdoch highlights the rise in websites, iOS apps, and GitHub code. One firm’s internal data suggests AI-generated code going from 3,000 lines in March 2025 to 2.26 million by August, of which 40% is shipped to production and PR cycle times having gone down. Blick, Blandin, and Deming find that industries where workers report more AI time savings (including information, finance, and insurance) also report higher detrended productivity growth. It’s hard to make sense of all of this together, by my read is we have real gains in specific workflows (especially coding-related), which are overstated in simple demos, and are slow to bubble up to the aggregate level.

On the question of distribution across workers, a consistent finding is leveling. Brynjolfsson, Li, and Raymond’s Generative AI at Work looks at the effects of AI access for customer service workers. Everyone benefits, but the largest largest gains go to less-experienced, lower-skilled workers. Cruces, Fernández Meijide, Galiani, Gálvez, and Lombardi find something similar in another randomized trial outside of firms; AI closes about three-quarters of the gap in productivity due to education at a problem-solving task.

Note there is an interesting tension here with the macro-style results Anthropic has reported overall. AI reduces skill inequality in the micro data, while increasing it in the macro data because at the macro level workers decide whether not to adopt. If you can force all workers to use AI, the skills gap might decrease; but higher-skilled workers tend to be faster adopters, so the net impact on inequality is a bit ambiguous.

Learning from History

It’s helpful to look back at history for other guides as to what “normal” technological disruption led to. One interesting historical relic is the “knocker-upper.” As factories moved to production shifts, workers had to start getting up at specific times. The knocker-upper was a person hired to wake up these workers, typically using peashooters or sticks to knock on people’s doors (knocker-uppers themselves sometimes relied on other knocker-uppers to wake themselves up; or were night owls). Obviously this went away as we developed alarm clocks. There are a whole set of professions like this: elevator operator, etc.

However, sometimes the pace of disruption is slower than you might think. The rise of ATMs did not kill the role of bank teller, despite worries at the time (i.e., the New York Times suggested in 1973 that ATMs might replace three fourths of tellers). Instead, cheaper branches meant that bank tellers specialized in providing higher-value services and actually grew in employment. It took until the iPhone before we saw a more complete disruption of bank tellers.

Another great example of augmentation-replacement-upgrading is the spreadsheet. The rise of new spreadsheet software: VisiCalc, Lotus 1-2-3, and Microsoft Excel first lead to a rise in employment for bookkeepers (a complementary or augmenting technology), until the technology got good enough to replace these workers. However, having well-kept financial books now enabled new occupations: accountants, auditors, management analysts, and financial managers to really take off. Andreessen and Horowitz make a more general point: the vast majority of today’s jobs are in occupations which didn’t exist in 1940.

These are all positive reasons from historical experience that suggest technological shifts have generally both expanded the pie as well as created totally new job categories along with them.

But there are a few more negative analogies too. There is the Engels pause: the fact that British GDP per capita rose from 1790-1840 while real wages were flat. There is deindustrialization in Northern American cities, which led to an urban doom loop as white collar work only slowly replaced blue collar work in cities. We haven’t yet seen labor displacements on this scale with AI. There is plenty of research suggesting that young workers in particular seem to be losing out, but even the mechanisms here are unclear (ChatGPT coinciding with a hiking cycle doesn’t help; then there is the congestion in labor market applications, partially AI driven).

A Normal Technology After All?

We close with Keynes’ famous essay on Economic Possibilities for our Grandchildren, written right around the same time as Rivera’s murals were painted. He wrote that in the midst of the Great Depression, a huge episode of what he correctly understood to be temporary unemployment. He predicted living standards four to eight times higher within a century, and forecasting a fifteen-hour week, since “three hours a day is quite enough to satisfy the old Adam in most of us.” He was more or less right about the living standards, but way off on the leisure part.

One part of being a “normal” technology is, ultimately, producing enough riches so we can enjoy our free time. In that area, at least, we have some interesting work by Blank, Schubert, and Zhang who find ChatGPT adoption raises individual leisure browsing, while leaving total digital time unchanged. So at home, at least, the Keynesian leisure dividend is there. But at work, the old Adam in us is intensifying efforts.

Whether AI is liberating or further entangling humans remains as enigmatic as Rivera’s murals, and there is room for many views. So far, AI is replacing some directly substitutable tasks, augmenting many more, and running into the same frictions that have throttled other general purpose technologies before it. These frictions will limit the gains, but also mercifully the losses as well. But there are a few risks on the horizon difficult to fully ignore: the loss of entry-level work, the social and economic ramifications of broad white-collar job loss, and the ever-advancing nature of scaling laws which make it ever harder to assume the comforting historical analogies hold.

I cast the lone no vote.

This vote is the occasion for two pieces I published this week: one making the case against the freeze, and one arguing what the city should do next to deal with it. The backdrop behind these pieces is the fiscal strength of the city’s economy, which is being potentially disrupted itself. This relates to another piece I wrote earlier this month about AI and city’s economy. All three pieces, ultimately, are about one question: whether New York City remembers the lessons of the 1970s.

1. The Case Against the Freeze (City Journal)

Article: The High Cost of New York’s Rent Freeze

The basic reason I voted against the freeze is straightforward: it would hold rents flat while expenses keep growing, and a large share of the stabilized stock has no other way to raise revenue:

These buildings have few ways to increase revenue beyond the adjustments allowed by the Rent Guidelines Board. Freezing their rents while their expenses remain uncapped is therefore a threat to their long-term financial viability. The nonprofit lender Community Preservation Corporation estimates that roughly a third of its rent-stabilized mortgages do not generate enough income to cover mortgage payments.

We know how the process ends:

The city has been through this cycle before. During the 1960s, 1970s, and 1980s, large numbers of rent-regulated properties fell into financial distress, which resulted in foreclosures and eventual takeover by the city. Further freezes could produce similar results by making more properties financially unviable…

Turning the ratchet further, without corresponding reductions in costs, threatens to turn a large portion of the rental stock into an expanded version of the New York City Housing Authority: more buildings owned or effectively controlled by government and increasingly dependent on public support…

The lesson is that freezing the price of a service indefinitely while its costs continue to rise does not produce cheap or abundant service. Instead, it produces deteriorating assets and, eventually, public bailouts and takeovers.

Second to First Generation Rent Control

The core issue here is the city’s rent control has slowly morphed from a 2nd Generation Rent Control system to a 1st Generation one. The distinction is:

• First Generation Rent Control — Price ceilings that keep prices basically nominally flat within and across tenancies.

• Second Generation Rent Control — Allow price increases within tenancies based on operating cost or inflation considerations, and accommodate larger changes across tenancies to reset close to market rates (vacancy decontrol), or in response to maintenance needs.

Prior to 2019, New York City had a largely functional second generational rent control system in its Rent Stabilization system (alongside a smaller first generation system). We have now seen, thanks to the tenant movement, two big shifts in this structure.

The first is the HSTPA, a 2019 law which removed many of the escape valves within stabilization. Previously, rents could be raised when apartments turned over; units could exit the stabilization system entirely after they crossed a deregulation threshold; and owners could recover the cost of capital work and apartment improvements. These provisions were not uncontroversial, and the tenant movement was correct they were sometimes abused. But collectively they provided some slack on the revenue growth to the system.

Mark Willis was one of the first people to realize what this actually means for the rent board. By shutting down most of the other available sources of revenue, it means that the RGB is now essentially the only source of revenue increases for a large chunk of the rent stabilized stock. So, structurally, you can’t really persistently offer below-inflation increases (as happened in the de Blasio) years without now affecting substantially the viability of the buildings.

The huge challenge for the rent stabilized stock is this effect happened at the same time as Covid: which put massive upward inflationary pressure on buildings, while also affecting the income of tenants.

This put the RGB in an impossible position during the post-Covid years. To follow our mandate, we need to accommodate larger revenue increases than typically; at the same time that the ability of tenants to pay remains challenged. The Board tried to balance these interests by voting for increases which were below our estimates of building costs; below inflation; below the rents set by essentially any other housing system in the country; and also below average income gains in the city. This was a “real rollback” in the sense that rent setting was much lower than broader price trends.

Trying to split the difference brought us no goodwill; in general, as you can tell, the Board faces an impossible and thankless task. This brought us to zero: the rent freeze which Mamdani campaigned on and the Board voted on. These are often compared to the rent freeze in the de Blasio years, but you can see below that was actually a very different context for two reasons. First, the actual inflation was low (this is the “PIOC” or the RGB staff’s assessment of costs), and even hit negative one of the rent freeze years. Second, actual rental revenues (the “RGB Rent Index”) typically ran ahead of the RGB guidelines, because buildings had other revenue sources (such as deregulating apartments).

Now, inflation and cost increases run red-hot, and the rent freeze has closed almost all of the opportunities for revenue growth available for many buildings. Absent other political intervention, this will accelerate all the other well-known pathologies of first generation rent control systems: building neglect, piling up vacancies as it is no longer profitable to lease out units, and ultimately financial collapse and takeover by banks and the state.

2. Where we go from here (Vital City)

Article: After the Mamdani Rent Freeze

So what are the other interventions which could help stabilize the system? I try to think more constructively about that in a piece for Vital City. The starting observation is that the “rent stabilized” stock is really two different objects. You have mixed buildings where market-rate units cross-subsidize regulated ones, which can tolerate a freeze. And the pre-1974 buildings which are almost entirely regulated which cannot, and that are in real danger. I set out an agenda there which tries to fix the situation without litigating the debates of the past:

The agenda outlined here addresses the underlying problem in rent-stabilized housing by fixing the cause: giving every distressed unit a new revenue source in the form of new market units, tax breaks, or vouchers to generate economic and social diversity across this stock. Doing so will not require repealing the state’s 2019 rent reforms, litigating the freeze or displacing a single existing tenant. The only question is whether the City can act now, while the distress is evident and there is still time to shore up building health, or whether we will have to wait for a rerun of the Bronx in the 1970s to learn the lesson a second time.

One of the mechanisms I have in mind borrows from what the city is doing to NYCHA campuses: add market-rate buildings and have them help pay the bills of struggling units:

We can pair upzonings along the lines of City of Yes targeted at the distressed stabilized stock, with an as-of-right pathway to add floors, infill development with potential teardowns, or rear additions to generate new market-rate units. These market-rate units would then become a cross-subsidy for rent-stabilized units. This would yield a rare housing win, which can add supply and preserve affordability at the same time.

Another idea I have would target the risk of worsening maintenance directly. It’s already the case that buildings with a higher rate of rent stabilized units have worse building conditions, as measured by housing violations. The idea is to condition RGB increases on those violations directly:

Finally, the rental adjustment process itself needs a change to address building conditions. The underlying issue with the current system is that rental adjustments are unconditional: A slumlord who allows the building to fall apart collects the same rental increases as a property owner who successfully upkeeps the building. This is important because we see neglect concentrating in the highly-stabilized, pre-1974 segment of the stock. The pre-1974, highly rent-stabilized segment has more building violations, an indication of worse building conditions. These buildings have more violations when the fraction of stabilized apartments is higher…

For these units, it is critical to use regulatory tools to achieve the same incentives for building maintenance that prevail in the open market. The cleanest way to operationalize this is to tie eligibility for the rental guidelines increase to a reliable, already-existing distress signal: enrollment or eligibility for the city’s Alternative Enforcement Program. AEP is HPD’s enforcement tool for the worst-maintained buildings in the city. The city defines objective eligibility criteria based on the ratio of open hazardous (class “B”) and immediately hazardous (class “C”) code violations issued over the preceding five years, plus unpaid emergency-repair charges. A building that winds up in this poor condition would forfeit its guideline increase until it cures the conditions and is discharged. This shift would amount to a rent freeze for slumlords, and would build a maintenance incentive directly into rent stabilization: letting a building decay would now cost an owner real revenue, year after year, until the violations are fixed.

This is an idea I’ve heard from so many tenants over the years in public testimony. They have reasonable concerns about whether the rental increases in their buildings actually would translate into improvements in building conditions; and their lack of belief that is happening has contributed to a failure in legitimacy of the entire rent setting operation. This is a suggestion to try to change these dynamics in a way that might get us to a more normally functioning housing system.

Vacancy Decontrol

Ultimately, though, most the of the fixes I suggest in the piece are pretty short-term. In the longer-term, we really need some more structural fixes to avoid finding ourselves with the problems of first generation rent control.

The simplest one is some version of vacancy decontrol: allowing higher rental increases after a tenant departs, revisiting the 2019 HSTPA. This avoids increasing rents on any one existing tenant, and ensures the building has increases in rental income over time even in the face of low rental adjustments for existing tenants.

I suggest a version of this in the Vital City piece (in a way that preserves mix-and-match housing affordability at the building level). In California, Costa-Hawkins protects landlord rent setting on vacant units, which allows cities like San Francisco and Los Angeles to algorithmically set rental increases at a formula below CPI (60% of CPI in SF, 90% in LA), but as tenants leave the rents readjust to market. By contrast, the rent control system in Montgomery County, which had a vacancy control and CPI + 3%, was sufficient to nearly collapse housing production. Something similar happened in St. Paul before it was rolled back.

What I take from these examples is you have to choose one. You can tolerate low, even below inflation, increases in rents on existing tenants as long as you allow increases across tenancies. But if you want to freeze rents across tenancies, you really need a robust system of rental adjustment to avoid crushing housing entirely.

3. The Backdrop: New York City in AI (Vital City)

Article: How New York City Can Come Out a Winner in the AI Age

This last piece ran a few weeks ago, on a totally separate subject (how NYC should manage AI). But it connects to the same decade, the 1970s, and the fear is that AI does to white collar work what automation did to blue collar work in that period:

Together with Stijn Van Nieuwerburgh, I have referred to the urban doom loop as the challenge of coping with dynamic responses from a smaller set of employed workers. While the full scope of AI-related impacts on white-collar employment is yet to be seen, the risk is that AI simply automates a large chunk of professional work just as we saw with the automation and outsourcing of blue-collar work, which led directly to the fiscal crisis of the 1970s.

My argument is there is a real opportunity for NYC to try to get ahead of these shifts by targeting application firms deploying AI in the real world:

New York City can capture the highest-value AI jobs by becoming the key hub for what the industry calls the “application layer” for AI. This is where AI’s capabilities meet the demands of real-world industries, most of all in regulated sectors. Think using AI to reshape law, finance, medicine, real estate, media, insurance and other sectors. These are all sectors where New York City already has an unmatched level of density and agglomeration of talent, where the economic value of AI will ultimately be realized. The opportunity for New York lies in becoming the primary destination for firms seeking to build the plumbing and deployment for AI across these different domains. This is where the high-paying jobs of the future are likely to be located, and hence New York City’s future tax base.

I think these are defensible industries and jobs with a moat because they rest on regulation and proprietary data. However, I’m not sure what the scaling potential is of this new economy.

In seizing this opportunity, New York has the potential to remain a global center for white-collar and professional work. The AI economy is unlikely to be geographically distributed the same way the pre-AI knowledge economy was. If foundational models continue to absorb routine cognitive tasks, the work that remains will concentrate in a handful of cities that already have the institutional density to support it. There may only be room for a few such hubs globally. New York is the natural candidate to be one of them, if we play our cards right.

The issue at the heart of both of these topics is whether NYC will re-run the mistakes of the 1970s, when we allowed the job base to hollow out while watching the housing stock fall into distress until the city itself became a landlord. There are a lot of strengths the city enjoys today so I don’t expect the situation to get that dire. But it requires careful and thoughtful policy, and for open discourse.

• You can also see a video and transcript of an event I did with Vital City to discuss the future of the NYC Rent Stabilized Stock

• I also did a podcast with Sandeep Ahuja on the Real Estate Meets AI podcast by Cove, along with James Scott at the Real Estate Transformation Lab at MIT

I’d also like to thank the excellent staff of the NYC Rent Guidelines Board: Andrew McLaughlin, Danielle Burger, Brian Hoberman, and Charmaine Superville. As well as the four Chairs I’ve worked with: David Reiss, Nestor Davidson, Doug Apple, and Chantella Mitchell; other Board Members, and the many New Yorkers who came out to testify in our meetings.

The latest research on this question concludes the answer is yes: there is enough of a personalized signature in Pollock paintings for a machine learning algorithm to distinguish them from forgeries 93% of the time. One particularly distinctive feature is that Pollock’s paintings are self-similar, i.e. they are fractal, in the sense that the whole painting is similar in some sense to small pieces of the painting.

This is the basic premise of AI applied to asset management and trading: that there is enough signal, amidst all the noise, that it’s worth delegating more and more of these “core” financial capital market decisions to machines. To examine how that works, this session asks three questions:

1. Is there a “scaling law” for alpha the way there is in other LLM applications?

2. If so, who captures the value?

3. How do we make sure this alpha is “real” and survives the translation from backtesting to trading, and dealing with adversarial market participants?

Scaling for Trading

Trading in markets involves three parts: first, you estimate expected returns. This can be as simple as an analyst using a multiples based approach to project winners, or in the quant version it involves connected likely future returns to numerical aspects of securities (their “factors” or “characteristics”). Then, when we have some sense of how stocks are likely to do, we run these through some sort of portfolio optimizer to balance the correlations and other risks across this basket of securities. And finally we trade the portfolio. Each of these steps can be augmented with AI, but this brings a host of new challenges as well.

Let’s start with the basic question: does the prediction of excess return scale with better models? The answer, in a great paper by Timmerman and Vulicev, is that return prediction scales with better models much as language does. Forecast performance follows stable power laws in training compute, and 25% increase in compute would have raised an investor’s Sharpe ratio by roughly 10% over the last thirty years. Throwing more and more parameters at the problem of return prediction finds more and more complicated patterns that eke out more gains.

They find a couple of other results worth emphasizing. One is that they validate the concept of double descent in financial contexts. This is a concept we covered in our first session in which model error can get worse as you approach an interpolation point, but then gets better and better as we blow past it. This is completely contrary to standard intuition in classical statistics, as well as standard finance practice; which is that we want to have a relatively small set of robust factors explaining return performance (a one factor model in the CAPM, or maybe a three factor model as in Fama-French).

Even standard machine learning and empirical Bayes methods don’t quite get us all the gains, because through regularization (i.e., setting some variables to zero if they appear noisy) we throw out some valuable signal along with the noise. To really see the gains you need to do the full neural net method of allowing for as many model combinations as possible to massively over-parameterize and train our model. Of course the industry has figured this out, even if it’s new to finance academics: the paper goes through a laundry list of major quant firms bragging about their compute and research datasets.

You may be nodding your head along if you are familiar with AI already in other applications; this is another rehash of the Bitter Lesson we also covered in that first session. But Finance, particularly capital markets trading, is different in ways that makes it harder to capture value through the AI solution.

Most obviously; the combined size of all the quant funds using sophisticated algorithmic strategies is surprisingly small: Renaissance Technology’s Medallion Fund for instance, the paradigmatic example of a sophisticated compute heavy strategy, has roughly $15 billion in assets under management. I’ve written about this elsewhere: this is because of sharply diminishing returns to actually capitalize on the computer edge. As a result, the Medallion Fund is operated on a pretty small scale, and runs it basically exclusively with insider capital to maximize the appropriable returns. Outsiders get access to lower-alpha strategies that run on a much larger scale, and have less alpha, and they are harvested primarily for management fees.

The problem is trading costs and competitors copying your strategy and trading against you adversarially quickly eat the returns from a better strategy. The world is non-stationary, i.e., always changing, so the model has to track a moving target, and distinguish genuine signals from patterns that just fit the training data and fall apart when deployed.

What we are left with is a relatively small number of quant funds heavily competing on compute, which substantially dilutes the profits they can earn. The scarce complements, in this world, are data, infrastructure, and latency, which are able to capitalize on some of the rents.

Taming the Factor Zoo

This issue of data mining has really plagued the whole academic literature on return predictors. Academics have put together over 300 published factors which supposedly predict stock returns, which is the factor zoo problem. Many of these factors, when tested rigorously out of sample, tend to fail. Even when the factors work, it’s unclear how to best combine and trade them.1

The landmark paper here is Gu, Kelly, and Xiu’s “Empirical Asset Pricing via Machine Learning” which compares a range of ML and AI methods in figuring out what combination of factors works best to predict returns. They find, naturally enough, that neural net methods work very well, with out-of-sample R² around 1.8%. Unlike in other domains, they find that a neural net with a relatively small number of layers tends to perform best, which they interpret as a sign that finance is prone to overfitting because of a relatively low signal-to-noise ratio. The right amount of model complexity, however, has increased over time. A fascinating question which no one has tackled yet is what the impact is of more sophisticated trading strategies on the prediction problem itself. Is finance going to get inherently much more complicated as people use more AI to trade?

