Take three snapshots of a share’s price chart around a results announcement. The first is the moment a report far ahead of expectations lands, before the market has processed it. The news exists, but the price still reflects the old information. The second spans the weeks that follow. The price jumps at the open without finishing the job. The shares keep travelling with the surprise as estimates and slower investors catch up. The third is a year later: the surprise is ancient, its implications priced, and the chart wanders on no news.

Ask in which snapshot the price itself holds a pattern an active investor can trade. Not the first: the opportunity there is in reading the report, before any pattern has reached the price. Not the third: whatever was knowable has been incorporated. Only the middle holds a pattern, and only in passing. The pattern is the price absorbing the news, and absorption ends.

What kind of problem is this?

It looks like a fact about one anomaly. The shape underneath is more general. There are two instincts about where opportunity lives. One prizes the finished end: the fully efficient market, where randomness is the sign of a job well done. The other prizes the untouched end: the report that has landed and not yet been read. The three snapshots show that both miss the same thing. The pattern peaks in the middle. Saying anything precise about that needs a way to measure structure. The disciplines that have thought hardest about that measurement are algorithmic information theory and the physics of complexity.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

The lens we are borrowing is called apparent complexity. The physicist Sean Carroll built it with Scott Aaronson and Lauren Ouellette. It puts a number on how interesting a system looks: the structure it shows at the scale an observer sees. The recipe has two steps: blur the picture, then measure how hard the remainder is to describe.

The Static And The Swirl

Carroll’s demonstration is a thought experiment that repeats the three snapshots. Photograph a cup of coffee at three moments. First, the cream sits in an untouched layer on the black coffee. Half a minute after the spoon, the cup is a tangle of pale tendrils and dark channels. A few minutes later, it is a uniform light brown. Suppose the three photographs are idealised: same resolution, same format, no camera noise. Compare their file sizes. The first and third compress readily. Compression exploits repetition, and those photographs are mostly uniform regions with short descriptions. The middle file resists. Every tendril differs from its neighbours, so every part of the picture needs its own description.

The measuring stick comes from the Russian mathematician Andrey Kolmogorov, who had already given probability its modern axioms. In 1965 his paper gave the definition: an object’s complexity is the length of the shortest program that produces it. A string of a billion zeros is enormous but simple: “print zero a billion times” generates it in a line. A string of a billion random digits admits no such shortcut. The shortest program is no shorter than the string itself. File size is a working stand-in, because compression hunts for exactly those shortcuts.

The blur is there because of what happens without it. By Kolmogorov’s count, the random string is the most complex thing there is. To any observer it is just noise. Zoom in far enough and every cup is that string: at molecular grain, all three photographs are static. So blur the image first, averaging away the molecular detail. Then compress what remains and read the file size. Random static now blurs to uniform grey and compresses away. The layered cup and the blended cup stay short, as before. Only the half-mixed cup resists, because its tendrils survive the blur. The measure now agrees with the eye: structure scores high, while order and noise score low.

Three photographs are only three points. Build a simulation of the cup and let it mix, and you will see the whole curve. Entropy, the physicist’s measure of disorder, rises from start to finish. Apparent complexity climbs as the tendrils form, peaks around the middle, and falls back as the cup blends. Disorder only accumulates; structure rises and dies.

Before leaving the cup, take two properties of the measure with you. First, the measure reads the present, not the peak. Nothing in that moment says it is the top of the arc. Settling that takes a model of what remains, or the rest of the run. Second, the answer depends on the blur. A different coarse-graining gives a different complexity, and there is no blur-free view.

The Blended Market

In 1965, the same year as Kolmogorov’s paper, Paul Samuelson proved that properly anticipated prices fluctuate randomly. Within his model, earlier price changes carry no forecast of the next one. Anything they did forecast would already have been traded into the price. Once a piece of information has been absorbed, its price trace offers no shortcut to the moves that remain. To the trader, that resembles incompressibility, though the mathematics differ. For everything already known, the textbook efficient market is the fully blended cup. The third snapshot photographed it.

The first snapshot was the unmixed layer: the filing nobody has read, the asset nobody has repriced. The middle snapshot is the canonical tendril, the post-earnings announcement drift. Ray Ball and Philip Brown first observed it in 1968. Studies since have measured it over subsequent months. The news is public and the price is still travelling.

Other tendrils trace the same arc, each for its own reason. Momentum lasts while incorporation is slow. A valuation spread lasts while capital is constrained. A merger spread lasts until the deal’s uncertainty resolves. The cup lends the arc, not the number. Apparent complexity measures description length, and a profitable pattern can score low on it. The value is alpha: the extra return a structure pays, measured against a benchmark and adjusted for risk. Only what survives implementation costs is investable.