This group followed that up with another paper “(Re-)Imag(in)ing Price Trends” which is a really crazy paper trying to use AI to resuscitate technical analysis through image analysis of price charts. Though academic finance and many practitioners tend to look down on this kind of chartist analysis, people also do generally believe strategies like momentum and reversal which also rely on trends. The AI here, interestingly enough, seems to find a lot of complex patterns in the charts that go beyond validated signals and seem to transfer across domains, and even time scales. This goes back to the fractal example we started with: financial markets also appear to be somewhat self-similar, in that a model fit on very high-frequency data (which is ideal, since we have a lot of this data) can extrapolate somewhat to say daily or monthly returns.2

Map vs Territory

All of this gets into a map versus territory debate: when is the model giving us a true, but highly complicated, representation of the world, and when is it picking up spurious relationships that only work in-sample? There is an interesting debate here, which I want to highlight to stress how this remains a contentious issue even in frontier research.

On the one hand we have a continuing thread of authors Kelly, Malamud, and Zhou, who argue for the “Virtue of Complexity in Return Prediction.” This argues in a similar vein as earlier that over-parameterized models can outperform in return prediction; where they use Random Fourier Features to blow up the dimensionality of the problem and document better performance (in terms of alpha, Shape ratios, information ratios). Then we have Stefan Nagel who argues against this Seeming Virtuous Complexity. He argues that when you get so many more variables (or “features” in ML-speak), the apparently complicated strategy actually just boils down to a weighted average of recent returns, which is just a volatility-timed momentum strategy. So what seems like a sophisticated new complex strategy is really just a well-known one, and the apparent performance in-sample is sort of a coincidence, rather than a fundamental aspect of the strategy per se.

Novy-Marx and Velikov highlight some of the broader challenges here in a nice paper “AI-Powered (Finance) Scholarship” which gets into the scope for data mining to produce spurious AI-powered results faster than referees can triage them, complete with made up narratives to support the strategy.

This debate will continue on, and the benefits (or lack thereof) of complexity in return prediction don’t hinge on any one paper or method. But this illustrates the challenges of interpretation when the model complexity blows out. If we don’t know what our strategy is really doing, we don’t know what risks we are really taking and what therefore might go wrong as the world changes.

Portfolio Selection

Even if we “solve” the problems above in figuring out what stocks or aspects of stocks meaningfully predict returns, we are then left with the problem of how to best sort these stocks into portfolios. This is the basic Markowitz portfolio problem: with N assets, a bunch of expected returns µ, and the covariance matrix across stocks Σ; we want to find the combination of stocks which minimizes return for a given amount of risk.

The problem here is the Σ. A covariance matrix for N assets has N(N+1)/2 parameters, which is 55 for ten assets, but 125k for the S&P 500, and about 4.5 million for the Russell 3000. Estimating this reliably requires a time series much larger than the number of assets, which we don’t have in monthly data. This is of course the dreaded “curse of dimensionality.” Michaud pointed out back in 1989 how brittle the mean-variance approach is to any kind of estimation error in this process, which has led to conclusion that it’s vary hard to beat indexing (i.e., 1/N for each asset).

A foundational application of AI here is “Asset Embeddings” by Gabaix, Koijen, Richmond and Yogo. The idea is that rather than estimating all of the pairwise correlations, we do a lower dimensional reduction of securities into an embedding space. The question is what is the natural analogue of relatedness between two stocks (akin to words being sequenced in a sentence), so you can produce something like next-stock-prediction in a series.

The solution they come up with is to look at portfolio holdings. So they mask a position in a portfolio (an example in the paper is to blank out Zoom in ARK’s holdings), and see if the model can recover the holding based on all the other stocks in the portfolio. Repeating this at scale produces a network of securities which seems pretty reasonable (i.e., Apple is “near” Adobe, in the sense that funds tend to co-own them, and Citibank is near other banks and financials), and they do some interesting work using earnings transcripts to isolate some of the narrative factors that seem to drive location in embedding space, and the embeddings themselves explain valuations. This is a very neat and natural application of the embedding methodology to finance, and the narrative or expositional layer also helps us to make sense of the underlying black box.

There are a few other approaches. Zhang, Zohren and Roberts’s “Deep Learning for Portfolio Optimization” skips the return forecasting step entirely to train portfolio weights which maximize Sharp ratios. López de Prado’s “Hierarchical Risk Parity” clusters assets by similarity and allocates risk across and within clusters instead of dealing with the covariance matrix.

Also, I think it’s worth thinking about the economies of scale here, which is Grinold’s fundamental law of active management: the information ratio (additional return per unit of risk) is roughly the information coefficient (correlation between actual and expected return) times the square root of breadth (number of independent bets): IR ≈ IC·√BR. This means that the skill per bet you make matters a lot, but the number of bets also stacks, albeit at a diminishing rate.

This is the basic economic logic I think for the quant shops (Two Sigma, Citadel) as well as the multi-manager pod shops. If you have nine independent signals, but can find a tenth one; you can stack them and earn additional returns uncorrelated with each other. This means that the limits of exploiting any one strategy or niche can be partially addressed by layering unrelated signals, at least up to some point. So we end up with medium-scale economies that account for quant and pod shops not taking over the entire market, but still having some fairly substantial size or AUM.

Transactions Costs

Even if we solve the other problems, we have to then translate an optimal portfolio into trades. Esakia and Goltz look at how all the complexities inherent in ML solutions translate when you incorporate realistic trading costs. They find the value of the nonlinearities in particular seems to go away, while the information breadth advantages seem to persist. To really get the value of the nonlinearities just entails so much trading activity that realistic frictions eat away the true value you can get.

A big lesson here so far is that the zero-sum nature of financial trading places huge limits on the value of AI deployment. It also means the social value gained is also maybe not so large, to the extent AI just fuels competing compute-heavy organizations. So I want to also stress two areas where I think we see more broad based social gains.

First, where AI also seems to be paying off is through augmentation. It’s hard to tell how much of this is marketing copy, but Norway’s Sovereign Wealth Fund reports something like 20% productivity gains through automation of newsflow monitoring, analyzing earnings call transcripts, and trading efficiency gains.

Second, personal wealth management seems like another high elasticity corner of the investing world. One simple example is tax loss harvesting, in which you sell a losing stock to bank a capital loss, and swapping to hold the exposure with another stock that is virtually the same. This strategy has well known alpha; but AI lowers the cost of delivering this and so personalized, tax-aware rebalancing shifts from being an expensive bespoke service to one that is nearly free, and so can scale rapidly across personal accounts.

The Future of Asset Management

So let’s put this all together and think through the equilibrium. As we get better and better prediction engines, the first order thing we expect to happen is that markets get more efficient and alpha is harder to find and sustain. That’s good for passive market investors and hard for discretionary stock pickers.

The quant heavy approach will still reward compute and scale, but the role of adversarial trades leads to a double-edged sword. On the one hand, it means that gains are likely to get sniped and diluted quickly, diminishing alpha further. However, there may be fresh new gains from forecasting the ever-changing strategies of others (one example would be trading around index fund reconstitution. This is an example in which new algorithms, i.e., index fund creation, results in predictable new order flow variation and hence trading opportunities). In the long-run, we expect the gains to accrue to whoever owns the relevant scarce complements to this activity (data, infrastructure, latency, etc.).

So AI will likely continue to shape the market, but it’s going to be very hard to capture the value in doing so. One promising area to try to make that happen is through activities which do new things to try to expand the pie: taking services which used to be costly and reserved for the rich, and finding ways to distribute them at a much broader scale. It’s not enough to figure out whether a painting is recognizably Pollock or not; can we then create a million more examples and place them in the homes of anyone who wants one?

Readings

• We have two cases; one on Citi and Aladdin wealth ([1], [2], [3]), and one on iBuyers ([1], [2], [3]).

The Base Rate Problem

We start off this week with Caravaggio’s The Cardsharps, which features a young man being cheated at cards through an accomplice looking at the cards signaling to his partner who has a few cards tucked behind their belt. I like this painting because it features all the everlasting elements of fraud: asymmetric information, coordination among bad actors, and the innocent mark. Fraud and financial crimes have since drastically shifted through scale and ease of execution through technology. Consumers report $12.5 billion in fraud to the FTC; global reported payment card fraud is maybe $30 billion a year; and global compliance costs on financial crimes is about $60 billion a year.

The core technical problem in fraud detection is the base rate problem. Despite the absolute magnitude of fraud, it’s quite rare in any given transaction, say less than 0.1%.

This means that the most “accurate” fraud detection algorithm is always going to be the classifier which tags all transactions as legitimate. It’s going to be 99.9%+ accurate! But this would be a terrible algorithm to deploy because failing to tag anything as fraudulent will be an open invitation for more crimes.

This is the fundamental tradeoff at the heart of fraud detection. Tightening the rules to capture more fraud (lowering false negatives, Type II errors) inevitably means flagging more legitimate transactions as fraudulent (raising false positive, Type 1 errors). These errors have complicated costs which vary based on contexts. Customers really dislike false positives, so tagging more fraud may lead to more customer churn. But letting real fraud through can also lead to huge losses, regulatory penalties, and reputational costs.

Beneish and Vorst have a nice paper illustrating this tradeoff using a range of fraud prediction models. Even the best models they consider have false positive rates in excess of 100:1, meaning that they tag 100 legitimate transactions as fraudulent for every one true fraud they capture. These tradeoffs are so bad they estimate it’s generally not cost effective to even adopt the fraud detection models.

From Rules to Representation

The broader history of fraud detection is similar to other areas of AI deployment in finance, particularly risk management which we had discussed in week 3. Prior to the 1990s we typically had manual review, which was expensive and slow to scale. From the 1990s and 2000s we had rule-based systems which were more scalable, but rigid and easy to game. The 2010s brought supervised ML techniques, such as random forests, which brought in labeled data to let the model learn about fraud drivers in more adaptable ways.

The impact of AI here has been to move towards richer representations of transaction data in ways that allow for more sophisticated analysis of fraudulent patterns of behavior. Purda and Skillicorn for instance show how even simple bag of words applied to the management discussions of financial reports can distinguish fraudulent from truthful filings. This shows that deceptive language carries detectable statistical signatures in word patterns which can be carefully mined for detection.

Dal Pozzolo and others writing using data from a real world credit card issuer, with 75 million transactions, highlight some of the challenges here. You want to develop real-time fraud detection algorithms, but are faced with the challenges of 1) concept drift (consumer and fraudster habits change over time), 2) class imbalance (the base rate problem we discuss above), and 3) verification latency (you have a delayed feedback loop in verifying fraudulent status).

One interesting AI solution to this problem comes from Capital One. Bruss and their colleagues developed DeepTrax, which is a graph embedding approach to financial transactions data. The idea is to treat sequences of merchant transactions just like words in a sentence: merchants which appear in similar transactions contexts should also be “close” in embedding space. Their derived embeddings seem to show intuitive clusters: KFC is similar to Taco Bell, Little Caesars, and Burger King; because apparently consumers tend to purchase these products in a similar enough context. The embedding dimensions even encode some attributes of branding or price point; so there is a directionality to the embeddings going from Ritz-Carlton to the Fairfield Inn among hotels, and Banana Republic to Old Navy for retailers along the same axis.

This is a good illustration of the fundamental objective of AI from session 1: generating maps or representations to simplify decision making. By doing this dimension reduction across firms, we can more easily figure out which purchases are likely to be “anomalous” and the authors suggest the embeddings improve the precision-recall AUC by about 1% over baseline, and also allows for a bit more simple ML implementation.

A natural question, given the succession of rules to ML to embeddings is whether LLMs can also process fraud directly. Tan, Ma, and Zhang look at that in this FinFRE-RAG paper. They translate transaction features into language prompts, select the most important attributes of a transaction, and retrieve similar transactions (using RAG). Interestingly, this approach still lags behind specific purpose-built ML classifiers, and the raw LLM approach (without RAG) is also quite bad. But LLM plus RAG has the added benefit of interpretability: it reasons about specific fraud patterns, and so can open up the dreaded “black box” of ML classifiers with some interpretability.

Graphs and Dark Fleets

An important development in fraud detection methods has been exploring graphical methods. This takes advantage of the reality that fraud is fundamentally relational, not just transactional. Chang, Zou, Xiang, and Jiang have a good review of graph neural networks for fraud detection. The basic idea is that graphical methods can tag fraud where the transaction itself looks fine in isolation, but the broader pattern of who is transacting with who, what’s the device structure, etc. can help surface novel fraudulent patterns. These methods operate at the level of the node (i.e., is the account fraudulent?), the edge level (is the transaction suspicious?), and the graph level (does this cluster of accounts form a fraud ring)?

There are some striking examples of using these methods. Fernández-Villaverde, Li, Xu, and Zanetti built a chip clustering model to detect “dark shipping.” These are ships which turn off their transponders to evade sanctions, and so are naturally enough challenging to measure. Their model incorporates aspects of vessels, their behavior during signal gaps, port visits and ship-to-ship transfers, to assign each ship and trip a “dark score.” They estimate that dark ships transported almost 8 million metric tons of crude oil each month from 2017-2023, with China taking about 15% of the supply. Dark shipments grew dramatically for instance after Western sanctions on Russia.

Another grim but illuminating example is the work by John Griffin and Kevin Mei on pig butchering scams. These are crimes in which victims are gradually lured through fake crypto investments which pay off in the short-run before the victim is finally defrauded for large amounts. They measure the digital signatures of these crimes through blockchain flows, which allow them to tag entire networks and clusters of fraudulent actors. Criminal networks sent over 32,000 small trust-building payments to exchanges used by US and European investors, before moving the amounts elsewhere and existing, typically through Tether. Blockchain is an interesting application here in that the digital ledger provides both the anonymity to engage in these criminal activities, while also providing the digital breadcrumbs to reconstruct the crimes after the fact.

Opening the Black Box

While there is some promise to these newer fraud detection problems, they face serious gasps in deployment. Bhatt and others survey firms and find that most of the deployment of ML is there to serve internal users, such as engineers debugging models, rather than the end users affected by model decisions. The lack of transparency and explainability behind complicated model decisions is a key barrier and friction around their broader adoption and use.

We saw above that LLMs can help to address one aspect of that explainability gap, through model rationalizations of classifications. Another important approach in practice is Shapley values, which decompose the model prediction into the contribution of each input variable. This helps to compute counterfactual explanations for what factors led to a credit card payment or money order being flagged as fraudulent.

Addressing the Compliance Burden

Figuring out how to improve the deployment of these tools is important because fraud has costs which extend far beyond the actual dollar losses. Financial firms spend billions to comply with Anti-Money Laundering (AML) and Know-Your-Consumer (KYC) regulations. The regulatory environment around finance has generally tightened after the financial crisis, where real challenges in fraudulent behavior led to a substantial regulatory response and the exit of banks from lending in mortgages as well as small business lending.

This leads to a situation I’ve discussed before: the advancement of technology in finance has been a disappointment in many dimensions. An easy way to see this is to simply look at mortgage rates and spreads over the last 26 years. Despite the rapid advancements in computing and technology, we’re really no better (arguably, we are actually worse) in pricing and originating mortgages.

This situation partially reflects the rising costs of regulatory compliance, which now exceed $10,000 for a mortgage loan. Naturally enough, the real costs and risks of fraud have set off regulatory demands for tighter paperwork, which are costly to provide.

This is where AI has the ability for broader impacts: through organizational and document burden changes which can automate KYC checks, accelerate regulatory document review, triage suspicious activity reports, and more broadly try to address the effective “tax” that compliance imposes on financial activities.

AI as the Sword and Shield

But we also have to be careful about the dangers here, following the “cause of and solution to all of life’s problems” dynamic. If the cost of meeting regulation falls down; than we might expect regulators to come back with a whole new stack of regulations to address, leaving the net effect more uncertain. This is a general equilibrium principle of what happens to the economy overall when the costs of some functions go down a lot, which we still have yet to think through.

AI also creates new vulnerabilities even as it addresses the old ones. Deepfakes and social engineering attacks are becoming extremely effective. The $25 million crime perpetrated in 2024 when an employee joined a seemingly routine video call populated with deepfake agents was just an early signal that the costs of phishing and other attacks has been dramatically lowered by AI.

Fraudsters have in fact access to the whole suite of AI technologies that banks do, resulting in another “arms race” dynamic. They can use the same embeddings, graph analysis, and language models to understand what detection systems are doing and figure out how to game them. The use of AI also creates a whole new set of attack vectors; from vibe coded software that lacks critical safety checks to novel prompt engineering attacks on the AI system themselves.

Whether the defenders or attackers benefit more from the dramatically increased scale of AI remains I think very much an open question.

Readings

• We have two cases this week; one Stripe (readings here and here), and Plaid (here, here, and here).

Subscribe now

I have a new piece out in National Affairs with Steve Teles, “Industrial Policy for Housing Construction,” which makes the case for supplementing our land reform efforts with strategic government support for factory-built housing. This would entail a mix of further deregulatory pushes in building codes, along with government funding to collectively address the coordination failures which have prevented the emergence of a competitive factory-built sector.

Why Housing Isn’t Built in Factories Now

We start the piece with the basic facts around construction productivity. As Cardiff Garcia had already outlined over a decade ago, and many others have written and discussed: measured productivity in the construction sector actually appears negative. This is a really surprising fact in light of the tangible productivity growth in other sectors of the economy, and the basic fact that if you visit a construction site, it does look like a lot of things have advanced over time.

The root cause of this, on some level, has to be the fact that most houses are built on site, rather than in factory environments which have formed the basis of productivity advances in other fields. Factories, of course, enable specialization, regulated environments, quality control, capital intensity, and the simplification and standardization of the production process which make productivity improvements over time possible.

And, of course, housing was once built in factories. The Sears homes are just the most prominent example of what used to be a far more expansive system of factory-built houses shipped on site. So why didn’t these techniques take off more?

I think it’s instructive to compare housing to automobiles or planes. These goods all share durability, which is relevant because consumers cut back on durable purchases most of all in downturns (since they can more easily delay these purchases). This generates large boom-bust cycles in production which is deadly for firms considering large capital investments. For this reasons, the housing, auto, and airplane sectors all faced large existential risks when the Great Depression began.

However, autos and ships were able to tap into much larger markets, and importantly enjoyed substantial government support. We did industrial policy for these sectors through R&D, guaranteed purchase support, and and so forth, because they had national security rationales and export potential.

Housing by contrast fell by the wayside. The mid-century developers at Levittown were able to bring back some of the economies of scale in housing construction through massive land assembly. But now we face a world in which land assembly near urban center is much more complicated; the regulatory environment (both land use and building codes) is far more complex; and transportation costs further limit the size of construction marketplaces.

The result is a construction industry which is fragmented geographically and across tasks. Firms have reacted to the boom-bust cycle with capital-light, contractor-heavy workloads which shift the risk of housing ramp-ups and declines onto a more precarious workforce and suppliers.

The Industrial Policy Solution

Our recommendations to change this situation and move us to a different factory-built housing equilibrium fall into a few buckets:

• An ARPA-style R&D program for housing innovation, to kick-start innovation.

• Government procurement commitments to smooth demand (think FEMA stockpiling modular units, military housing contracts), and incentive subsidies for factory-based construction.

• Financial risk mitigation modeled on DOE clean-energy loan guarantees to address some of the downside risk and enable scaling up.

• Reform of the legal treatment of manufactured homes (they’re currently classified as personal property rather than real estate, which limits mortgage access).

• Harmonization of building codes across jurisdictions so firms can more easily access different markets.

• Direct government land banks to help the process of parcel assembly and accumulating land for large-scale development.

The piece also talks about the political economy of this proposal at length. One key insight is that this is an unusual form of industrial policy. Because construction is a purely domestic industry, there’s no global marketplace to distort and no risk of creating a “national champion” dependent on subsidies. The goal is pro-competitive: use federal coordination to break down the patchwork of local protectionism that fragments the market and insulates incumbents from competition.

Broader Connections

Here I want to go a bit beyond the piece to think about some other linkages

The Bitter Lesson in Construction

In my AI in Finance course, I teach students about Rich Sutton’s “Bitter Lesson,” which is the idea that general methods using scale and computation always win out over human-designed domain-specific approaches.

Construction has consistently not applied the bitter lesson. We do construction using the craft approach, with on-site assembly, because that’s flexible and fits local conditions. But the same underlying principle that suggest that neural nets should beat out chess grandmasters suggests that factory methods which leverage standardization, capital intensity, and scale should eventually dominate bespoke on-site construction, if we can figure out the bottlenecks to this deployment.

The Swedish Comparison

One of the things our piece gets into is the comparison with industrial policy experiments elsewhere, particularly in Sweden, where the vast majority of multifamily is at least partially constructed off-site. So how well does Sweden build?

The best evidence I can find suggests that Sweden does indeed have relatively cheap building costs, though it has “spent” some of the gains on relatively demand climate and emissions requirements.

Sweden reports purchase prices of 44,401 SEK/m². This includes VAT (25%) and land costs (let’s say 18%, which was their estimate from 2018). We also need to convert this “usable floor area” definition to the US standard (which reports gross; an adjustment of roughly 17%). This gets to a range of $250/ft², which is maybe more like $290-320/ft², in Stockholm, despite the stringent performance based codes (triple glazing, mandatory HRV, airtightness testing, etc.).

This is comparable to mid-range US markets, but is considerably lower than expensive US coastal markets which range from $400-600/ft². These benefits also don’t fully account for the greater speed of Swedish prefabrication. These aren’t radical differences and I’m open to people pushing back on these estimates; but at a rough glance it does suggest that countries which have coordinated to a more factory-built approach are able to build a bit more cheaply despite high local labor costs.

Remote Work and the Geography of Construction

Another thing which didn’t make it into the piece is a big theme of prior discussion here: remote work. The shift from fully remote and hybrid work are increasing housing demand in secondary metros, suburbs and exurbs, and smaller metros where construction costs are increasingly the binding factor behind housing construction. The geographic dispersion of demand from remote work is therefore expanding the addressable market for factory-built housing in the the places where it’s also most viable.

There’s also a political economy angle here we did talk about in the piece: factory-built housing creates a reciprocal economic relationship between urban and rural areas. Factories located in lower-cost regions (central Pennsylvania, the Central Valley) could ship panels and modules to higher-cost cities. This benefits both rural areas (which get more manufacturing jobs) as well as cities which enjoy cheaper housing. It’s a rare win-win relationships here which might, optimistically, help tie together shared economic prosperity across geographies.

Value Capture for Land Assembly

Our piece proposes land-banking authorities to assemble parcels for large-scale factory-built housing developments. But where does the money come from? This connects directly to the property tax and value capture mechanisms I’ve been thinking about. Tax-increment financing districts, where future property tax increases from development fund upfront land acquisition, can potentially make land assembly self-financing. Some Sunbelt states already aspects of this, such as MUDs in Texas. The question is whether we can get more jurisdictions to adopt these instruments in conjunction with factory-built housing programs that would simultaneously bring down the the cost and speed curve for housing construction.

The Factory-Built Future

The broader goal we have for the piece (again link here) is just to raise the salience of the construction productivity problem in the housing discourse. Which is not to say that zoning and regulatory reform aren’t important on their own: see our earlier discussion on this, and we’ll have more related data products here soon. Instead it’s more that success in addressing some of the bottlenecks around housing production just shifts the constraints elsewhere in the housing pipeline. We are putting out a few ideas to try to address this construction problem: maybe people agree with them, but hopefully it at least inspires people to think creatively as well.

The opportunity I think is huge. Even modest improvements in the cost, speed, or quality of housing will allow for a broader set of housing investments to pencil out, expanding the size, affordability, and scope for our housing future.

Other Links

• You may also want to check out this podcast with the New Bazaar, which went into industrial policy for housing and other topics.

  Subscribe now

AI and LLMs are disrupting this sector by processing text at a scale and speed which fundamentally shifts the core economics of business analysis. Previously, this was a labor-intensive process bottlenecked by the speed of human reading capacity. Now, some of the core analytic functions have become commodified due to the rapid pace of AI advances. At the same time, faster and cheaper information doesn’t always help people make better investment decisions if the bottleneck shifts elsewhere. AI also enables completely new forms of intelligence functions: in particular in silico agent simulation. But are these information tools accurate?

So the key questions this week are: what is going on with the quality of information we summarize or simulate, and does it help us make better actions? And even bigger picture: where does the alpha go if everyone has access to AI tools?

The Arms Race in Textual Analysis

The history of text analysis in finance is a good illustration of the “bitter lesson” of scale economies combined with the “follow the price” principle from Session 1. Each generation of tool analysis commodifies one layer of analysis, pushing the alpha or edge further up the complexity stack.

The first generation was simple dictionary-based sentiment analysis. Tetlock’s classic 2007 paper counted words in one WSJ column using the Harvard psychosocial dictionary, estimated a simple pessimism factor, and showed it predicted Dow Jones returns. This was a big advance at the time, even though it built on a pretty simple measure. As we discussed back in Session 1, further advances from here developed finance-specific dictionaries (Loughran and McDonald) and chained together word combinations in n-grams and bag of words.

Then we get to LLMs. Lopez-Lira and Tang showed that GPT-4 can classify news headlines for stock market impact with pretty high accuracy (capturing 90% of the hit rate for initial reaction). The really interesting result though was that the Sharpe ratio of the LLM classification trading strategy was steadily decreasing over time alongside rising LLM adoption. The information edge from reading headlines was apparently real, but got competed away and is now largely priced in.

As that initial reading layer is commodified, we can move up the complexity value chain through other analysis opened up by LLM analysis. Leland Bybee takes this in a really interesting direction by using LLMs to generate economic beliefs or expectations by examining historical headlines, allowing him to build 120 years worth of economic sentiment data from newspapers. This also starts to move us in the direction of “representing” economic beliefs through their textual breadcrumbs.

Another advancement of the complexity ladder is Hansen and Kazinnik who show that ChatGPT can decipher “Fedspeak;” which is to say the deliberately ambiguous or obfuscated language the Federal Reserve in monetary policy. You see some interesting advances across models in this textual analysis. GPT-3 tends to analyze the actual text of the statement itself, while GPT-4 is able to correctly map some of the subtext of statements as well with reasoning which lines up with that of a human analyst. The emerging model capacities here allow AI to escalate all the way to complex narrative interpretation (which starts to replicate sophisticated Romer-Romer style interpretation of complex texts).

As an interlude, I also have to mention Joe Weisenthal’s vibe coded tool, Fedlock. Rather than direct classification, this tool takes an elo-style competition in which pairs of Fed reserve statement are evaluated by an LLM-judge as to which statement is more dovish or hawkish; which produces a complete ranking of all statements. So we have these two application types of LLMs: direct classification, and pairwise elo-comparisons, which can be useful in different contexts (it seems like the elo approach produces a bit more of a continuous and smooth ranking, while the direct classification approach has some tendency to produce extreme outcomes).

We can move on from here to analyzing equity analyst valuations, in some work by Bastianello, Decaire, and Guenzel. They draw on 2.1 million equity analyst reports and use LLM tools to diagnose the topics and mental models analysts use in their representations of firm valuation (they have a followup paper which focuses on the comparisons between accuracy and complexity on DCF vs multiples-based valuation models).

At the current frontier of AI representations, we have Suproteem Sarkar, who generates economic representations of entire firms. To do this, he estimates vector embeddings of firms from financial news discussions that quantify the economic features and themes of each firm’s coverage. To do so requires solving one important problem in historical bias: the LLM has future knowledge, which potentially contaminates its assessment of historical data. There are two general ways to address this: “masking” historical data by swapping names to de-contextualize the LLM, or training models on the basis of historical data alone, which is what Sarkar does here.

The output from all of this is an interesting “map” or representation of firms which captures interesting relationships between firms. Cross-sectional comparisons show that “similar” firms in this embedding space also tend to see similar stock price co-movement. Shifts in the embedding representation also predict stock market changes, suggesting that part of the variation in valuation comes from shifts in market perception of the firm’s representation (including possible misperceptions such as periods when attention-grabbing features like “the internet” in the 90s or “AI” in the 2020s dominate the embedding location).

Beyond textual data, there is a whole alternative data ecosystem which represents another dimension of how AI has disrupted market intelligence. This includes everything from satellite imagery, web traffic, app downloads, credit card data spend, and so forth.

The common pattern here is that every stage of LLM scaling speeds up one layer of textual analysis. Initially, this produces some alpha for people with access to the analysis tool, while also unlocking new layers of complexity for people to analyze. However, that next set of documents ultimately becomes the raw material for the subsequent LLM deployment, and so we have an escalating level of textual and market intelligence analysis.

Attention Reallocation

A deeper question here is what all of this data does to the structure of forecasting itself. A big challenge is that for all the virtues of “big data;” we will inevitably have more information about the cross-section than the time-series (i.e., we can get a lot of information about conditions all over the country today, but are limited in our ability to analyze historical data before we were collecting all of this alternative data).

Dessaint, Foucault, and Fresard analyze some of the implications of this bias on forecasting horizon. Their analysis suggests that analysts have grown more accurate on shorter horizons, where all of the alternative data helps generate near-time accurate forecasts, but have actually gotten worse at long-term forecasts.

We see similar trends in the attention across stocks. Farboodi, Matray, Veldkamp, and Venkateswaran find that the improvements in forecasting accuracy are concentrated in large growth stocks. These are stocks whose value grew dramatically from the value of data and other intangible assets, and the market is increasingly focused on trying to understand the most valuable data. This leaves out small and value stocks, whose pricing is actually stagnating or falling behind.

An even broader version of this attention reallocation situation is discussed in this paper by Hao, Xu, Li, and Evans in Nature. They find that AI adoption in science expands the scope of individual scientist impact, but reduces the collective breadth of scientific focus. In the long-run, AI advances might eventually expand our reach to other currently data-poor environments, but in the short-run it seems to contract our focus.

The implication for finance is that while AI might get very good at reading 10-K’s and earnings calls, and dilute the alpha from that information commons, there are likely to remain pots of alpha from being able to mine hard to access information.

We see a version in this phenomenon in Kim, Muhn, and Nikolaev who find that humans outcompete machine analysts in domains involving institutional knowledge, such as intangible assets or financial distress, but lose in other domains characterized by broad information. This is consistent with many other results we look at in this course. The “bitter lesson” applies to domains with voluminous data that can be effectively mined for clear insights. But interpretation, judgement, and the ability to act on private or tacit information remain valuable, at least for now.

Simulating Your Customers

This discussion so far is about AI advancing its capacities to replicate or replace part of what humans typically do. But many of the most exciting AI applications are about using LLMs to do completely new things. One domain where this has really grown is the idea of having AI simulate economic agents. This idea has a few different intellectual grandfathers, but a very influential strain is attributable to John Horton and gets at the idea that become LLMs are trained on human-generated data, they contain implicit representations of human behavior in other contexts. They can be considered a homo silicus with the ability to be given endowments, preferences, and information and their behavior explored in simulations.

The results so far are very interesting, if early. Manning and Horton build “general social agents” combining social science theory and empirical data, and show it matches human play in novel games better than some standard models. Park et al. create generative agent simulations of 1,052 people by applying LLMs to qualitative surveys, and showing these surveys replicate participant responses in the General Social Survey 85% as accurately as the human agents themselves in a followup (another framework here is the “digital twin” methodology).

This technology, if it pans out, potentially unlocks a whole new layer of understanding consumers, voters, firms, governments, and other agents. Rather than running expensive focus groups, we can trial in a digital laboratory how people would respond to new products, pricing, or marketing campaigns.

A lot of this literature has thought about the careful ways LLMs need to match human behavior to ensure their responses to novel questions are in-line with their responses to other questions. Gui and Toubia for instance find that variations in the context provided to the LLM for apparently unrelated information tend to bias the LLM, and the resulting demand curves we get don’t match human demand curves well unless we fully “unblind” the model to the full experimental structure. This is a big problem: the whole point of running an experiment is that subjects shouldn’t know the hypothesis, but LLMs may need that precise context to work well, and may be distracted by unrelated context.

Some things that seem to help are demographic matching humans to specific personas. Fine-tuning machine responses on actual choice behavior also seems to get us closer to realistic willingness-to-pay estimates for existing products and features, but it remains challenging to extrapolate these improvements to novel product categories or consumer segments. Translating machine output into free text responses, and connecting those to Likert-scales, also seems to help (this result is a bit reminiscent of the elo idea above that the LLM may be better at comparing two texts rather than classifying them in total).

We’re left with a deeper question of whether synthetic agents are really representing the full structure of human beliefs, or if the map here is a poor guide to the territory. Barrie and Cerina find that LLM personas are considerably more coherent in their beliefs than are human agents. People tend to hold a mix of somewhat contradictory (Walt Whitman: “Very well then I contradict myself, I am large I contain multitudes”). The LLM agents sort of don’t contain multitudes and are a bit too consistent, which of course is a natural limitation to their deployment to understand human ideology.

Where we stand today I think is that agents work best when we have natural scaffolds to help structure their behavior: either a lot of background text on the personas we want to match, actual choice behavior from those agents, or some sort of relevant theory of agent behavior. They also seem great to deploy for low cost of error trialing purposes. But we are probably not yet at the point when they substitute for experiments on actual humans, especially when we thinking about novel scenarios which require extrapolation rather than interpolation, or politically polarized populations (I’ve written more of this challenge in out of sample prediction here; and do think we can do more creative deployment in the future).

Where is the Moat?

If intelligence becomes cheap and broadly available, what remains scarce? What are the new bottlenecks, and where do rents flow to? I think there are three broad answers:

1. The first, most obviously, is proprietary data. Data is key to the entire analytic platform here, and especially in finance contexts we care most about new data which is always arriving. Data becomes the scarce input from which new algorithms can squeeze fresh insights.

2. Second is the speed of action, rather than the speed of reading. In finance applications, this is going to be your investment committee process or portfolio management decision, rather than your analyst reading time. If these remain your bottlenecks, then it doesn’t really matter how fast document reading gets (Amdahl’s law again).

3. Finally there is judgement under uncertainty. For now, humans seem to retain an edge in contexts of soft information, broader context, long-horizon prediction, novel situations, private information, and stochastic environment. This is going to be increasingly important when we bear in mind the adversarial response in finance. Even if machine-parsed text is quite accurate now, you can bet that market participants will strategically alter their words to break these relationships). This inherently adversarial and zero-sum nature of financial markets means that we are looking at a moving target, which is going to be a challenge even as model capacities improve.

A last image to leave you with: the Astronomer, by Vermeer. The astronomer has the heavens through the window, but is studying the globe: a map representation of the world. The question we’re always left with is whether this globe (or any map) is good enough to navigate by, or whether it misses something essential about the world.

Cases

• We have two cases this week; one on AlphaSense (see here, here, here, and here); and Morgan Stanley (here, here, and here).

Subscribe now

The key punchline from this session is that AI in risk management, to quote Homer Simpson, is the cause of and solution to all of life’s problems. Better models can improve prediction, expand credit access, lower losses, and help address operational challenges for financial institutions. But they also introduce new competitive dynamics, help enable adversarial financial actors and cybercrime, and generally add to new risks. Some of the best ways to tackle these risks, naturally enough, entail adopting yet more AI.

Trust and Verification

Credit scoring is a long-standing information problem in finance, and has seen substantial automation over time. Historically, credit access was deeply tied to one’s reputation and social standing, as trust played a key role in repayment. As early as the 1840s in the US, this information started to get codified into ledgers. Bradstreet was founded in 1857, and commercial credit started in an alphanumeric scoring system in 1864 by Dun (now Dun & Bradstreet). Even at that point we had the basic ingredients to put together an algorithmic model of credit scoring: some degree of private sector surveillance to collect information, information sharing to make it accessible, and a ratings system which turned that information into actionable content.

Consumer credit took another half-century to develop, while computerized systems started to hit banks as early as the 1950s. Anti-discrimination laws in the 1970s (Fair Credit Reporting Act in 1970, the Equal Credit Opportunity Act in 1976, and the Community Reinvestment Act in 1977) actually helped to boost credit score adoption. This legislation had the effect of limiting the scope for discrimination by lenders and requiring them to provide specific reasons for credit denial. Given the technology at the time, this was actually easiest to achieve through algorithmic approaches which allows lenders to make systematic lending decisions in a clearly observable way. We advanced to the FICO model in 1989, which became used for mortgages in 1995, and the VantageScore as a competitor from 2006.

The standard credit model developed around this time was a logistic regression, or logit. This is your basic workhorse model of binary classification to estimate whether a borrower will default or not on a loan. We have a sigmoid relationship between inputs and the probability of default, which provides a nice interpretable model with clear coefficients that can be presented to bank managers, regulators, and customers.

A standard Machine Learning counterpart, by contrast, is the random forest. This segments the data through as series of decision trees. This introduces many more parameters, and so the key trick is to avoid overfitting through a variety of design choices (such as measuring performance out of sample).

The basic tradeoff this then introduces is between a simple model, which is easy to implement and interpret, and a more complicated model which has greater predictive power but is more of a “black box.”

Logit vs the Forest

Fuster, Goldsmith-Pinkham, Ramadorai, and Walther have a really nice paper — “Predictably Unequal” — going through this comparison using US mortgage data. For the task of pure prediction, the ML random forest models deliver modest outperformance (in terms of ROC, Receiver Operating Characteristics, which plot the true positive rate against the false positive rate; and AUC, Area Under the Curve, which measures the area under the ROC curve).

The interesting part, however, shows up in contour plots of predicted default probabilities as a function of FICO score and income. The standard logit produces linear level sets: default rates trade off between FICO and income. The random forest, however, produces non-linear level sets which allows for a more complex representation of the underlying default risk. In particular, the model finds a region where borrowers with low credit scores, but high incomes, actually have less risk than expected; while borrowers with low income and high credit scores are actually pretty high risk.

This matters for fairness because the impacts on credit access of switching to a different statistical technology depends on how the model predicts risk as a function of borrower characteristics, and what the distribution is of those characteristics across groups. This team shows that different racial groups cluster in different parts of the FICO-income space, and so the nonlinear interactions that the ML models pick up help some groups while harming others. While the ML model outperforms in general here, it actually performs worse in a measure of credit racial disparities.

Other problems we have to worry about here include the issue that backtests might look great on historical data, but a new world or regime shifts could drastically change the situation. Of course this was a problem even before ML or AI; but the extent to which new AI methods gain their predictive edge from complex nonlinear interactions may make them brittle to structural shifts in how the world works.

Invisible Primes

So those are some of the downsides. What are the upsides of better model accuracy? A big category has been trying to expand the information set beyond traditional credit bureau files.

Berg, Burg, Gombović, and Puri have a paper on “digital footprints:” the information users leave just by accessing a website. They find this simple and easy to verify information matches the information content from credit bureau scores. The different in default rates between iOS and Android users, for example, is comparable to the difference between a median and 80th percentile credit score. Customers who arrive from a search engine (more likely to be impulse buyers) are more likely to default than those who arrived through price comparison websites. So these digital footprints can help proxy for additional aspects of income, character, or reputation; that “soft information” which credit scoring has been trying to capture for a while.

Di Maggio and Ratnadiwakara examine these trends more directly in fintech lending with alternative data. This platform is able to find “invisible primes”—consumers who have thin credit files and low credit scores, who are classified as high risk under traditional underwriting, but the model is able to actually classify as low risk. We see similar results in another paper here examining LendingClub data. Other work establishes the savings from improved algorithmic modeling might save costs or result in lower required regulatory capital.

We can push this further through AI by expanding the set of inputs and including, unstructured texts. Loan files, narratives, conversations with customers are all fair game now to be processed through LLMs and fed into prediction models. One example of this analyzes CFPB consumer complaints using ChatGPT, and finds this produces valuable signals. For instance, complaints with higher “resolution expectations” associate with higher deposit outflow, suggesting that AI can help triage which complaints associate with bank operational problems.

It’s harder to see so far how AI will further change the underwriting process, but one interesting sign comes in this paper by Gambacorta, Sabatini, and Schiaffi who explore the impacts of AI investments among Italian banks. The key finding is that AI banks are able to lend without a long relationship depth. So traditional banks rely on relationship history to build data and relationship value, while AI banks are apparently able to do a sufficient job of extracting signals that they can give you a loan officer without the soft information from a long-standing relationship. These AI firms also sustained credit supply to firms through Covid, helping firms sustain investment and employment.

AI Generated Risks

The last example, comparing AI banks to non-AI banks, also brings up an important category of AI-generated risks: the competitive dynamics. If a bank has a better model, they can cherry-pick the best borrowers, leaving other banks stuck with lemons. This creates a Red Queen-style arms race problem: you have to keep running to stay in place. Lenders have to keep investing in model improvement, lest they be sniped by an entrant who uses a better model to undercut incumbents for the borrowers where the existing model is most wrong.

But if everyone adopts similar models, you might instead get herding behavior which can amplify systemic risk. This is amplified by recency bias: the models we currently have are disproportionately trained on recent data, and it’s not so clear how they might handle a novel downturn. And as mentioned above, the explainability problem is a real concern. More sophisticated models may perform better, but explaining to a customer or regulator why a loan was denied becomes much harder when the model involves thousands of interacting features.

One way to mitigate some of these issues is a “hybrid” approach in which we use a complicated model to fit the model, but then approximate the complex AI or ML application through a simpler rule. i.e., in the mortgage lending example above, we could simply fit in a non-linear interaction of credit score and income to match the model predictability.

We also spend some time talking about SVB as a case study. Cookson, Fox, Gil-Bazo, Imbet, and Schiller show that banks in the SVB bank run with more exposure to Twitter lost more market value. Interestingly the issue wasn’t negative sentiment per se on Twitter, it was the attention on the social media platform, which created a coordination mechanism for depositors to flee. Other work has argued that misinformation can be created even more cheaply, which might be the basis for an adversarial attack.

This points to a novel class of AI-accelerated risks: liquidity risks on banks through coordinated misinformation or adversarial attacks. Other AI risks include errors from black-box models which propagate through the system, concentration risk on third-party vendors, and the risk of data privacy leakage from model outputs. Naturally, back to Homer Simpson, AI itself is likely to be an important tool in helping to diagnose and mitigate these risks. For example; by helping banks monitor social media chatter to check if depositors appear unusually flighty.