Individual tendrils get exhausted, yet the market as a whole never finishes blending. Part of the reason is fresh cream: news and flows keep pouring in. The deeper reason came from Sanford Grossman and Joseph Stiglitz in 1980. In their model, information costs money to acquire. If prices fully revealed what the informed had paid to learn, nobody would pay to learn it. Then prices would reveal nothing. So a perfectly efficient market is impossible on its own terms. Markets settle instead at what they called an equilibrium degree of disequilibrium. The name is paradoxical, but it describes an equilibrium. The cup must stay partly unmixed to pay for its own stirring. The investors who pay to be informed are the spoon.

The spoon erases mispricing. Where a tendril is mispriced and tradeable at scale, each position against it chips away the return it promises. In the cup’s terms, trading shortens the description of the opportunity, not of the price series itself. Crowding can even swell a tendril before smoothing it. This is why any backtest shows patterns as they stood before you arrived. And no present moment certifies that an edge has peaked. You learn where it was by watching the returns fade, with capital already committed.

Living In The Middle

The blur is the deeper of the two properties, because in a market the blur is yours to choose. Every description of a market blurs it. A factor model does. So do a sector label, a monthly return, a price chart. And what you can see depends on the blur you chose. A £100 million position unwinding through a thin book leaves a wake that lasts minutes. In quarterly data it does not exist. A slow rotation between regimes takes decades. On a trading screen it never shows. When two investors argue about whether a market is efficient, they are often describing it at different blurs.

The choice of blur carries two hazards. The first is invention: a blur can add structure that was never in the process. The Trend The Gaps Invent showed how assets observed too rarely arrive pre-smoothed. The simulation of the cup gave the same warning. The builders’ first blur put a complexity hump into a control cup where nothing interacted. They had to replace it. The second hazard is luck: genuine noise sometimes compresses by chance. The standard is survival out of sample.

Three implications follow. First, budget for the fade. Each tendril’s return should diminish as its cause is used up, and your own trading is part of the using. A strategy that keeps finding new tendrils need not fade with any one of them. Second, search where the mixing is. Structure concentrates where information or flow is in transit: new instruments finding their clientele, forced sellers, calendar flows. It thins towards the most watched corners of any market once costs are counted. Third, audit the blur. Why One Model Is Never Enough argued for holding many models. The same case holds for resolutions.

The Mixing Question

The pattern the mixing makes, then, is one kind of structure the market can pay an active investor to find. Where the structure is a mispricing, trading it helps finish the job.

The same arc turns up beyond markets, though as a rule of thumb rather than a law. A scientific field is most fertile after the founding insight and before the textbooks close it. A technology yields its richest variety between invention and standardisation. An organisation is most adaptable between the garage and the procedure manual.

Knowing the arc does not place you outside it, because every trade against a tendril is part of the blending. An edge made by the mixing is a wasting asset. The question is never whether it will be smoothed away, only how much of the smoothing you are paid for.

The discipline is in asking: not whether a pattern exists, but where in its arc you have met it, at what resolution it is visible, and what your own use of it will do to what remains?

References & Further Reading

1. Three Approaches to the Quantitative Definition of Information: Andrey Kolmogorov

2. Quantifying the Rise and Fall of Complexity in Closed Systems: The Coffee Automaton: Scott Aaronson, Sean M. Carroll & Lauren Ouellette

3. The Big Picture: On the Origins of Life, Meaning, and the Universe Itself: Sean Carroll

4. Proof That Properly Anticipated Prices Fluctuate Randomly: Paul Samuelson

5. An Empirical Evaluation of Accounting Income Numbers: Ray Ball & Philip Brown

6. On the Impossibility of Informationally Efficient Markets: Sanford Grossman & Joseph Stiglitz

In the session of 1846, Westminster passed 272 Acts authorising roughly 4,500 miles of new railway. Across the peak sessions of 1844 to 1846, promoters won approval for some 8,000 miles; a decade earlier the operating network was a few hundred. They issued the shares partly paid: a deposit of five or ten per cent secured the certificate, and the company could call the remainder as construction proceeded. By modern estimates, railway construction absorbed close to seven per cent of national income at its 1847 peak.

An index of railway shares roughly doubled between 1843 and the summer of 1845; by 1850 it had surrendered the whole advance and more, to around two thirds below the peak. The companies made their calls into the falling market, so many holders owed fresh instalments on shares worth less than the sums paid. George Hudson, the “Railway King”, controlled more than a quarter of the network, and his dividends proved to have been paid partly out of capital. He was disgraced and driven from his chairmanships in 1849, and the claims that followed completed his ruin.

Yet by 1850 more than 6,000 miles of railway were open in Britain, the greater part built during the convulsion. Within a generation the network had opened national markets for fresh food, collapsed the cost of moving coal and people, and pushed the country onto a single railway time.