Then we have Knight Capital. In 2012, a software flaw in their trading code led to a $7 billion purchase spree in the first hour of trading, which ultimately led to the failure of one of Wall Street’s largest trading firms.

Thinking in General Equilibrium

It’s common when evaluating AI to think in partial equilibrium: we hold fixed everyone else’s technology, competitive response, regulation, and prices; and just imagine changing one thing. But a world in which AI is freely available will have broad changes in general equilibrium; a world in which prices, competitors, and regulators also move.

There is a wonderful graph of the length of the tax code which appears to hit an acceleration point upwards after the introduction of the typewriter. Making it cheaper to produce text, naturally enough, may have drastically increased the amount of regulatory documents out there. This same dynamic is likely to play out with AI in risk management, which is just an application of Jevons Paradox in our first session. The demands for documentation, model validation runs, compliance analysis, and stress testing are all likely to increase as the cost to produce them falls.

We are still at an early stage here, so it’s hard to assess the scope of these general equilibrium effects so far. We know that better models improve prediction, that alternative data can improve credit access, and that speed and automation can reduce costs. But the feedback loops: what happens to competitive dynamics, organizational responses, regulatory responses, and novel systemic risks resulting from common adoption of AI tools, are really hard to think through, but will likely account for a lot of the pain points in the coming years to get right.

Readings

• We have two cases this week; one on SVB and AI bank runs, and one on Blackrock’s Alladin product.

Subscribe now

This week we apply this to financial document intelligence. The main issue we wrestle with is that LLMs are great at processing text, but by their inherent nature are unreliable at telling the truth.

Documents All the Way Down

Finance is sometimes seen as mathematical in nature, but really it’s about documents. 10-Ks, earnings calls, prospectuses, credit agreements, regulatory documents and filings: the raw material financial analysts have to work with is text. A lot of text. Historically, turning this text into actionable insights has been hard and rate limited by human processing capacity.

This makes financial document intelligence the most immediate and obvious AI use case, which has already seen a lot of adoption. The basic workflow for anything in this space has a few basic steps. First you extract and reconstruct a document base (for an enterprise use case, this may consist of your own proprietary documents). Then you do semantic understanding, which maps the underlying document base into stable concepts.

Next comes the retrieval and reasoning part. This is typically done through something called Retrieval Augmented Generation (RAG). RAG is, at the moment, one of the main AI enterprise use cases. The value of all of this is unlocked by automating and triggering actions, such as triggering an investment review, flagging a covenant violation, or surfacing a fraud risk. For finance applications especially, there is also ideally a reliability and governance layer on top of this to track provenance of claims and controls on deployment.

Customizing AI

To really build on these capabilities, we typically need to do some amount of customization. Pre-training entails building from scratch, and is now the domain of the major AI labs. BloombergGPT was a notable exception here, discussed in class as a case, and since then the industry has typically moved on to customizing general-purpose models rather than training domain-specific ones. The pace of advances of the frontier models has generally outpaced the benefits of training a model in a particular area.

That means fine-tuning a model by tailoring it for a specific task where you need a consistent style or the base model lacks specific domain expertise. Or prompt engineering, which is the main AI customization tool used today. This used to be a really important tool to unlock LLM capabilities, but has been growing less important over time. At this point in the cycle, the most important components here are to provide the right context to the model in solving your problem.

The basic reason this works is that foundation models today already have a lot of financial knowledge baked into them, and so most work in practice is about activating and directing that energy. This paper by Lopez-Lira and Tang shows us how a basic prompt asking GPT-4 to classify headlines as good or bad news works well enough to capture about 90% of initial market reactions.

The Jagged Frontier

One of the critical features of AI capabilities is they are uneven. Dell’Acqua et al. ran a classic field experiment with BCG consultants which introduced the evocative term of the “jagged frontier” to characterize AI abilities. As you’ve probably seen if you’ve played around with AI, it does really well at some tasks while totally missing on others.

This has important implications also for how humans use AI. This team found that consultants working on tasks inside the AI frontier saw large improvements in quality, which were stronger for performers in the bottom-half (i.e., AI helps to level performance, which is what most of this literature has found). For the tasks that fell outside the frontier, AI users did worse than the control group.

Even when we are working with high-quality AI tools, we face important problems in deployment based on how humans interact with them. The Dell’Aqua falling asleep at the wheel paper suggests recruiters collaborating with higher-quality AI tools were less accurate than those with lower-quality AI because they stopped engaging critically. Shen and Tamkin argue that developers using AI assistance complete tasks faster (though not that much faster), while scoring substantially worse in comprehension quizzes.

All of this suggests a few modes of interaction with AI. We have the distinction between “centaur” and “cyborg” forms of interactions. “Centaurs” divide tasks between human and machine. It’s important here for the human to set the context and let the AI execute with clear boundaries for what each side contributes. The “cyborg” model instead blends human and machine work in an iterative fashion, going back and forth. The basic challenge is that AI is going to be most valuable in places where we are already strong enough to spot mistakes, and risks harming our own learning when it skips a necessary cognitive struggle.

Ultimately, what we need to realize as a society is that mass deployment of AI has the capacity to dull the mind as much as processed foods and easy transportation led to an obesity epidemic. We’ll need to figure out the right heuristics, the mental equivalent of working out, to ensure cognitive discipline in a world of AI slop.

Truth and Fiction

This brings us to the the core problem with AI deployment: hallucinations, and the challenges of figuring out truth from fiction.

Ultimately this gets to a philosophical question: do the scaling properties of AI development converge on a shared view of reality, or are the current problems with hallucinations going to be with us for a while because these tools are fundamentally stochastic parrots, doomed to remix their training dataset without converging on anything we would call true understanding?

This leads to one of the deeper lines of thought I’ve seen in the AI space: the Platonic Representation Hypothesis; which argues that models trained on different data and methods ultimately converge to a shared statistical model of reality which corresponds to the real world. There is some recent work which argues this is the pattern we see across scientific models in various domains.

It could be further scaling is all we need to get accurate world models; or it could be that Yann LeCun is right and we need a radically different model architecture. For the time being, however, we are stuck with hallucinations and errors in using AI. Benchmarks on long context retrieval show steep performance degradation when we expand the context, as is necessary to access a broad document corpus. When do these issues impact the ways we use AI?

My dean Bharat Anand (and his co-author Andy Wu) have a nice framework for analyzing this problem across the dimensions of how tacit vs. explicit is the data, and whether the cost of errors is high or low. AI does best when the cost of errors is low and data is pretty explicit. With higher cost of errors, but still pretty explicit data, AI is best at producing work for human verification. If the cost of errors is low but tacit knowledge is high, it’s best to use as a creative catalyst with humans still selecting the final option. Humans shine most when the cost of errors is high and tacit information is also high. Of course, one challenge in thinking through this framework long-term is the ability to write down the tacit information and context so AI can access these long-term appears to be growing over time.

I think here of the fictional essay by Scott Alexander which imagines the limits of the burden of knowledge. Scott describes a world in which scientists, at the limit, spend their whole lives just discovering the limits of knowledge, spending only a brief moment to expand those limits further. Does AI change this, as an ever-living entity unburdened by knowledge, and able to push ever further?

Grounding AI in Reality: RAG

I think so, one day. But we are a long way from that world, and our practical challenge is a lot more mundane. How do we actually solve the hallucination problem in practice, and get AI to give us accurate answers on a specific set of documents without making things up?

As alluded to above, the main solution to this today is Retrieval-Augmented Generation (RAG). The basic idea here is that LLMs working in document retrieval face a challenge of trying to find a needle in a haystack. So instead of letting the model perform next-token prediction using its entire training data, we are going to restrict the search space to a pre-specified corpus of documents and have the model generate answers grounded in specific references to that corpus.

The basic RAG pipeline works in two parts. First there is indexing which entails collecting documents, processing and cleaning the output, and “chunking” them into manageable pieces for future retrieval. Then there is retrieval itself, when the user submits a question, you retrieve the most relevant chunks, pass the question and retrieved context to the LLM, and generate an answer grounded in those references.

An example of this is from my own research with Alex Bartik and Dan Milo. Our objective was to measure housing regulations at scale across thousands of US municipalities. This is exactly the kind of task which was infeasible at scale before LLMs, because you would need large teams of researchers reading each code individually.

We built a RAG pipeline to process municipal codes, splitting each ordinance based on its own hierarchical structure. One thing we see here is that chunks of text within the same article tend to cluster together near other chunks in embedding space, which suggests that geometrical distance follows contextual meaning. This means that the embedding model also places questions that we want to ask near the part of the text relevant to answer the question. This in turn enables drastically higher accuracy at lower cost to actually answer these regulatory questions correctly at scale.

The final accuracy we get to here is about 96% for binary questions and a 0.87 correlation on continuous measures, using the AI methods available at the time. Is that good or bad? For a developer, that’s probably not high enough that you can skip the final verification step. This maybe puts AI in the category where you follow up to verify the output. But for a researcher, this is great: we can take the data and go ahead and run regressions, since we naturally have statistical processes that tolerate error.

Where this is Going

The takeaway here is that financial document intelligence, especially through RAG, is already useful for text-heavy workflows: research synthesis, compliance review, client communications, document Q&A, etc. Tasks that entail working with the AI to draft reports, navigate documents, and explore data are “within the frontier” of AI capabilities. Whereas fully automated decisions which entail high-stakes autonomy, responding to a changing environment, and real-time trading decisions are still (for now) outside that domain.

Should that change your workflow? This brings us back to Amdahl’s law we discussed last time. If the bottleneck in your workflow is document processing itself (or you can adopt a new workflow based on document processing), then RAG-based tools can meaningfully help. If the real problem is elsewhere, than faster document processing may come with real costs and possible errors that don’t actually solve your real problem.

Readings

• Generative Regulatory Measurement

• Generative AI for Economic Research

• We cover two cases this session:

  • There is a really interesting postmortem here of a company called Tract

  • We also discuss S&P’s acquisition of Kensho

Subscribe now

So I’m going to use this newsletter to share some of what I’m developing for this course. I’ve created an open course page here, where you can find a syllabus and lecture notes. On this Substack, I’ll be sharing weekly course summaries. My motivation here is Josh Rauh, who has also opened up his Energy Finance course on Substack, which I’ve found very helpful. I’ll be sharing weekly updates on my main newsletter here over the semester.

This has been a fun experience for me to figure out what is worth teaching about AI and how to do it. The class itself features a lot of case discussions and applications of AI (in assignments and a final project). The content here is going to be more about the conceptual framework and economics.

I start with three principles which I think helps capture where AI helps to add value in finance.

1. Turn Insight into Action

First let’s start out by summarizing what AI does: it builds compressed maps, or lower dimensional representations of reality so we can make decisions. This is really valuable because the world is complicated and we need simplifications in order to take actions.

Borges has a famous short story about an empire which creates a map so detailed it’s the same size as the empire itself, which is eventually abandoned. The lesson here is that maps are valuable precisely because they are more simple than reality (there is some dimensional reduction).

A couple of examples from this: the famous set of lithographs by Picasso showing progressively more abstracted depictions of bulls, and the 1972 Vignelli Subway Map. This subway map in particular tells us that the point of the simpler or more abstracted representation is to help us make a decision: how to get from point A to point B.

LLMs are, in that sense, very complicated map-making machines. They compress their training data and inputs into something which captures basic patterns and relationships. That can be valuable insofar as it helps us to make a decision in the real world.

In Finance specifically, the gap between understanding and action is pretty crucial. The usual benchmarks of AI performance on narrow and specific tasks are less helpful than evaluating them along other dimensions:

• What are the cost-latent-quality tradeoffs?

• What are the costs of inaction vs. the costs of flawed action?

i.e., we need to be more Bayesian in our decision analysis, and think about when a simplified representation facilitates a better decision.

2. Fix the Slow Part

One of the biggest challenges in deploying AI tools, especially for finance applications, is that AI advancement is capped by the slowest link of the system.

This is just Amdahl’s Law, taken from computer science: system speedup is fundamentally limited by the time you actually improve. If you have a process where 90% of the time is spent waiting for compliance review, and you build an AI that makes document drafting 10x faster, that’s going to improve your overall process by maybe 5%.

This is a pretty straightforward idea, but I think it helps to reconcile some of the conflicting stats on productivity (discussed there by Alex Imas). AI can result in really impressive improvements in discrete and specific tasks in the finance workflow, but are those really the main bottlenecks to getting stuff done? It all depends.

This is another drawback of standard benchmarks. We should think instead whether and how AI helps to speed up the really slow and constraining parts of a particular workflow; and what the frictions and barriers are towards that kind of speedup. Often, the constraints there aren’t technical, but regulatory, organizational, or incentive-based.

3. Follow the Price

Automation drives price adjustment. Things that have a near zero marginal cost are going to get very cheap. What are the downstream effects? The answer here depends on substitution and elasticities.

• When something gets cheap, the price of substitutes falls and the price of complements rise. So if AI makes basic financial analysis commodified, the value of having basic financial analysis declines, but the value of things which complement it (private information, relationships, judgment) might increase.

• What happens to the overall size of the pie? If automation improves production, does demand increase in tandem or not? This is Jevons Paradox: improving efficiency, if something is in high demand, can actually lead to more overall consumption.

One example of this is in the graph above—tradable goods have been getting cheaper, relative to services, as productivity advances in more capital intensive-sectors drive growth and output, which then concentrates more rents into other sectors with lower productivity growth. The big thing that’s changing now is that even services are getting disrupted by advanced automation.

My guess is a lot of finance work has relatively elastic demand, and is currently rate-capped by the human complexity required to produce it. Firms might do a lot more due diligence, scenario analysis, and valuation if it was cheaper to produce the analysis. Doing so will wipe out certain tasks which get completely automated, but be an advantage for higher value add analysts.

Of course, the distributional effects could still be brutal here, depending on whether your job is made up of easy to replace tasks or not. But the takeaway here is to follow the price action to determine what gets commodified, and what stays scarce. The challenge is to use the low cost products in your production function, while generating outputs that are more scarce complements to automation.

The Bitter Lesson (and why Finance resisted it)

To understand how these principles apply, it helps to understand how we got here, and why finance was in some ways slow to adapt.

In 2001, the statistician Leo Breiman wrote a famous paper called “Statistical Modeling: The Two Cultures” arguing that the field had broken into two camps. One side assumed the data came from an interpretable model: linear regression, logistic regression, or something you can write down and explain. The other culture treated the data-generation process as fundamentally unknown and used tools like decision trees and neural nets to just find algorithms which worked well at prediction.

At the time, finance was in the first camp, along with 98% of statisticians (as Breiman estimates). There were good reasons for this. We tend to favor simple, interpretable rules. We think they are going to be robust out of sample (less overfitting), and they are easier to explain to regulators and clients. Finance cares about causality, not just prediction (though of course these two things are linked). Our “gold standard” is the RCT or A/B test. We write down a model: the CAPM, Fama-French model, whatever it is, and test it against data while being careful about the biases which show up in estimation.

Then came what Rich Sutton calls the “Bitter Lesson.” The lesson from the last 70 years of AI research, Sutton argues, is that general approaches which leverage improvements in computational ability and declining costs scale up very well and defeat human-designed approaches, over and over again across domains.

You can see this playing out in the history of text analysis in finance. The first generation used simple dictionaries to classify words as positive or negative, count frequencies, and assess sentiment. One of the challenges, as Loughran and McDonald identified, is that finance words are different from other words (“liability” is typically a negative in the world, but can have varying meanings in a 10-K). They showed that three-fourths of words tagged as negative in the commonly used Harvard dictionary are not actually negative when used in a finance context.

So to address that we first used some human intuition to develop a finance-specific dictionary. Then came n-grams (pairs or longer sequences of words like “strong demand” vs “weak demand” which help to disambiguate language). Finally we get to word embeddings, which turn words into vectors where similar meanings are closer together in geometric space. This enabled contextual embeddings, whereby “bank” gets a different representation depending on whether we are talking about a river or finance. Then we get the landmark attention paper, transformers, and LLMs. Tim Lee and Sean Trott have a fantastic overview walking through the basics of how LLMs work, which I highly recommend you check out.

The key point though is that each step involves less human curation and more letting the algorithm just learn from data. The models get bigger, more opaque, and more capable.

The weird thing here was that classical intuitions about model fit turned out to be wrong in important ways, and discounted the levels of emergent ability that have shown up. Traditional statistics says that if your model has more parameters than data points, you’re going to overfit horribly and perform poorly out of sample. But ML models exhibit properties like “double descent” whereby errors rise as you add parameters, but then fall again with improved scaling. These massively overparameterized models generalize and perform better than we thought possible.

This doesn’t mean interpretability and causality don’t matter at all. I just had a discussion on this with Alex Imas in this newsletter all about the ways in which purely AI approaches can and cannot learn from data. We have to be careful about the ways in which generative models have implicit world models that may be very wrong in ways that inhibit performance. At the same time, sometimes prediction is all you need. Just as one illustration, the FT recently published a geopolitical tracker which estimates sentiment across articles, correlates with the traditional dictionary based geopolitical risk approach, but captures a broader context. You could imagine this being useful to track sentiment, manage risk, and potentially even make trades.

Putting this Together

The tension here is that the bitter lesson says to bet on scale, but finance is full of constraints that don’t easily get computed away. So the basic strategy for this course is to figure out 1) how to identify and address the critical bottlenecks; 2) use the improved predictions to make better decisions; and 3) keep our eye on price theory because the resulting surplus is going to accrue to the scarce factors.

That’s it for this week: subscribe if you want to follow along (or unsubscribe if you’re tired of hearing about AI; it’s okay I get fatigued too).

Subscribe now

Readings

• Ethan Mollick has a nice guide on getting started with AI right now.

• Tim Lee and Sean Trott have a great explainer on the basics of LLMs.

• Dario Amodei, the CEO of Anthropic, has a podcast episode with Dwarkesh Patel discussing scaling laws

• The Bitter Lesson, by Rich Sutton

• Case: BloombergGPT. An initial paper; evaluation of different models; and an open source version

This leaves open the question I get the most: what does this really mean for productivity? All of this reallocation seems well and good for workers, even if has some costs for urban centers, but for remote work to really be sticky we want to know the ultimate implications for firms.

The debates on this question are all over the place. Jamie Dimon famously thinks remote work doesn’t work for innovation or management. Companies like Amazon and Google are pushing workers back to the office. Meanwhile, a bit more silently, many other companies are staying remote friendly (here for instance you can see remote job postings at Y Combinator-backed startups).

So what’s going on here? Is remote work good for firms or bad? Are executives right to push for RTO or they optimizing a bit too much for the illusion of control and biasing for face time?

Together with my co-authors Elena Simintzi and Abhinav Gupta we tried to build a comprehensive dataset to answer this question, linking productivity measures from GitHub commits to firm remote working measures and hiring patterns.

In our new paper, we find is remote work is not uniformly good or bad: but the productivity effects vary substantially by age. Startups do well with remote work, while established companies struggle.

Heterogeneous Impacts of Remote Work

To establish our core findings, we ranked workers by their GitHub commits, and did some validation to verify that more commits associate with better firm outcomes. It’s a natural, albeit imperfect, way to try to proxy for white-collar productivity across coders, who are often evaluated and promoted on the basis of their demonstrated commits to company projects (we proxy for these specifically through contributions to restricted repos). We then instrument for remote work status based on how much the firm’s pre-pandemic occupational mix is conducive to remote work, and see whether employees at firms which go remote only because of their occupational suitability see higher or lower productivity.

The key result is that remote work increases productivity at startups (+12 percentage points out of 100; for firms younger than ten years) relative to others, while reducing it for established firms (-9 percentage point decline in worker productivity). Averaged across all firms, the effect of remote work is small and slightly (insignificantly) negative, but this disguises substantial heterogeneity which firm age is a good proxy for.

Why Does This Happen?

We explored many reasons why startups and large firms would differ in how they cope with remote work, but part of the answer turns out to be hiring constraints.

Before remote work was widespread, a key friction for startups was competing for talent against established firms which have bigger brands, more offices, and more established recruiting channels. One way startups responded to these challenges was clustering into expensive business hubs like New York or San Francisco, which allowed them to access many workers with a single location, but at a high cost.

Remote work changes this situation by allowing startups to hire from anywhere (or indeed all over the world), and thereby accessing talent across a range of geographical markets. This relieves frictions to scaling up for firms, which is exactly what we see in the data; remote startups grow much faster (they also post more jobs and see more hiring success per posting).

Larger firms also do see slight increases in growth rates, but they are almost exactly matched by increases in departures as well. These retention challenges for large firms might relate to lower frictions in outside job market search (i.e., it’s easier to take Zoom recruiter calls) or lower value of corporate culture for large remote firms.

About half of the productivity gains for startups seems to be accounted for by the increased ability to hire and scale. We see these productivity benefits materialize for new hires upon joining; they also appear to spillover and increase the productivity of existing team members, too. So the overall pattern suggests better matching of firms to remote startups.

What This Means

In my view, there are a few key implications

1. The RTO Wars Make a Bit More Sense
   When Jamie Dimon or Andy Jassy push for return-to-office, they are likely responding to real productivity and retention challenges faced by large firms. In person work isn’t the only possible solution: just as startups figured out how to make remote work function, there are probably further technological innovations which could make it work better for large firms. But there do appear to be genuine challenges getting large, complex organizations with established cultures and important coordination challenges to adopt remote work effectively.

2. Remote Work Could Boost Business Dynamism
   The U.S. has seen declining business dynamism for decades, a trend which seems to have reversed since the pandemic. Even if remote work doesn’t stick around for the entire economy, being a remote-first startup may be an important new technology which tackles a key barriers young firms face in accessing talent and competing against incumbent firms. So it could be an important new bootstrapping tool helping entrepreneurial entry, even if most JPMorgan employees are back in the office.

3. Innovation Might be Less Geographically Concentrated
   Our results should be a bit reassuring in terms of economic geography, because many regions wind up winning. Large cities with established firms have a bit of a relief in terms of remote work and office demand, because the large firms they host may ultimately need to get back to the office. Meanwhile, startups and other remote firms can hire remote workers outside of major job hubs, which helps to disperse innovation and economic activity outside of superstar cities. This helps to alleviate some of the issues we have seen after decades of consolidation of knowledge-intensive work in particular in these hubs.

The basic takeaway here is that remote work is a technology with important tradeoffs. It seems to come with important productivity challenges for many firms, while changing drastically where your employees need to live, which winds up impacting firms very differently across their life-cycle.

Subscribe now

So what happened? Expectations of inflation made workers and firms change their behavior: knowing that inflation was going to be higher in the future, workers demanded higher wages and firms hired fewer workers today. It turns out that a statistical correlation between lower unemployment and higher inflation is not exploitable by policymakers as an intervention, because their policy action induced a different set of behavior by agents.

Enter Robert Lucas with his famous 1976 treatise “Econometric Policy Evaluation: A Critique.” This Lucas Critique, as it came to be known, argued that economists had been interpreting correlational relationships from historical data as structural, meaning they were invariant to policy changes. But they weren’t. The economic agents which generated the data may change their behavior in reaction to changes in policy, which—as the Phillips curve example showed—can shift the observed relationship between variables. The Lucas Critique essentially argued that reduced form methods were only able to estimate the local equilibrium and hence would not be able to predict outcomes to a policy shift.

His suggestion was to use structural models of the economy as these will better predict responses to policy changes. Specifically, if one microfounds the structure of people’s preferences and firms’ objective functions and scales up, then it is possible to predict the impact of policy because the model will take agents’ responses to it into account. This Critique had an explosive impact on economics: the “microfoundation revolution” resulted in a complete shift in how macroeconomics was done. Out were the reduced form regressions, in came the representative agent models.

But the structural approach comes with costs. For one, the predictions will only be as good as the model. If the models were misspecified or the exclusion restrictions were “arbitrary and incredible,” then the predictions were not going to be very good either.

Transformer models and Learning Causal Structure

In this post we revisit the target of the original critique: the model estimates from running fairly simple correlations did not take into account the structure of the economy, and therefore make the wrong predictions when policy shifted. Now consider transformer models. In principle, there is little in the architecture of such models to suggest that they can learn the structure of the data generating process (DGP) from seeing the data. But recent work has shown that transformer models do seem to learn DGPs–-encoding the causal structure during training—and are robust to distributional shifts at least in the case of “nearby” DGPs (e.g., models of the same class). Recent work by Ben Manning, Kehang Zhu, and John Horton shows how transformer models can automatically propose structural causal models and test their implications through in-silico experiments with LLM-based agents.

This emergent property suggests that transformer-based LLMs may, in principle, be able to make predictions for policy shifts in economic settings. How would one test this? One test was suggested by Pontus Rendahl: take a transformer model and train it on economic data simulated by a known structural model. Will the model be able to predict responses to policy shifts?

So that’s what we did. We took the New Keynesian (NK) model and simulated a bunch of data. We then trained a transformer on it, took a hold out policy regime and predicted the response. How well does the transformer model predict the policy response?

Before getting to the results it’s worth clarifying what this experiment is doing, as well as its limits. The question it’s answering is the following: let’s assume that the real model of the economy is NK. The transformer model sees data from this economy and predicts a policy response. Like the econometrician, it has the history until time t and must forecast the state of the economy in period t+1. How well does it do? If it does well, what does this mean? It means that given the underlying model of the actual economy (which may or may not be NK), the transformer will be able to “learn” the DGP at least well enough to make well calibrated predictions. Here is what the experiment cannot answer: 1) will the predictions change if the relationships between variables change, 2) does the model actually encode the DGP in a way that can be elicited (research by Keyon Vafa and others makes us pessimistic on this), 3) are the predictions robust to extreme policy shifts? Beyond these limitations, the obvious one is welfare analysis: the black-box nature of the transformer model prevents the researcher from saying much about welfare given various counterfactuals, which the structural approach allows for by design. But here is one aspect of the exercise that we do not view as a limitation: the NK model has a fairly simple structure—perhaps the transformer model is able to make predictions from data simulated in this framework, but it will do a lot worse if the structure is more complex. While this would certainly be a limitation for the transformer-based approach, the same critique can be applied to the structural approach. Specifically, if the NK model is too simple to generalize from the transformer approach, it is also too simple to model the economy.

Transformers tracks the NK model well

Ok, with these qualifications out of the way (and we’re sure to be missing some), here are three sets of results.

1. The Transformer tracks the NK model’s realized dynamics across the three key observables: output gap, inflation, and interest rate.

To evaluate model fit, we picked some out-of-sample parameters which the model never saw in training. We then simulate a NK model under these parameters for 50 periods, for which the key three observable variables are the output gap, inflation, and the interest rate. We then give the transformer the parameter vector and innovation, and ask it to forecast the endogenous economic variables, plotting the truth (solid line) against the transformer (dotted line).

The transformer is basically on top of the truth, suggesting that it is able to match the dynamics of the real system quite closely. It even manages to get the turning points and large deviations right, and matches the levels quite closely as well. The main deviations are at the extreme peaks and drops, where it slightly underpredicts the magnitude of the shift.

By itself, this evidence doesn’t establish that we can just switch over to transformers instead of solving DSGE-style NK models. The objective is a little more limited: that the transformer model appears to do well in out-of-sample forecasting in a simulated NK world.

2. The transformer model has some success in forecasting in response to policy shocks.

This is really the central question for evaluating success at the Lucas critique: how well does a model trained in one set of parameters match the counterfactual economy’s trajectory after being hit by a different set of shocks? That is, how well can a transformer model predict policy counterfactuals from a different regime?

We test this by drawing the holdout parameters, comparing across three types of shocks (natural rate, cost push, and policy), and plot the true impulse response functions (solid line) against the transformer prediction (dotted line). The test is whether the transformer, having been trained in one regime, can accurately trace out the effects of a clean shock out of sample.

We find the transformer generally gets the sign correct in the short-run, and is close in magnitude in the long-run. Ie, if a policy shock raises the output gap and lowers the interest rate, the transformer generally accurately predicts the direction of the effect at onset and gives us a rough magnitude.

One place the transformer model does more poorly is in estimating the dynamics of the impulse-response. The transformer exhibits a bit too much oscillation and overshooting, including sign-shifts.

This suggests that the model has not completely “learned” the true state of the economy in a structural sense, and therefore cannot provide a strictly accurate guide to causal effects.

However, how much do these effects matter in practice?

Theory versus practice

In a famous essay, Milton Friedman argued that the key to evaluating the methodology of an economic theory lies not in the realism of its assumptions, but purely in the predictive power of its conclusions. He drew the analogy to modeling pool players assuming they have access to the accurate mathematical formulas determining billiard ball trajectories. In reality pool players use heuristics and mental shortcuts, but this assumption might nonetheless be accurate in predicting shots by expert pool players, who have converged closer to the prediction of an optimal physical formula.

Of course, the intuitions from the two UChicago economists Lucas and Friedman are at complete odds in terms of practice. If the only thing that matters is prediction, then we can afford to be quite agnostic about the precise data generating process as long as we are happy with the model fit. But if the main thing that matters for causal inference is having the “right” representation of the state of the economy, then the predictions alone are not good enough and we really need to nail down the true economic model.

The results from our experiment suggest that transformers sit somewhere between these two extremes. The fact that a transformer can take a new policy regime and a clean impulse shock and still get the sign at impact and rough magnitude is already a non-trivial achievement that advances beyond simple reduced form correlations. However, the volatile dynamics are a reminder that getting the rough direction correct doesn’t necessarily entail learning the true structure. Whether this matters or not depends a bit on the application.

The Friedman-friendly interpretation, in other words, is that transformers do seem to learn some useful and regime-transferable predictive technologies which enables a limited degree of extrapolation out of context. Despite not having “learned” the true state of the economy, the model has internalized some regularities which allow it to approximate the conclusions we care about.

The Lucas critique does still bind here, however, because the model has not yet learned the true state representation, and hence the true nature of shock propagation in the economy. The shape of the IRFs can be wrong even if the endpoints are correct.

Moving beyond the Lucas Critique

3. Transformer models represent a huge advance on the purely reduced form models criticized by Lucas.

Another natural benchmark to consider the success of the transformer models is through a direct comparison of the Cowles-style regressions Lucas was criticizing. We take our same simulated NK data and estimate a linear reduced form dynamic system in which the output is predicted using lags of historical variables.

We observe substantially worse model fit, as judged by comparing the true data (solid line) to a reduced form prediction (dotted line). The mean square error is about an order of magnitude higher in this predictive exercise than in our transformer.

When we look at impulse responses, the reduced form model also really struggles to get the magnitude of the impact right, especially at inception. Generally the reduced form model drastically overestimates the true magnitude of the effect.

You can see based on these results where Lucas was coming from: if the economy genuinely did follow an NK structure, you would be pretty far off in estimating true policy counterfactuals through a purely data centered approach. Opting for the structural route was a reasonable way to try to improve policy credibility with the technology available at the time—so long as one was prepared to believe the implied structural model of the economy.

The Lucas critique logic does hold today as much as it did in the 1970s. What has changed is the technology of approximation.

When Lucas made his critique, he was responding to an econometrics which was essentially linear, based on a small number of parameters, and reliant on strong identifying assumptions. The toolkit consisted of lagged variables, VARs, and some simultaneous-equation systems which had a hard time representing the true, non-linear, and forward-looking nature of the economy. It’s easy to see why such models would struggle at getting out-of-sample policy forecasting right.

Transformers, and the new AI machinery around neural nets more broadly, represent a huge advance in this kind of representation. With much deeper context windows, these models can keep track of long and complicated histories and do a much better job at prediction. If the model can infer something close enough to the true hidden state of the economy from observable variables, it can transfer that knowledge to nearby policy regime shift states because the underlying economic logic is the same even as the states change. This is what we see in our forecasting plots and (somewhat imperfectly) in the IRF graphs: the model is learning something like a state-space model, at least enough to forecast, without knowing the true state-space.

So what do we conclude from all of this?

First, that advances in artificial intelligence do weaken the practical impact of the Lucas critique relative to the models criticized at the time. Newer transformer models can learn predictive relationships which remain partially stable across regimes, at least holding fixed the true structural nature of the economy and for some “nearby” policy shifts. This is something already that Cowles-style models struggled to do.

However, simple transformers alone do not abolish the Lucas critique. If the goal is accurate counterfactuals for policy, rough approximations aren’t enough. You probably want greater assurance that the model’s internal representation is close enough to reality so as to be comfortable relying on the model for guidance. Of course, the same challenge applies equally to the DSGE objects in modern macro, which have had mixed success in forecasting and prediction.

All of this suggests a natural research agenda going forward. Structural models have their place in economics, and will retain important advantages in terms of legibility, being able to cleanly trace mechanisms, and evaluate welfare and counterfactuals. But we should be increasingly willing to explore transformer-style models and other tools from ML and AI to explore purely data-driven model generation in our field. As this technology is likely to just get better over time, we should grow more comfortable in thinking of “structure” as something which can be learned, rather than assumed.

Edit: See here for underlying code generating the transformer model; and here for the Cowles-style reduced form benchmark.

Edit 2: In response to this post, we got some good comments by Olivier Wang and Nathaniel Bechhofer, among many others. Our design here was intended to be as simple as possible to illustrate the use of transformers, but there are two potential issues with the comparison. One, the transformer model does have access to the parameter vector and innovations. Note that while this helps the model understand structure, the model does not assume linearity in the underlying DGP, so there is still a non-trivial step mapping the parameters and latest data point to a forecast of the next period. Second, the linearized NK model is already a pretty good fit for the VAR approach, so it is hard to improve on that baseline too much in this case.

To address some of these issues, we created another experiment in which we have a causal transformer as before, but it observes only the y variables, i.e. it has none of the benefits of observing structure. We compare this to a Kalman filter, which is a reduced form model with a similar handicap of not inherently matching up to the precise design of the NK experiment. We then show the MSEs of the two approaches below, which show the transformer (TF) substantially outperforms the Kalman filter.

To be clear, we are not asserting that the (simple) transformer model here is the last word—there are obviously many other implementations of these methods which can do better. Nor will the results necessarily extrapolate to other economic frameworks, or to areas in which the economic regime changes entirely (though we are not so confident our existing DSGE methods work for such drastic changes either!). Our goal here is pretty modest: we are just highlighting the improved predictive power of transformer models relative to the technology available in the 1970s in the specific framework of the NK economy, which has some relevance in thinking about the bite of the Lucas critique in shaping the research we do.

On housing, we could talk a little bit about Wall Street speculators buying up housing and making them less affordable.

This has led to proposed legislation targeting these investors, based on the idea they exercise market power and are bidding up house prices and rents. In some housing circles, the role of these investors (alongside the role of algorithms for price setting) are really viewed as the central drivers of housing unaffordability.

Are these reasonable views from an economist perspective? At first blush, the fact that these investors still own such a small fraction of the overall stock of single-family homes would suggest their impact is still pretty minor. But what accounts for the entry of these investors, and is their impact so far good or bad?

I have a great PhD student on the job market this year, Josh Coven, who has written a topical and well-executed Job Market Paper which sheds light on these issues, which I’ll summarize below.

The Role of Housing Scale

The first thing to note about the rental landscape, especially the single-family rental world, is that it is pretty fragmented — you have a lot of mom-and-pop landlords who operate at a small scale. This started to change after the 2008 financial crisis, with the entry of large-scale landlords like Blackstone. At first, these investor groups targeted areas with high foreclosure rates, which allowed them to amass large portfolios at deep discounts.

The key difference with these new large-scale investors is that new rental management technologies, financial instruments, and bargaining power allow Wall Street investors to operate with lower costs in levels and also exhibit much better economies of scale relative to the mom-and-pops.

Josh shows that large investors enjoy cost advantages by paying fewer operating expenses, property taxes, and lower mortgage costs. One particularly interesting one is insurance costs, which have of course been rising all over the country. Large landlords appear to enjoy bargaining advantages and pay substantially lower insurance expenses as well.

The only way that many small landlords can remain cost competitive is by avoiding mortgages and outside management companies entirely. However, this inherently limits the scale at which these landlords can operate, and these strategies are also pretty time-intensive. Those smaller landlords which do try to scale up (by taking out mortgages, etc.) quickly hit profitability barriers, and in fact Josh measures a surprising fraction of such landlords appear not to be profitable at all.

Professionalizing management, better bargaining over expenses, some degree of appeals over property tax, and better access to capital markets therefore enable institutional investors to improve their cost structure, which results in their ability to provide many rentals at scale. They are also able to maintain these lower costs even as the number of properties they manage goes up.

Market Power of Institutional Landlords

However, another consequence of this greater scale is large landlords do appear to have some market power, as they are the sole residual source of additional rental properties. So in the basic supply and demand graph above: large landlords have a flatter and lower supply curve (due to lower costs overall and flat marginal costs), which moves the housing market equilibrium from A to B. However, if they optimally exercise market power on the supply they provide above what small landlords can, the markets wind up at an equilibrium point C instead.

So both effects are going on here in this stylized model: institutional landlords are more efficient, and their ability to provide more rental properties lowers rents and increases rental availability. However, they take a fraction of the gains back as market power. Notably, renters still benefit overall from Wall Street investors, as rental availability and rents are still higher. One thing that’s going on here is that Wall Street investors help to arbitrage around financial frictions: many renters lack the down payments or credit scores to buy on their own, so they rent instead from an investor company who can buy and is renting out to them.

Why Do People Care About Speculators?

I think this gives you a sense of the key economics of the paper, though of course there is a lot more in there. Josh quantifies the intuition from the stylized setup with a much more detailed model featuring BLP estimation on the housing demand side, and a careful consideration of landlords as well. There are a number of important frictions to keep track of: there is a whole construction side of the model, in which builders respond to price signals, and a margin of adjustment between small and large landlords. All of these are important to get the quantification of investor impact right.

One important tradeoff in the model results is that institutional investors seem to improve outcomes for renters — who have better access to “high opportunity” neighborhoods characterized by better economic mobility and test scores, as well as lower rents overall — against higher prices for houses which make it harder for prospective homeowners.

But I want to focus on two key implications which get at the broader public debate over these investors. The first is about why people seem to care so much about these investors.

One key set of results Josh gets (Figure 7) is the relationship between institutional investor shares and prices/rents; both as observed in the data, compared to what the model predicts.

Essentially, the model suggests that house prices do rise a bit when Wall Street investors enter, but the effect is fairly modest. Two key reasons for this are 1) housing supply responds a bit to institutional investor entry, and 2) investors partially crowd out mom-and-pop landlords. Both of these effects mean that when institutional investors buy homes, they don’t reduce the availability for owners 1-1, and so the price effects are muted as well. And, of course, the model actually predicts a decline in rents.

By contrast, you see that house prices and rents went up quite a bit, in actuality, in areas where Wall Street landlords entered. So to the casual observer, it looks as if the entry of institutional investors is driving really enormous price action.

What is happening here, basically, is the standard economic story of selection: consistent with Josh’s modeling of them, institutional investors are picking areas likely to see future rental growth. Therefore, even if their own actions slightly lower rental pressures, it’s going to look as if they correlate with higher house prices and rents. It’s sort of a similar situation as housing supply skeptics who don’t believe that new housing supply can reduce rents due to some sort of folk economics-OLS.

The Role of Housing Supply

The discussion above suggests that housing supply is a key part of how we assess the impact of institutional investors. Indeed, as many have recognized, landlords themselves emphasize these factors, ie Blackstone’s Jonathan Gray:

The one benefit to existing owners is that construction lending is getting tighter. I think we will see less new supply and long-term that's a positive… Fundamentals around supply and demand are what drives value. That's why I have confidence going forward

Josh simulates some of these changes in counterfactual simulations. A world without housing supply responses would feature even higher price impact from the entry of institutional investors. By the same token, policymakers concerned about the impact of these investors on homeownership and house prices can increase housing supply, which strongly mitigates those consequences while preserving the benefits of these investors for renters.

Of course, another way to mitigate these impacts is for institutional investors to build housing themselves. This kind of build-to-rent market seems to be growing quite quickly, and can be another important way to address housing shortfalls without crowding out homeownership.

I hope this gives you a flavor for the paper — again, there is a lot of other stuff there, so be sure to check it out (and interview Josh!).

There are two main ways researchers tried to measure housing regulations. The “Wharton” approach, led by people like Joe Gyourko, tried to get a sense of the nationwide variation in housing regulations by sending surveys to planners around the country. This has the benefit of a nationwide scope, and has been an incredibly valuable resource for researchers, but does not always give you the granular understanding you might want.

The second approach, the “Harvard” approach led by folks like Ed Glaeser and Jenny Schuetz, tried to instead do a deep dive into housing regulations in specific areas through the Pioneer Institute. This allows for a much more precise and accurate classification of housing regulations, but is limited by the cost and expense of scaling up the approach, and was only done for 187 municipalities in the Greater Boston Area.

Together with Alex Bartik and Dan Milo, we release a heavily revised version of our paper on Generative Regulatory Measurement, along with associated code and data, which argues that LLMs are capable enough to solve this problem of accurate regulatory measurement at scale, the benefits of Harvard granularity at Wharton scale. We’ve put together a publicly available AI-generated national housing regulatory dataset and look forward to your feedback.

Generating Housing Regulations

If you upload a municipal zoning document to ChatGPT and ask some questions about housing regulations, you’ll notice that LLMs are not always accurate at the task of reading through large complicated documents and providing the right answer. Models are improving rapidly, but at this point I think it’s fair to say they do a much better job of accurately categorizing the content of a single sentence, rather than consistently picking that right sentence out of a large document. This naturally makes it hard to consistently and accurately parse regulations from lengthy documents.

We address this problem through RAG — retrieval-augmented generation, which has become standard in computer science and is just getting started in economics. The basic idea is to “embed” regulatory documents, which entails applying a mapping which transforms the document into a vector. You can think of a simple embedding as a vector the size of the number of words in the English language, which simply counts the frequency of each word in the document. Modern embeddings are more sophisticated, and vectors that are close in embedding space are more similar in meaning.

This allows you to compare the embeddings of municipal zoning documents against the embeddings of our regulatory questions, and then figure out where the answers likely lie. To get a list of regulatory questions, and put together a validation dataset, we start with the Pioneer Institute’s set of questions and answers from 20 years ago. Applying RAG, and some other tricks outlined in the paper, we think we can get pretty accurate regulatory classifications: binary questions line up between the LLM and human answers 96% of the time, while we estimate minimum lot sizes against ground truth with a 0.92 correlation on our validation sample.

These numbers are likely to only get better over time, as LLM models improve along with our ability to use them. This suggests that we can likely use LLMs to go through large chunks of textual documents — regulations, court cases, earnings call transcripts, newspapers, etc. — And categorize not just sentiment or vibes, but also their semantic content.

We do other checks in the paper, including making sure the results don’t just hold in specific geographic areas, and then scale up by collecting a large sample of municipal zoning documents to generate a national housing regulatory dataset.

What we learn about housing regulations

In the paper, we analyze this housing regulatory dataset to establish five facts about American zoning:

1. US housing production disproportionately happens in unincorporated areas

2. Housing regulation limits density. This happens through:

   1. Bans on building or converting to multifamily buildings

   2. Prevalence of single-family zoning

   3. Minimum lot size requirements in single family areas

3. Zoning is more restrictive of density in suburban areas, particularly in the Northeast

4. Housing regulations can be organized into two principal components. The first associates with higher housing demand and is associated with rent extraction in Blue cities

5. The second regulatory PC captures exclusionary zoning

You can read the paper for the rest, but I want to focus here on the last two points about rent extraction and exclusionary zoning; which were two big things I learned from working on this.

Rent Extraction in Blue Metros

We often think of housing regulations as a force associated with housing supply. Which they are! But we also find housing regulations seem to associate with housing demand. Specifically, areas that are both expensive and have relatively high housing production have a distinctive set of housing regulations adapted to environments of extracting value.

Some of these regulations are administrative and process related: like more zoning districts and process requirements to build housing. But another typical example is mandates or incentives to build affordable housing in the form of inclusionary zoning. This is a form of redistribution of value away from developers in these areas to some residents.

Areas scoring high on this dimension share some common characteristics:

• They tend to have a higher proportion of college-educated residents

• They have higher job density

• They have lower poverty rates

• They have a substantially higher share of Democratic voters

These areas are also more likely to have retail outlets like apparel stores and dining establishments, as well as professional services including educational institutions and healthcare facilities. In contrast, they have fewer gas stations, utility services, and truck transportation businesses.

This is all consistent with many YIMBYs emphasizing standardization of process and and simplification. These do seem to be key issues in high demand cities; which otherwise actually allow density and building more than other areas.

Exclusionary Zoning

We also identify a factor which looks a lot like exclusionary zoning — suburban areas which have high minimum lot size requirements, frontage requirements, and public hearings requirements for housing. These are areas that strongly limit density and apply minimum quantity standards for housing, which have the effect of limiting entry by low-income residents, thereby exacerbating income and racial segregation. I think we all know exclusionary zoning when we see out: our goal here is to try to measure it more precisely, and connect with other outcomes.

One of the features of exclusionary zoning which really surprised me are the associations with educational outcomes: test scores are better in areas with exclusionary zoning, spending per pupil is a lot higher, as are Chetty opportunity measures — these are good places for social mobility.

I think one way to rationalize this set of regulations is the decentralized nature of school funding and administration. In many developed countries around the world, there don’t seem to be massive differences in the funding or curriculum of public schools. As a result, you don’t often hear of people in, say, Italy or France moving to specific neighborhoods to access local schools. School quality is, at least compared to the US, relatively homogenous, at least at the within-city level (there do seem to be large regional variations in some of these countries).

As a result, with less scope for educational sorting, you typically see rich people live in the center of the city. And some of these people in the city center, in places like Madrid or Buenos Aires, vote for the right wing party. This basically never happens in the US, where the rich have decamped in many cities, and city centers are invariably left-wing strongholds.

So what explains this differential sorting of people and partisanship across countries?

We can’t prove it in our data conclusively, but I think a very plausible hypothesis is that tying school quality to local conditions generates incentives for rich people to decamp for suburban areas, exclude the poor through high bulk regulatory barriers, and thereby produce enclaves of high quality educational.

On the one hand, you might think that this is the sort of “fiscal zoning” advocated by people like Fischel and Hamilton. However, it is a pretty inefficient way to get there — we have to leave large tracts of land less developed to achieve this exclusion — and it results in other undesirable features of segregation in cities.

The Midwest has some exclusive suburban enclaves — but it turns out the Northeast is really where these bulk regulations really matter. By contrast, the West regulates housing more through process. You can see some of the associations here; because municipalities typically report a range of lot size requirements, we focus on the minimum binding one:

We encourage you to use our data and see what else you can find; suggest any other housing regulatory questions you’d like us to add to our dataset, and adapt the approach to categorize other types of regulations and unstructured texts.

Subscribe now

I have a new paper with Josh Coven, Sebastian Golder, and Abdoulaye Ndiaye on property taxes which sheds some light on what’s going on here, and points to some policy tools to address them using traditional property tax instruments.

The Housing Mismatch Problem

The basic issue here is that a majority of bedrooms in owner-occupied homes are owned by people between 50 and 70 years old—empty nesters who often have more space than they need. Meanwhile, younger families with children frequently find themselves in crowded living situations, unable to afford homes with adequate space.

This is also highlighted in Redfin research, which finds empty nesters own 28% of large homes, while millennials with children own just 14%.

What this means for the typical lifecycle for the American family is that households are struggling with cramped living space as they start families in their 20s or 30s. Eventually, many households accumulate the down payment (or else are given them as a gift) to buy their own place. But they are often forced to choose between a large suburban house in a worse job market; or cramped living quarters in better job markets. These pressures are probably responsible, on some margin, for rising ages of marriage and childbearing; greater cohabitation with parents; and declining fertility rates, which are a big macro concern these days.

Households do accumulate more and more housing over the lifecycle; but this means that they are left aging in place in large, spacious homes. In our model, households accumulate these assets due to a bequest motive to pass them along to their children. However, the collective consequence of everyone investing in homes for intergenerational purposes pushes up the price of housing, and results in a lot of empty bedrooms in the hands of the elderly. By the time owners pass away, in their 70s or 80s; the next generation are frequently empty nesters themselves in their 50s. It also makes life harder for aging populations, which may lack a local younger workforce for essential care services.

These challenges are all amplified in high housing cost markets, like California, where you see extremely low ownership rates among younger households:

The “Forced Mortgage” Mechanism

So what’s going on in California? The central force, we argue in the paper, is low property tax rates resulting from Proposition 13. The key insight is that property taxes function as a kind of a “forced mortgage.” Higher property taxes tend to lower the upfront purchase price of homes through capitalization. Essentially, future tax obligations get priced in to the current home value. Property taxes therefore result in a lower initial out of pocket equity investment, as well higher ongoing costs in the form of annual property tax payments, which is equivalent to what mortgages typically do.

The lower prices attract young buyers, as they make down payment requirements more attainable for young, cash-constrained buyers. The higher user cost of owning housing pushes out aging empty nesters and encourages them to downsize. Overall, this tradeoff from higher property taxes is particularly beneficial for young families who have steady incomes to cover the ongoing payments but struggle to save the large lump sum needed for a traditional down payment in expensive housing markets.

We show that, across the United States, spatial variation in property tax rates is generally consistent with this story going on. Areas with higher property tax rates feature:

• More young homeowners

• Fewer empty bedrooms

• A higher percentage of children in the population

• Lower house prices and price-to-rent ratios

All of which are consistent with the idea that property taxes play a role in shaping housing allocation across generations. This is in contrast to a standard intuition you might hear from, say, cranky homeowners in New Jersey or Connecticut, who insist their property taxes amount to a large cost and burden. While this may be true for such homeowners now; these same cost burdens likely facilitated their entry into that housing stock in the first place. Meanwhile, hopeful homeowners in, say, San Francisco are looking at local housing inventory that is mostly locked up and priced out of range for them, and so often depart to other states to buy homes.

California vs. Texas: A Natural Experiment

To further quantify these effects, we build a detailed structural economic model calibrated to match the housing markets of California, as an example of a high property tax state and Texas, which is a classic example of a high property tax state (has no state tax instead).

The model successfully replicates key features of both housing markets, including the steeper relationship between age and homeownership in California compared to Texas.

Using their model, we then run a policy counterfactual: What if California raised its property taxes to match Texas levels? We find:

• Overall homeownership in California would increase by 4.6%

• Young household homeownership would jump by 7.4%

• House prices would fall by about 18% due to capitalization

• In-migration to California would increase, especially among younger households

• Out-migration from California would decrease

• Total wealth would decline (due to lower house values), but incomes would rise slightly as people gain access to California's higher-wage job market

Policies like California's Proposition 13, intended to protect homeowners from rising taxes, may therefore have had the unintended effect of making it harder for young families to enter the housing market. Homeownership for the young would be considerably more attainable under higher California property taxes.

Implementing Georgist Solutions

Our paper is positive rather than normative, in that it focuses on the role for asset taxes like property taxes on equilibrium prices and quantities in the presence of financial constraints but doesn’t take a stance on the desirability. It’s a close cousin of other work which has highlighted the efficiency benefits of wealth taxes; the role for depreciating licenses and Harberger taxes; and the value of dividend rather than corporate taxes for corporate taxes; and of course the Henry George land value tax tradition broadly.

All of these papers have a similar flavor: identifying some market inefficiency, and focusing taxes on correcting this distortion in a way that raises as much money as possible with as few costs.

But I think it’s also worth thinking about the implementation challenges to making something like this happen, should policymakers want them to. On the one hand, there is a clear constituency of possible supporters — the young should favor this policy, as they are prime beneficiaries. It offers them a way to both tap into the increased wealth in real estate generated over the last few years, while also easing on affordability burdens into the future. Of course, there is a just as motivated a set of opponents in the form of existing owners.

The key aspect of political economy which seems to help property taxes in the US is the tie to the funding of local public goods. Property taxes make up something like half of local government revenue in the US, which is actually substantially higher than rates in Europe. Many European countries focus their real estate taxes on transactions based taxes, which instead further increases lock-in.

In California, the key issue, as Jake Krimmel has argued, was school finance equalization, removing local control. The anti-property tax backlash emerged afterwards, and then municipalities saw new construction would give them additional liabilities (school and infrastructure expenses), without a corresponding property tax stream, and they turned to exclusionary housing regulation to keep out newcomers.

By contrast, many Sunbelt states have value capture property tax instruments (MUDs in Texas, SIDs in Nevada, and CFDs in Arizona) which use property tax revenue in the future to pay off municipal bonds issued to build infrastructure today.

So one possibility would be to innovate on the use of property tax revenue to help finance necessary investments today (maybe in childcare, eldercare, mass transit, the green transition, or lowering other local taxes); which might build a broader political economy coalition to allow for more taxes on land value.

Elsewhere

Vital City has a great new issue out on the prospects for an “urban doom loop,” a term that has come out of our research on the impacts of work-from-home and the pandemic on cities. Together with Stijn Van Nieuwerburgh, we have one essay evaluating where we stand on the urban doom loop thesis; and another one analyzing the viability of office to residential conversions to tackle urban decay.

Typically, rapid transit is grade separated from other uses, often underground (though sometimes elevated), with frequent stop spacing. The lines typically converge in a central business district, though often meeting at distinct points to allow interchange between urban corridors. Think the MTA in New York City or the German U-Bahn. These systems facilitate commuting and tourist trips within dense urban areas.

Suburban rail systems, as the name suggests, serve the suburban towns and hinterlands around major cities. They are sometimes not grade separated, and often run alongside or share tracks with other rail lines, with larger stop spacing. The best run systems fan out from suburban areas and concentrate in a urban core, where they provide rapid transit-like services, and fanning out again on the other side of the city. ie, they use through-running of rail operations. Think New Jersey Transit or Metro North in the New York area (though these lines don’t through-run), or the German S-Bahn. There are other nuances and complications here (ie, China has largely avoided traditional suburban rail, instead choosing to run metro lines at longer lengths into suburban areas) but I think this is broadly reasonable summary.

This post describes an exciting new layer of rail services for urban areas, what I’ll call rapid regional rail. This is primarily taking off in Europe and Asia, though it would be ideal if we could get it to work in the US.

What is Rapid Regional Rail?

Rapid regional rail is basically an extension of commuter rail services to towns and regions even more far-flung from the urban core. They typically feature substantially faster trains (max speeds of at least 140 km/hr or 87 mi/hr), with fewer stops across much greater distances (75-100km); across a mix of above and underground routes. The function is to reduce the length of commutes into urban areas to greatly expand the travelsheds around cities and increase their effective size.

We can see this in one of the first rapid regional rail systems — the Réseau Express Régional (RER) in Paris. RER A for instance is an east-to-west line stitching together suburbs far to the east and west of Paris along a 109 km line, serving over 1.2 million passenger trips a day. Other lines are even longer — 172 km for the D line and 187 km for the C line, with speeds reaching 140 km/hr. Like metro lines, these converge in a triangle at the city, which helps to avoid overcrowding any one station.

Why Rapid Regional Rail

So why does it matter? Well, cities have always grown in conjunction with the availability of transportation options. The growth of subways and metros boosted the population of major cities like London or New York in the 19th and early 20th centuries. Adding highways and suburban rail further grew cities into metropolitan areas, with the now distinctive core-suburban separation, especially after WWII.

Rapid regional rail basically helps cities scale in the next transition: towards even larger megapolis regions with huge populations over large scales. City sizes around the world have hit new dimensions in both physical size as well as population. The economic agglomerative forces of having this many people live in the same area and participate in common job markets, dating markets, etc. can therefore boost idea generation and economic output.

Connecting many people in the same area can also relieve housing pressure on the center urban areas — by dramatically expanding the set of houses that lie within a normal commuting distance to the urban core to include the peripheral towns at greater distances from cities.

It’s for this reason we are now seeing rapid regional rail systems take off in the rest of the world.

The New Wave of Rapid Regional Rail Systems

One example of the new rapid regional rail is Crossrail in London, which provides rapid commuter rail coverage — connecting two major east-to-west mainline train lines — with a rapid transit metro link in the middle (Elizabeth line). The line carries over 200m annual rides, many of which are “new” to the entire transit system.

There is another line in Istanbul: M34 or HızRay (Fast Rail), which is about to start construction and would connect 13 stations across a 75 km length around the Istanbul area with 140 km/hr trains.

In Copenhagen/Malmö you have the Öresund Metro — a plan which would link Copenhagen and Malmö with a driverless metro line with a top speed of 120km.

But the area where you are really seeing rapid regional rail take off is in Asia.

Seoul is a good example of this new system, with the first (A) section already opened. This line currently stretches 82km across 11 stations, with speeds reaching 180km/hr. The goal is to hit even far-flung townships like Dongtan which is already a fast-growing master-planned community. Building out multiple of these GTX lines, so it is hoped, can decongest some of the pressures on the center area of Seoul, while retaining and even adding to the overall agglomerative economies of the metro.

In Delhi, there is the Regional Rapid Transit System (RRTS) in the National Capital Region around Delhi. Delhi’s metropolitan area has something like 28 million people, and this would further expand the metropolitan area through three proposed lines. Cities like Meerut (~1.8m people, 82km away) would be connected with a commute time of an hour, which would make this second tier of peripheral cities commutable to the NCR and so connected through labor markets. Guangzhou has also proposed a high-speed metro line (up to 160km/hr). I’m sure there are many other lines I’m missing, but this at least gives you a sense of the rapid regional rail boom across the world.

Why America Needs Rapid Regional Rail

Okay that’s all well and good for Korea; but what’s the relevance for America? Isn’t America already notoriously auto-focused in its transit choices? I think there are three reasons why America is actually a good location for these rail services:

1. Remote and hybrid work has led to a further wave of suburban and exurban sprawl. Free from the need to come into the office every day, workers have been more willing to move to the far suburban and exurban fringes of cities, and commute in more irregularly. This pushes out the effective commuting window of workers further out beyond traditional transit networks, and into the next set of cities, which could instead be met by a new wave of transit networks reaching into these areas.

2. American cities face an affordability crisis. It would be great if we could respond to housing shortages in cities like New York by building more — and I hope we do. But the single easiest way to increase the number of affordable housing units near metros is to build additional transit links. This will bring existing cheap housing units directly into the commuting shed of metros — thereby dramatically expanding the number of cheap housing units commutable to major job centers.

3. Auto focused commuting faces inherent capacity constraints and externalities. As more people are driving into urban cores from further away — that places necessary strain on car-centered infrastructure. Even if that doesn’t mean that suburbs are literally unsustainable, we could likely reach much higher throughput with fewer externalities like vehicle emissions by relying more on rail solutions.

With that in mind, here are a few sample crayons which illustrate what rapid regional rail could look like in the NYC, LA, and Bay Area metro areas:

We haven’t yet touched on the cost. Most of the other countries discussed here have been able to construct rapid regional rail services at an affordable price, making them positive NPV investments.

The same would likely not apply to new rail investments in the US: it simply costs too much to build in the US, and it is far too difficult to address local NIMBY opposition, to get any of these projects off the ground. Which is a shame! Because rapid regional rail seems to be taking off around the world — and it would be great for that to happen in the US as well.

[Edit:] Since I wrote the post, I found these plans for a “regional rapid transit” system in the Bay Area from 1955, with travel time projections and projections of large population increases in the area. Really illustrates the far-sightedness of planners back in that era, and how everything ground to a halt in the 1970s as NIMBYs/downzoning/end to transportation construction.

You see these claims around all the time around urbanist discourse broadly, but they haven’t attracted as much academic engagement one way or the other.

I think this thesis is broadly misguided and does more harm than good: suburbs are not Ponzi schemes, they are not inherently fiscally unsustainable, and in fact cities are the ones host to a lot of cost bloat issues. It would be better for advocates to tone down the more inflammatory aspects of their rhetoric in favor of their more balanced policy agenda (which does not require the Ponzi language); or else do the data work to actually support these strong claims.

Steelmanning Strong Towns

There are a lot of things Strong Towns folks say which I agree with. They generally advocate for higher density; safer transportation patterns; and addressing housing shortages, all things I'm happy to support. Relative to other urbanist groups, they have a focus on municipal finance questions. That, too, seems broadly reasonable and appropriate.

I think they are also correct that legacy infrastructure — think sewers and bridges in particular — will face problems in the future as they reach the end of their effective life cycles and require maintenance and upgrading. The costs of doing all this have likely increased from just a few years ago, given higher inflation, labor costs, materials costs, etc. I can easily imagine these costs become quite burdensome for various municipalities, especially poor suburbs, which are often operating with tight budgets and without as much state capacity. So again I am happy to believe there are real issues here which require some coordinated policy responses.

The “Ponzi” Scheme

However, Strong Towns advocates typically go far beyond these claims. They assert, quite literally, that suburbs are a “Ponzi” scheme — and they seem to mean this. The argument is that suburban sprawl literally does not pay for itself when considered on a lifecycle basis: that the deferred maintenance and lifecycle costs of suburban developments are higher than the benefits, such that bankruptcy is inevitable. More development can generate short-term cash, but the long-term trajectory is the whole system collapsing.

The costs that Strong Towns advocates focus on relate to lower economies of scale with respect to suburban developments. Lower densities require sprawling road networks, and high repair costs. Service provision costs for things such as snow removal, police, fire, etc. all snowball, so the argument goes, as the service area expands.

These economies of scale imply, for Strong Towns advocates, that dense urban development is inherently sustainable, while suburban sprawl is inherently unsustainable and can only be propped up in the short-run.

Advocates Exaggerate Infrastructure Spending

The first basic reason this narrative overstates suburban gloom is the simple reality that government budgets — at federal, state, and local levels alike — are primarily about social spending, rather than infrastructure.

Judge Glock and Tracy Loh pass along this Urban Institute report, which shows that all road and highway spending accounts for just 5.6% of state and local budgets. Per capita spending growth has been fairly low in this category so the budgetary share is actually down over time, even as the past infrastructure backlog has gotten somewhat addressed. Similarly, sanitation (3%) and sewerage (2%) are small items as well.

What do local governments spend money on instead? The main categories of expense you should be thinking about are social services like schools, police, and healthcare.

To be sure, this is still perfectly consistent with some municipalities facing concentrated shocks, and existing infrastructure systems facing higher costs down the road as they inevitably need to get maintained and refurbished. But it’s hard to tell a story about the inevitability of suburban doom when main drivers are just fundamentally pretty small potatoes in the overall budget.

Strong Towns advocates typically do not provide straightforward math here to back their claims — and there are important questions about the validity of the numbers they do provide — but you can multiply highway spending by 2 or 3x without eroding suburban budgets completely.

The other complication here is that many municipalities rely on capital budgets to cover their infrastructure backlog, not just their operational budgets. But here, too, cities also use capital budgeting to pay for other expenses, like school spending, and those capital budgets are often larger. Additionally, these capital plans are often collateralized or backed directly by, for instance, water fees. So again cities may need to raise some user fees down the road, but it’s not obvious how this inevitably turns into a Ponzi scheme.

Urban Costs Are Typically Higher Than Suburban Costs

The real crux of the issue however, is that infrastructure costs are typically higher in urban areas than suburban ones. A good and typical example here is suburban Roseville, MN (where I spent my elementary school days — a great place to grow up). As Ryan Radia helpfully suggests, the municipal per capita spending is about $2,000; the county is about $1500, mostly health and police, and the school district is about $2400). Enterprise operations and public works — the typical nuts and bolts of infrastructure spending — make up about $720 a year, or roughly 12% of this total spend.

By contrast, municipal spending in Minneapolis is about $4,200 per capita, of which public works and water make up about 30%. Minneapolis city schools have a budget closer to $20k/student, and Hennepin county is another $2k/person as well.

So whatever are the theoretical benefits of amortizing infrastructure spend over a more densely packed population; in general cities have greater diseconomies of scale which lead to both higher infrastructure spending as well as more spending in the main social service categories.

The sources of excess urban expenditures are varied. For things like sewers and water — a lot of the costs seem to stem from things like environmental review, permitting, and so forth — which are really regulatory barriers and not so much about the nuts and bolts of just building the infrastructure. Urban wages are higher, especially union wages which are going to be more binding in urban areas. And then finally you just have greater cost bloat and inefficiencies in urban governments in general, which face fewer competitive pressures relative to their suburban counterparts. Or in Philadelphia; you see that streets and sanitation are just $155m out of a budget of about $5 billion; with the largest item reflecting benefits and pensions.

Here in New York City for instance, we are on track to spend $39k per pupil in K-12 schools. This is not even counting the capital budget. Even though NYC has the best transportation system in the country, which handles 7/8ths of the students in public schools — the busing expenses for the 1/8 of other students takes up over $10k a student traveling on bus, exceeding the total education budget of many states. We also see all the time the costs of dealing with and maintaining legacy infrastructure.

Now, there may well be good reasons for cities to spend more than suburban areas. Perhaps cities take care of different populations or offer more in the way of public services. This point is debatable — there are good reasons to think social mobility and even social integration can be good in many suburban areas.

But if the question is one of pure fiscal sustainability, on average, the answer is pretty clear: cities are more costly to run than suburbs. The pro-urban bias of Strong Towns advocates leads them to obfuscate this basic fact in favor of a doomer narrative of suburbs which lacks a strong foundation.

Do Suburbs Exploit Cities?

I think a somewhat better argument concedes the greater cost structure of cities — but instead argues that suburbs are extractive through a different route; because suburbs are full of people who commute to the urban core yet do not pay local taxes.

This is a better argument because cities do bear many costs on behalf of suburban commuters — financial costs in terms of infrastructure and maintenance for items used primarily by those commuters.

However, cities also don’t bear a lot of the costs related to those residents — in particular, the education, services, etc. costs which dominate city budgets.

Additionally, the whole idea of a monocentric urban core, with all the jobs, does not fit as many jobs have left for suburban areas themselves. For instance, in Cook County, half the jobs are already outside Chicago city. Remote work has shifted the balance even further, with more suburban commuters staying at home all days.

I am more sympathetic to cities here, and obviously mechanisms like congestion pricing are one way to try to collect more tax revenues from suburban commuters to pay for associated expenses. But I still suspect the narrative is not as one-sided as you might think at first glance. The environmental and traffic death unsustainability of this whole arrangement — with suburban drivers commuting in all the time — is also more persuasive, though I suspect those costs do go down substantially as we shift towards autonomous and electric vehicles and a more renewable electric grid.

I think the best argument against suburbs is exclusionary zoning — minimum lot sizes and other instruments are used by suburbs to push out poorer residents into the urban core.

Towards Better Urban Governance

This is an admittedly cranky post, and I’m sure urban advocates who have read this far will find a lot to complain about. So I do want to emphasize again that I have no problem with the idea that infrastructure costs money, and that we probably face substantial deferred maintenance issues with our infrastructure stock (in both suburban and urban areas). I’m also actually on board with “urban doom loop” mechanisms which also impact small cities — through mechanisms that involve population declines and legacy pension and service obligations, which are again the real budgetary costs as well as the infrastructure.

But more broadly I’d like to ask the urbanist community to take more seriously the problems of urban governance, and see the role of suburbs through a slightly more sympathetic lens. Whether you like it or not, most Americans live in suburbs — they do so because they find it an attractive and low cost of living. Many people would rather live in cities instead, but are dissuaded by low housing inventory, high taxes, high costs of living, and other aspects of poor urban government. The competition between suburbs and cities is one of the main restraining forces keeping cities as functional as they are.

So rather than construct narratives about how the seemingly placid suburban existence is actually on the verge of disaster, and defending cities — I think urbanist efforts would be better spent on greater advocacy and monitoring of urban areas. Obviously YIMBY activity is helpful; but I think there is much more to be done in complaining about urban areas to improve them.

Subscribe now

The 1980s CRE Boom

The 1970s and 1980s were an interesting time. As I’ve talked about here previously — this was the era of stagflation: shocks to financial constraints and the supply side of the economy which hit production in various ways. Modern treatments of monetary policy over this period think of this as a time of Fed overstimulating the demand side, but more recent work confirms that the supply side was a big part of the story as well.

The graph illustrates housing production trends, with multifamily production highlighted in blue. It's crucial to understand these trends in the context of economic fluctuations. The economy experienced recessions due to tightening financial conditions, which severely impacted housing production. This impact is visible in the inverse relationship between housing production and interest rates, represented by the 10-year Treasury rate shown in red. When housing production declined, rents (indicated in green) would increase, eventually leading to a resurgence in housing construction as economic conditions stabilized. High demand for housing, driven by the desire to invest in tangible assets as a hedge against inflation, also fuels this cycle of booms and busts in the housing market.

That was the macro backdrop heading into the 1980s: when interest rates continued to stay elevated (the 10-year Treasury averaged over 10% in the decade, with mortgage rates of course being above this). Based on everything I mentioned before, you would expect to see multifamily housing production would crater over this period given those high financing costs. After all, today we are also seeing declines in multifamily housing production even at far lower interest rates.

Instead, the 1980s were a booming decade for commercial real estate. We were producing more multifamily units in the early 1980s than even the post-pandemic apartment boom in absolute terms (even setting aside the fact that the population is a lot higher now). Commercial real estate, overall, you can see was growing as rapidly as 6% each year based on the size of the existing stock — a huge contrast to the construction in subsequent decades, especially the post-GFC bust.

So what happened? How did we boost commercial real estate production, especially in apartments, despite the high financing costs?

1981-1986 — Tax Depreciation Comes and Goes

This answer to this question comes in a great piece by Alex Muresianu who covers the impact of the 1981 and 1986 tax law changes on the depreciation of commercial real estate. The general idea is that when companies make investments in commercial real estate assets, they can write down the value loss or depreciation in the asset over time, recognizing these as a non-cash expense lowering taxable income, thereby increasing the after-tax returns for investment and encouraging firms to do more of it.

Just as important is the timing of when depreciation credits arrive. With accelerated depreciation, firms are able to write down more of the value of an investment earlier. Note that accelerating depreciation doesn’t change the total undiscounted amount of corporate tax receipts — but it means they arrive to firms much sooner to the time they make their investment.

Tax reform in 1981 had two big changes — it really lowered depreciation timelines from 36 to 15 years, and it increased the speed of depreciation from a 150% declining balance to 175%. This was a massive shock to depreciation schedules, and effectively brought forward heavily the tax benefit firms get from investing in Commercial Real Estate. The 1980s CRE boom in construction was the consequence.

Why does the timing of tax credits matter so much to firms? A great deal of corporate finance research has emphasized the importance of hurdle rates for investment: firms have fairly static rate of return requirements to engage in investment. Bringing forward the timing of tax benefits, even when it does not change the total tax liability paid over the life on an investment, raises the rate of return and can help projects pencil out. These are particularly important in real estate, given the large capital tangibility of the projects and financial constraints faced by developers. This is important enough that public housing advocates emphasize that using public funds for social housing can help projects which wouldn’t get funded by private developers employing high hurdle rates.

The 1980s CRE boom fizzled out by 1986, when the Tax Reform Act of 1986 (TRA86) again shifted depreciation schedules. While this tax reform package was good in many other dimensions (equating the highest end income and capital gains tax rates for instance), it turned out to be a disaster for real estate, by drastically extending depreciation schedules, which are now 39 years for commercial real estate (along with straight-line depreciation). The upshot is that developers experience drastically delayed cost recovery schedules for CRE investments. Inflation and the standard time value of money imapct lowers actual recovery rates further. Slowed development pipelines are the consequence.

The negative impacts of TRA86 were recognized immediately by commentators — James Poterba has an NBER paper from 1992 arguing “TRA86 reduced incentives for rental housing investment, contributing to the decline in new multifamily housing starts from 500,000 per year in 1985 to less than 150,000 in 1991. In the long-run these policies will lead to higher rents.” So the idea that housing affordability is impacted by the financing environment for builders isn’t new.

Now, there are a few other details I’m skipping over here. For one, many commentators have grown to think of tax rule changes as contributing to an “overbuilding” problem in CRE, especially as it was followed by financial distress in the asset class and the S&L crisis over this period. These were also related to other tax dodge issues in real estate, covered in Alex’s piece. There’s a lot more that can be said here, but suffice to say that it’s possible to boost the investment in CRE without adding additional tax haven benefits, and sitting here in 2024, with super high apartment rents, we have a different perspective on the problem of “overbuilding.”

TCJA in 2017

This brings us to another major revamp of the tax code — TCJA in 2017. This alleviated some of the tax code’s bias against investments, especially short-term investments. There was a broader debate over the tax bill passage over whether we should move to full expensing of all capital investments. The Tax Foundation estimates suggested that real estate would bear a large brunt of those tax changes, given how important long-lived structures are, and drive the largest increase in economic output.

As Paul Williams points out, the real estate industry itself opposed some of these changes:

The industry concerns with expensing are based on historical experience. Accelerated depreciation of real estate in the early 1980s led to tax driven, uneconomic investment. Tax-motivated stimulation of real estate construction that is ungrounded in sound economic fundamentals, such as rental income and property appreciation expectations, creates imbalances and instability in real estate markets. No other major country in the world has immediate expensing of real estate. The market implications of expensing real estate are risky, untested, and unpredictable. The negative consequences could harm state and local communities (through reductions in state and local property tax revenue), the financial security of retirees (through pension investments tied to real estate), and the banking system (through the declining value of real estate on bank balance sheets and systemic risk to the financial system).

In fairness to the industry — the real estate boom and bust cycles in previous decades were real, and you can obviously understand how increasing real estate production might shock rents. I’m not the first person to point this out, but it really is striking how NIMBYs are so skeptical of the link between lower increased supply and lower rents, in contrast to capitalists who see that link as crystal clear.

But just because real estate lobbyists don’t want more competition is not a good reason to give it to them. It’s important to fight against the real estate lobby and insist they receive tax benefits — that’s the best way to get new construction.

What Should Depreciation Schedules Look Like?

These accounting details are important in the context of development pipelines dry up across the country. Firms are looking at a weak rental picture in the next year or so, as the supply that was started during the pandemic comes online. But we’re facing the prospect of higher rents in the subsequent years — as the lack of construction today will ultimately result in a supply crunch and raise apartment rents in the longer term. But construction firms aren’t able to look this far forward and plan for more construction today, at least in part because of financing frictions, driven by higher interest rates and slower bank credit, which are shutting down plans for construction today.

One way to think about this shock is in terms of crowd-out. As federal spending has grown on projects like manufacturing through the IRA, interest rates have gone up and this has crowded out housing production.

A possible way to address the situation would be lowering interest rates through monetary policy, and there probably is some scope for that as inflation has come back down.

But accounting changes can also play a role. I think the optimal policy here would at least resemble the tax reform bill of 1981 — shorten depreciation schedules for commercial real estate and include bonus depreciation, so as to shorten investment cost recovery timelines and encourage a boom in construction. I also think the political economy for such changes might be more favorable than in the low interest rate environment of 2017. Back in 1981, we were looking at an environment with falling inflation but interest rates that remained stubbornly high, just as today. There is probably a greater willingness to consider changes in depreciation schedules in the context of addressing those financing frictions.

Even better would be full expensing — allowing firms to deduct the full cost of structures the year they incur the expense, the same way that is allowed with wages. The Tax Foundation estimates this would raise output by 2.8%, wages by 2.4%, and add 569,000 jobs. On top of that, it would likely reignite a multifamily housing construction boom which would bring down rents across the country.

It’s commonplace these days for YIMBYs to think about regulatory constraints — zoning and building codes — which hold back housing construction. I think increasingly they also think about operational costs for multifamily building operators which raise the cost structure and rents. Financing and accounting adjustments also need to be part of the pro-housing coalition agenda. The coming expiration of the TCJA by the end of 2025 is a good motivation to kick off the debate over how to structure the tax code to best support American prosperity.

Subscribe now

My argument is this has basically been a failure so far for a variety of fundamental reasons, and is unlikely to improve anytime soon.

What Does FinTech Failure Mean?

First, let’s do a quick gut check on one of the largest and most important consumer credit markets: residential mortgages. This is an industry that, over the time span from 1997-2023 shown below, has seen enormous technological changes. You saw computers being substituted for underwriting decisions (as Fannie and Freddie rolled out desktop software); credit scores used for underwriting decisions (this is a classic example of FinTech in the sense that we start to use algorithmic tools to collapse a variety of consumer credit attributes into a number for loan underwriting decisions); the rise of private label securitization and then FinTech mortgage companies; the availability of the internet and cost comparisons online; so on and so forth. This follows also the mass adoption of Option Adjusted Spread techniques in the 80s to quantify prepayment risk.

Suffice to say it’s a large, sophisticated marketplace with a high degree of competition at multiple levels. But what does that ultimately mean for consumers? We can look at the total price of intermediation: the mortgage rate that consumers pay above a risk-free benchmark, like the 10-year Treasury. This rate is actually high these days — close to 300 basis points — and you can see this is close to historic highs.

Now, you might say that’s a product of an unusually inverted yield curve, and implied prepayment risk in the future as future mortgage rates fall (that shows up as the green bar for prepayment risk). But you can also see the primary-secondary spread there: that’s basically the fees charged by mortgage originators, and you can see that’s also pretty high by historic standards. So despite the growth in FinTech mortgage lending; the cutthroat competition across the expansion of technology in the sector — the cost of mortgage origination has basically never been higher. So why are things getting more expensive even as technology gets better and better?

Steelmanning the Case for FinTech

So here, it’s probably helpful to define FinTech a bit more and think about why technological innovation is plausibly a disruptive force we should expect to have a bigger deal. As one friend wrote to me when I ran a poll to establish people are interested in FinTech as a topic on this substack: “I'm a longtime fintech skeptic; if you write about that, would love to see a non-ZIRP, non-crank explanation for why anyone thought it would succeed.”

I think the basic argument is that technology has pretty strongly disrupted many industries its entered — particularly industries that are focused on information, for which advances in processing are plausibly important. There are a few key domains in which you imagine this is important:

1. Loan Origination: Consumer credit is all about establishing risk to make an underwriting decisions, and there are basically two things you can do. Either you can require collateral (giving you an asset to repossess in the event of default), or you use soft information. Historically, this entailed sort of looking over the person as they walked in your bank to see if they were sort of the chap who would pay you back — basically a social reputation thing, and you can imagine all the pitfalls associated with that.
   
   Instead FinTech offers you the potential to use all sorts of “hard” information — ie numerical touch points predicting borrower risk — to replicate the “soft” information you previously arrived at through human judgement. The end product is being able to use a much broader range of information to accurately predict default risk, and so make more lending decisions. A pretty tangible way this matters is through something like consumer credit scores or corporate bond ratings — innovations in information which generate concrete measures of performance, thereby enabling more lending to creditworthy firms or consumers (who otherwise would have been pooled in risk with riskier borrowers).

2. Financial Valuation: You’d expect that improvements in data acquisition, data processing, and so forth would improve our ability to value and price financial assets, and indeed there is some evidence that financial markets have grown more informative (in the sense that current market prices better predict future cash flows).

3. Financial Competition: Think about banks in the 1970s. Banks were strongly regulated in their ability to open branches; and so you were stuck banking at your local community organization. Compare that today, where you can move funds across financial institutions with a click of the mouse. Surely, that amount of competition should erode away rents associated with financial accounts; just as ending geographical rents has resulted in arbitrage of other goods? Surely the cost of borrowing would fall as firms have access to a much broader range of intermediaries, and can more credibly demonstrate their creditworthiness through new data and analytics? And free entry should erode away rents that do wind up?

4. Technological Innovation: To further support all of this, you would expect to see new technologies and new services pop up, especially by new firms, in order to cater to customer needs and demands — as indeed you do see across all sorts of sectors. Some of this might wind up being technological disruption, in the Clay Christensen sense of new low-frill options to capture ignore markets, and ultimately wind up lowering costs.

And some of these positive innovations have indeed happened. If you’re accessing your bank services through a phone app, investing with a roboadvisor, borrowing through a DeFi platform on crypto, or using a nonbank company for your mortgage or student loan — you’re benefitting from FinTech.

But none of this is the main story. The main story is that the provision of financial hasn’t really changed that much, especially the cost of providing those services. So now let’s turn to the various reasons for that.

Asymmetric Information

The first problem is information asymmetries, and a great example here is iBuyers — companies like Zillow or Opendoor making house sellers take it-or leave it offers by leveraging their detailed housing information to make accurate valuation offers. It makes a lot of sense, and has the potential to disrupt the realtor network: why should agents continue to make 5-6% for every housing transaction in a world in which websites can list housing attributes, and algorithms can accurately assess housing values?

Greg Buchak, Gregor Matvos, Tomasz Piskorski & Amit Seru have a great paper which describes the key problem: home owners have a lot of detailed private information about their house value, and only sell to iBuyers when it turns out the algorithms have overvalued the property. So even when the algorithms are doing a pretty good job of assessing house values in isolation, they play a role in a complicated game featuring strategic interaction and adverse selection, and in the end the machines lose out.

Platform Economies

One of the key innovations technology has opened up in the last few decades is the notion of platforms — information aggregation devices which collect and integrate information and market interactions, thereby opening up a lot of economic value but also providing a centralization location for value capture.

One platform example, brought up with realtors, is the MLS platform: where realtors list properties for sale. It’s commonly thought that realtors sustain their 5-6% fee structure through their control over the MLS platform

Or take another specific example: interchange fees, and Visa specifically. The Acquired podcast has a great episode on them here and describes the fee structure for credit card processing:

Visa, let's round it to 0.2%, gets 20¢ of that $100 shoe sale. But the cool thing about their 20¢ is there are basically no variable costs. It's not dealing with fraud. It's not moving heavy data around. Merchants are allowed to have a 20-character name in Visa's network. This is tiny amounts of data. Stack as much metadata as you want on top of that, we are not shipping around huge payloads here…

The most shocking thing about the business is they have 50% net income margins. Of the $30-ish billion that they made in revenue, their net income was $15 billion.

David: This is absurd. All the picture we painted in the whole story, it was all building toward that climax of they have created something with essentially zero marginal costs in, perhaps, the largest market out there, certainly one of them is global commerce—both e- and non-ecommerce.

Ben: As Visa would argue, both consumer but also B2B commerce.

David: Fifty percent net income margins on $30 billion in revenue. There it is.

Ben: You might say, wait, if they have 50% net income margins, what is their gross margin? Because is it SaaS level good at 75%–85%? Their gross margins are 98%. There are no variable costs in this business. There are no cost of goods sold. It's crazy.

As their episode goes into in detail: setting up the whole credit card interchange system to create an interlocking set of interests did entail a lot of technological and business innovation. But now that it’s here: it’s basically pure profit in the form of $550b market cap for Visa, $413b for MasterCard, and $146b for American Express. So that’s basically a trillion dollars in the market value of pure rents extracted by the credit card companies for the cost of processing payments: a durable monopoly because it sits on the commanding heights of a payments platform that it’s in everybody’s interests to sustain with huge network effects — the Chase Sapphire-Americans who benefit from rewards points, issuing banks who benefit from customers — at the cost of retailers who pay a substantial tax to keep the whole thing going, ultimately passed on to consumers.

There is a better solution here, but it requires substantial government investment. Pix in Brazil, UPI in India, Faster Payments in the UK are all state intermediated payment processing technologies which create platforms for payment processing at close to zero costs, thereby removing the 2% tax on the whole economy extracted by the credit card companies. For India in particular, the payments layer serves as the base for whole additional layers (the “India Stack”) of financial technologies providers. Imagine, for instance, businesses and consumers who can credibly commit to loan repayment based on verifiable histories of sales volumes as recorded on payment transactions.

So what’s the takeaway here? I think the lesson, as Sergey Sarkisyan suggests, is that technology adopted by first movers, applied in financial contexts, can easily exclude rivals and grow monopolistic. To avoid that, you really need open protocols which ensure democratized information access.

Regulatory Burdens

Returning back to loan origination — I think a key driver here of those rising loan origination expenses have been genuine increases in paperwork and regulation. Now these may be good regulations which address the fraud which was common before the GFC. But it has steadily added to the cost and expenses inherent in loan origination, as lenders now undergo extensive income verification, and indeed many lenders have simply pulled back from the space entirely especially out of threat of lawsuits.

This is a tricky issue, because some degree of prudential regulation seems warranted (remember the point above; there is asymmetric information in this space) but it all reflects a bit of an informational arms race. People can use IT and technology to do more fraud, which requires more resources to clamp down, which triggers more regulation, which increases compliance costs and expenses. A related regulatory problem is systemic risk: regulators (naturally) want to avoid fragility concerns, and to an extent that’s mitigated by cozy insiders who sit on rents.

Arms Race Losses

Also in the category of arms races is the general zero-sum nature of a lot of financial activity, which generally results in a lot of sort of wasted effort to front-run others profits. This is of course most evident in high-frequency trading, which is highly competitive and features huge investments, but is basically entirely socially wasteful in shifting around some fixed set of profits among a small set of high-frequency firms (with dubious, at best, benefits on liquidity overall).

But you see similar trends across the FinTech space — here are a few firms (Affirm, soFi, Worldline, PayPal).

For companies like SoFi or other mortgage FinTech firms — the arms race that I have in mind is just cream-skimming good quality customers and offering them lower interest rates as rates continued to go down. That’s a zero-sum game, and in particular is only value as rates kept falling. And so that’s in part why I think you’ve seen the stock prices of many of these companies start to fall once we hit that rising rate cycle. It turns out there just wasn’t as much fundamental business innovation; so much as these more zero-sum arms race components.

Behavioral Biases

Let’s come back to banks. As Itamar Drechsler, Alexi Savov, Philipp Schnabl have emphasized, banks pay consumer interest rates on deposits far lower than true market rates. And the resulting bank betas — how sensitive consumers are to moving deposits with different interest rates — are if anything even lower these days than historical norms. The upshot is that consumers basically lose enormously on money sitting in bank deposits, the entry of other institutions hasn’t done anything to help, and this all winds up as bank profits (one caveat here is banks wind up spending a large chunk of the resulting profits on paying for bank branches and customer acquisition costs more broadly to get these people in the door).

It’s all fairly surprising given the ease of price comparisons and money transfers, especially given the internet, and broadly I think it’s hard to think what’s going on here aside from consumer inattention.

Especially in the wake of the zero lower bound period: the whole issue of deposit rates just really weren’t salient to consumers, and so people focused instead on other characteristics and features of banks to decide on, and so depositors were pretty “sleepy.” Similarly, people are “woodheads” or inattentive in their refinancing decisions; they pay outrageous fees for consumer credit cards; high rates for actively managed investment products; and creating low fee products just doesn’t move consumers the same way that it does in other product categories.

Why all this is happening is maybe a bit puzzling at some deep level; but for our purposes we can chalk this all up to the idea that consumers basically don’t seem to spend that much time or energy thinking about financial issues, get pretty strongly bamboozled by financial institutions as a result, and there’s probably not that much that technology is going to do about it.

One thing we could potentially do here is make it easier to handle payments through non-checking systems — ie, imagine you could pay rent by writing a check from a money market mutual fund invested in a high rate of interest. But that’s banned by regulators — those funds can’t access the ACH system of payment processing — in part (back to an earlier point) because regulators don’t want the systemic risks that might happen from those deposits fleeing the system for mutual funds.

How to Make Better FinTech

How to fix this whole situation — and help FinTech realize its initial promise — turns out to be a pretty tricky process. There’s probably room for some regulation to address prices which are not salient to consumers, which wind up driving up fees and expenses, but you have to be careful to avoid overregulation in ways that drive up costs. It seems helpful to have some public utilities control the commanding heights of certain digital platforms to enable information sharing rather than value capture, but you also want to ensure adequate incentives for private actors to build some necessary infrastructure themselves. You could imagine creating more public goods here — maybe Central Bank Digital Currencies which directly pay out a fair rate of interest, cutting out bank deposits entirely for the purpose of liquid transactions — but that would have pretty disruptive impacts on existing markets and raises all sorts of privacy issues. You could try to break up existing private exclusive platforms (like interchange or the MLS platform), but the incumbents would probably fight those pretty strongly.

So there are no magic bullets here, and so my basic expectation is the fundamental challenges in this space are so strong that FinTech will steadily move along without drastic or revolutionary benefits for consumers, and hence will broadly be a failure from the perspective of the lofty claims and hype made on its behalf. But at least it will mint a few billionaires along the way. And I could be wrong! I had greater hopes for innovation in this space, and it’s possible they’ll get realized eventually.

Subscribe now

In fact, it took about as much time between the origin of agriculture and the emergence of the first true cities, in the 4th millennium BC, as has elapsed since then. When cities did first arise, you saw the near-synchronous emergence of urban centers in the Bronze Age characterized by internal stratification and elites across Egypt, Mesopotamia, the Indus Valley, and Central Asia. Why did it take so long for these cities to form, and what explains the timing? Why did they start up so quickly all over the world?

It’s a foundational question at the heart of urban economics, politics, and trade, and a new paper by Matthias Flückiger, Mario Larch, Markus Ludwig, and Luigi Pascali provides a provocative answer: the key to the emergence of cities in the Bronze Age was, well, Bronze.

Tin and Copper

The key thing about bronze is, to make it, you need two minerals: tin and copper. Both metals are somewhat rare, though tin is considerably rarer. Bronze was sufficiently valuable in that era — for both agricultural production and to make bronze weapons — that it therefore makes sense to spend enormous effort to transport both copper and tin to agricultural areas to smelt into bronze for use in production and defense.

The origin of tin, in particular, is a source of huge scholarly debate in the historical community, and Flückiger et al. side with the current consensus: at least in the initial phase of the Bronze Age, tin was mined in the mountains around Uzbekistan and Afghanistan by the BMAC civilization. It might seem slightly odd that some of the best early signs of urban civilization come from Central Asia, but it makes much more sense if this was a terminus for an essential trade product. From there, tin travelled over mountains to the Indus Valley civilization, and from there by boat to Ur. Ur itself, the world’s greatest urban center at the time, was situated perfectly as a trade entrepot for transshipment of metals further upstream in Mesopotamia.

Cities like Ur grew by inserting themselves into this metal trade — they were basically stationary bandits who taxed trade at crucial nodes in the metal shipment network, which provided a surplus to sustain local elites. We then see the rise of local elites, new religions, and complex writing to manage this trade.

It’s a very neat theory, and the authors provide some compelling evidence. First, they can draw connecting lines from local agricultural populations to local sources of metal – note in particular the Central Asian tin sources.

Next, they draw a trade index — the lowest cost trips connecting agricultural land to the nearest copper and tin mines, producing a set of optimal trade lines. Areas higher in this index are essential more central to the metal trade network — and this is also where the cities generally line up.

This result shows up in a regression as well — ie, areas which are more dense in trade links are more likely to have a city, even controlling for other relevant factors. Interestingly, you see that local agricultural productivity, which you might imagine is important for local urbanization, actually enters in here negatively after controlling for transit.

A good example here, which the authors discuss in some detail, is Assur, the capital of the Old Assyrian empire, located at the confluence of the Tigris and two important tributary rivers. The land around Assur was fairly desert-like, even back then, but it commanded an important node in the trade network, and we have extensive records of Assyrian merchants trading all over this region, especially in Anatolia. The implication is that cities, and ultimately whole civilizations like the Assyrians, were founded on the surplus to be had from taxing and trading high value metals passing through their territory. Supporting the analysis further, they argue that the hold out power — something like the ability of merchants to skip a particular city and head elsewhere instead — should associate with taxation authority, and seems to line up with the size of the elite class. This general argument is also reminiscent of the paper by Bleakley and Lin – that cities are often found on the fall line, where portage is necessary up rivers. So centrality in trade networks (and probably in transhipment across modes of transport as well, ie unloading cargo) is plausibly a really important driver of cities across much of history.

Another piece of broadly confirming evidence here is the Uluburun shipwreck — a Late Bronze Age sunken ship with enormous troves of trade goods, including huge amounts of raw tin and copper, indicating the vast scope of trade that took place in this period, especially for those key metals.

The End of the Bronze Age

The origin of the tin in the Uluburun wreck is still being debated, but the general historical consensus seems to be that at least some tin in that time period was still being sourced from Central Asia, while some was starting to get sourced from Cornwall in Britain.

This brings us to discussing the cataclysmic end of the Bronze Age, culminating in around the period 1177 BC in Greece and the Middle East, but actually ending before then in South Asia.

The Indus Valley has a number of unique characteristics relative to the other civilizational zones in this era. It rose and fell quite quickly — a period between 2600 BC and 1900 BC. It was also unusual in social organization; here’s what Patrick Wyman says:

What we don’t see in the Indus Valley Civilization is evidence of strong hierarchies. As far as we can tell, no kings, exclusive cliques of nobles, or other elites dominated Indus society. There are no grand royal tombs or palaces, no monuments to the power and ego of a powerful figure determined to make his or her mark on the built landscape. Cities were filled with large complexes of relatively uniform houses, with rooms for domestic and craft activities arranged around a central courtyard. These houses were then grouped into separate quarters or neighborhoods. No one individual or group seems to have stood above the rest. For that matter, there’s precious little evidence to suggest that even the great cities somehow exercised power or sway over their smaller neighbors.

This doesn’t mean that everybody was totally equal within Indus society; instead, it means that we don’t yet have the evidence or the conceptual tools to understand how Indus society was organized. It was certainly comprised of different groups, who presumably cooperated and competed with one another, but we don’t really have any idea on what basis those groups were organized: maybe kinship, maybe occupation, maybe ethnic identity or religious affiliation.

Shedding some potential light on these identities is a new paper arguing that the as-yet undeciphered Indus Script were licenses issued by guilds to enforce standardized taxation rules, and the emblems were signifiers for the local guilds themselves.

The argument I would make is that it would make a lot of sense of these guilds were really jatis – ie endogamous caste organizations which are an essential part of South Asia today. It’s possible this institution was present already in the Indus Valley Civilization, growing to organize the emerging trade networks. While pharaohs ruled in Egypt, priest-kings held sway in Mesopotamia, maybe India innovated a kin-based method of social organization to handle trade, which then ultimately provided so durable as to last to the present day. It will take more research to sort that out, but it would be fascinating if the legacy of the Bronze Age can be seen now the social organization of the Indian subcontinent.

Then you have the rapid collapse of the Indus Valley Civilization around 1900 BC, centuries before the ruptures that happened further west. Many accounts emphasize climate shocks, but it would make a lot of sense if the ultimate cause were trade related. Around this time, the sources of tin in the BMAC civilization were facing serious stresses, with the entry of steppe nomads from the north. Potentially the steppe nomads cut off the tin trade to the Indus Valley; maybe the tin started to move directly west overland through Iran rather than over the mountains to the Indus Valley (maybe being shipped directly to Mari); or potentially new tin in Cornwall was brought online and proved to be cheaper than the Central Asian tin.

Whatever the reason; the collapse of the eastern side of the Bronze Age world was followed by wholesale collapse of the western part. Clearly, the end of international trade routes that kept the system going was part of the process of complete system collapse. But why did the trade routes fall in the first place?

One theory is that this is when mass infantry tactics were developed, which turned out to overturn existing states. A related theory, which Flückiger et al. seem to favor (as do others), is that iron was invented around this time period and started to diffuse. Because iron can be found in many more places, it produces a more decentralized trading network. And so you have a potentially catalytic effect — the introduction. of some iron weapons and warfare starts to eliminate some Bronze Age states, which may feed the demand for investing in more iron weapons, and so the whole system collapses. It’s certainly a very attractive theory from an economics perspective, though I gather direct evidence for actual iron weapons used in the Bronze Age collapse is more mixed (though iron weapons themselves may degrade more over time), and so it’s hard to reject the theory that the Bronze Age collapse happened for other reasons and iron weapons were the endogenous technological response (which is probably the dominant historical view today).

There is much more in the paper — including detailed statistical and trade analysis to support all of this, historical analyses of specific episodes and really neat datasets, and all sorts of new questions such as what role do amber, jade, obsidian, and developments in the New World have to do with it. In the end, cities and complex political organization reformed themselves in the Iron Age — but this time, the basis of trade (for whatever reason) was now bulk trades in primary agricultural commodities. But this whole paper was one of the most thought-provoking articles I read, and I hope it spurs more research in these fascinating areas of human history.

Subscribe now

The Services to Goods Rotation

The backdrop here is that goods prices have been steadily getting cheaper over many decades, relative to services prices. This is a natural consequence of Baumol’s cost disease and more rapid productivity growth in manufacturing.

The pandemic led to drastic shifts in this pattern, as many categories of consumption became risky or prohibited from the perspective of Covid risk. This affected every sector in different ways; but a rough way to categorize the impacts is to say that services were heavily shocked (think haircuts, dentist appointments, in-person restaurants, etc.) relative to goods (autos, luxury watches, housing). This shows up pretty clearly in the data, when you can see a large sectoral reallocation away from services towards goods, which slowly reverses over the next few years as the economy reopens.

Under an assumption of flexible prices, this kind of sectoral reallocation would work out just fine: people buy more cars and houses, buy fewer haircuts and dentist appointments; and so we might expect the prices of goods to rise while the price of services falls in comparison. Instead, price rigidity is asymmetric: goods prices rose, while service prices didn’t really fall, despite the large fall in demand. After a while, goods prices mostly flattened out, not declining much in absolute terms, and so the goods/services price ratio was restored through a steady rise in service prices.

Because of this interaction of price rigidity and sectoral impacts, you can get aggregate inflation even if total spending doesn’t change at all: the first rotation to goods increases the overall price level, and the second rotation back to services does as well.

The Role of Supply

This is an account which emphasizes supply shocks, but it’s important to emphasize these are supply shocks in the services sector as a result of the pandemic, which steadily got cleared up as mandated and voluntary restrictions on ordinary life were were removed — though it’s interesting that service spending has remained below trend, while goods spending has remained far above trend.

In addition to the services supply shock; there were additional supply shocks to certain goods producing sectors — autos is a great example; as many car manufacturers cancelled orders for chips, and took a long time to obtain enough replacement supplies. Over time, these pandemic goods disruptions worked themselves out, helping to explain a component of the drop in inflation. Energy is another tangible example of a supply shock; and seems to have left a more lingering impact in Europe.

Despite the partial role for supply disruption in explaining a component of inflation in goods; I think the balance of evidence suggests at least some, if not most, of the net inflationary effect was driven by the sectoral shifting of demand. To take a relatively niche category, the price of luxury watches shot up, despite production being higher in 2021 than in 2019, before slowly deflating over time. Container traffic in volume was higher in 2021 and 2022 than in 2019; the bottlenecks in shipping seem as a result to be the product mostly of even more goods trying to reach consumers, rather than reductions in net supply. It’s not completely confirming evidence: but it is interesting that total real consumption, even in sectors like autos, went up a lot, as did the stock prices and earnings of large auto manufacturers, retailers, and goods producers in general.

Real estate is probably the best example of this trend: housing completions continued to rise steadily throughout the pandemic; while housing starts went up too. Despite construction times rising in housing; it seems pretty clear here that rising house prices have primarily been the product of higher demand (driven in turn by remote work and negative real interest rates), rather than falling supply. This is a fairly clear example of a sector in which pandemic related disruptions in ordinary life resulted in large shifts in demand to categories of consumption which were more sheltered, ie a large personal home.

The Role for Policy

Even if we can explain the bulk of the pandemic related factors with the goods-to-services rotation story; what additional role did policy play? We engaged in substantial fiscal and monetary stimulus in the beginning of the pandemic, followed by policy tightening that coincided (but did not necessarily cause) the fall in inflation.

This relates to a fundamental question about what the nature of the pandemic shock actually was: was it a 2008-style of massive aggregate shock, from which we should expect large and persistent declines in output; or was it more like a natural disaster — a shock that tends to result in a short and somewhat voluntary declines in output, followed by quick recovery?

With the benefit of hindsight, I think it is more likely we assess Covid as being more like a natural disaster shock: the voluntary and temporary decline in consumption of services was inevitably going to be followed by a recovery as soon as that sector came back, just like turning the economy on and off.

Consistent with this idea, we’ve seen broadly V-shaped recoveries across the entire world; regardless of the scale of stimulus the country has engaged in. One good example here I think is India, which did relatively little stimulus, and has come out of the pandemic as one of the world’s fastest growing economies.

As Guerrieri et al. point out: there’s a clear role for policy under this situation: provide insurance to households facing drops to income. However, they suggest that expansionary fiscal and monetary policy might be less effective under these circumstances: when service consumption is shut down, incremental demand is going to be pushed into the goods sector, which are already facing capacity constraints. Pushing even further demand into this sector, already elevated as a result of sectoral reallocation, can potentially push up prices even further.

Consistent with this view is the role of excess savings: the increase in household cash buffers as a consequence of lower spending on service consumption; higher income (as a result of fiscal transfers), and higher wealth (as a result of lower interest rates and higher discounting). This facilitates “dissaving” from late 2021 onwards — spending out of income at a greater rate than previously.

People have called into question the presence and nature of these excess savings; but it’s pretty evident looking at consumer bank deposits that households saw a large increase in liquid cash, which they have slowly started to spend down/move into higher interest rate products over time.

So while some degree of stimulus was helpful to maintain incomes; the incremental increase in aggregate demand may have simply further boosted inflation concenrated in goods. As a result, ultimately increasing interest rates (at least to a more neutral position) may be part of a helpful process of normalizing inflation.

There are a couple of refinements here that are interesting. First, Justin Bloesch has argued that Phillips Curve thinking works if you use the quit rate as your measure of slack against wage growth. So imagine here that you have particularly tight labor market activity in the really active part of the labor market in 2021-2022, and things moderate a bit by 2023 (either because sectoral demand shifts into other sectors of the economy, or aggregate demand falls a bit). That gets you rapid wage growth in 2021-2022, also fueling inflation, which lowers by 2023.

Similarly, Pierpaolo Benigno and Gauti Eggertsson argue for a non-linear Phillips Curve. As labor market tightness gets particularly high; inflation really starts to shoot up. I think you could imagine labor market tightness here as a product of: 1) high labor demand in goods and other industries which remained operational during the pandemic; 2) low labor supply due to pandemic risk and other concerns; and 3) potentially a proxy for low “slack” in the economy overall, including also the goods side of the economy. The nonlinearity means that further boosts to economic activity in the inelastic part really push up inflation; but at the same time, incrementally cooling down the economy can achieve an “immaculate disinflation.” Benigno and Eggertsson also attempt to measure supply shocks — which they proxy through a few different methods (difference between headline and core CPI/PCE; and differences between import prices and the GDP deflator); and don’t find these have been particularly large recently by historical standards.

We can contrast these views with other aspects of “mainstream” macro thinking here which involves incorporating expectations about inflation, which don’t seem to fully fit this situation. Rather than inflation expectations leading the rise in inflation; instead you see expectations generally lagging contemporary shifts in inflation on the ground. The risk in the models is that inflation gets entrenched somehow, requiring even more proportionate increases in interest rates to re-anchor expectations.

However, it’s not obvious that sort of story even fits the 1970s; or the present day. Inflation seems more a product of conditions on the ground at the moment, rather than the more complicated forward-looking story. If anything, one stabilizing force might be that when faced with high and uncertain inflation, many people seem to respond on surveys as being fairly pessimistic in their economic attitudes. To the extent that influences consumption; that may lead them to actually lower their spending today, rather than accelerating it forward (which they would do if they really thought inflation would be higher in the future). Which is to say it seems there may be some stabilizing dynamics with respect to inflation, rather than the destabilizing ones that macroeconomists typically assume.

Another miss here is thinking about the labor market: Larry Summers and others were ultimately quite pessimistic about the labor supply side, and this was a key reason behind their prediction that the economy needed a recession (no way to cool off the overheated job market absent higher unemployment). This follows a lot of pessimism, in the 2010s, about the marginal productivity of laid off workers. Instead, it turns out the labor market was able absorb many additional workers, both in the 2010s and from 2021 onward, particularly into some of these “rotated” sectors.

Conclusion

The pandemic period of inflation deserves to be considered in its own right; not as a simple restatement of previous trends. With the benefit of hindsight; it seems like the economy wasn’t in as much risk of remaining either stagnation in demand (as it was in the 2010s) or in persistent inflation (as in the 1970s). We experienced a fairly historically unique set of circumstances resulting in a sectoral reallocation of spending towards goods and then back to services, and this pattern can explain a lot of the features we saw in the last few years.

Subscribe now