The verdict on the 1840s is therefore double: one of the great destructions of investor capital in British history, and one of the great creations of economic value. A cash-flow view says the subscribers mispriced the assets, and they did. A behavioural view says the crowd was carried away, and the shareholder records complicate it: the subscribers included the informed and the experienced, and the insiders lost alongside the outsiders. Both treat the investors’ outcome as the whole story, and both leave the same fact unexplained: the losses and the value creation came out of the same event.

What kind of problem is this?

It is tempting to file the episode under error, but the record resists it: the canal promotions of the 1790s ran the same course, electrification ran it again, and so did the fibre-optic construction of the late 1990s. Hyman Minsky described how stability breeds the credit structures that undo it, and his mechanics explain how any of these run-ups builds. They do not explain why the largest gather around new technologies, nor why the wreckage often has a golden age on the far side. A single episode can be read as a lapse. A repetition has to be read as a process: the same course every half-century or so, always attached to a transforming technology, always pairing private loss with public gain.

When what matters is how a new technology and the capital that funds it move through an economy, in phases, with value changing hands as they go, that is a specific shape of problem: diffusion. And the discipline that has thought most rigorously about the diffusion of technological revolutions is neo-Schumpeterian economics, the tradition that descends from Schumpeter’s account of creative destruction.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

The discipline’s evidence is historical: five great surges reconstructed from two centuries of record. The structural logic transfers: two kinds of capital move through an ordered sequence of phases, and value migrates between them.

Two Capitals, Two Clocks

The framework comes from Carlota Perez, who set it out in Technological Revolutions and Financial Capital (2002). She identifies five technological revolutions since 1771: water-powered mechanisation; steam and railways from 1829; steel, electricity and heavy engineering from 1875; oil, the automobile and mass production from 1908; and the information age, opened by the microprocessor, from 1971. Each is a cluster of technologies plus a techno-economic paradigm, a new common sense about best practice, and she argues the paradigm matters more.

Each surge, on her reconstruction, runs the same sequence. An installation period opens with irruption, when the new technologies appear amid the decline of the old paradigm. It builds to a frenzy, her term for the phase in which finance takes command. A turning point follows: collapse, recession, scandal, and the institutional recomposition they force. Then comes a deployment period. It opens in the phase she calls synergy, when the paradigm spreads across the whole economy, and ends in maturity, when the paradigm’s potential nears exhaustion and restless capital goes hunting for the next irruption.

Her distinction between two kinds of capital drives the sequence. Financial capital is mobile and uncommitted: it holds claims and can leave. Production capital is embodied: it lives in particular firms, plants and knowledge, and it cannot exit without ceasing to be what it is. During installation, financial capital is in command, and the frenzy is what command looks like: money pours into new infrastructure at prices detached from any defensible projection of cash flows.

Perez’s claim is that the detachment is functional. Prices that had stopped doing arithmetic routed seven per cent of Britain’s national income into iron rails within a few years, at a pace and on a scale that sober valuation was unlikely to match. The frenzy funds parallel experiments, most of which fail. It installs capacity years ahead of the demand that justifies it, and it breaks the commercial and political grip of the paradigm being displaced. Functional does not mean necessary or efficient: much of the loss was plain waste, and the capacity might have been built more slowly with less ruin. The turning point then demands an institutional catch-up: in Britain, the accounting scrutiny and reform that followed Hudson’s exposure and the standards that a national network required. Where the catch-up succeeds, command passes to production capital, which spends the deployment period making the installed base earn.

Owned Twice

The framework first separates two things the 1845 subscriber had fused: the return to the technology and the return to its financier. The railway was among the most productive technologies in British history, and railway shares bought in 1845 were among the worst investments of the century. Both hold because the value went to other parties: to users, through cheaper transport, and to the operators of the deployment period, who bought capacity below its construction cost. Infrastructure priced in a frenzy tends to be owned twice: once, expensively, by the financiers who build it, and once, cheaply, by the consolidators who run it.

The Room That Runs Out described carrying capacity: capital that crowds into a niche compresses the returns that attracted it. The inflow that validates an opportunity is the inflow that consumes it. The frenzy is that dynamic at national scale, with one difference: the overshoot leaves physical infrastructure standing for someone else to earn from. And the ergodicity problem from The Path The Maths Misses applies here. The aggregate gain describes no holder’s experience: each subscriber lived a single path. The partly paid structure meant many paths hit an absorbing barrier — a call that could not be met — years before deployment arrived to pay anyone. Society could afford to wait for the golden age. A holder facing a call in 1848 could not.

What The Price Is Doing

The framework implies a second separation: what a price is doing changes with the phase. In deployment, a price is mostly a forecast, a claim about the cash an asset will produce. In installation, a price is partly a recruiting instrument: it is how an economy commits itself to one paradigm rather than another. Frenzy pricing is the beauty contest of What Others Think Others Think: the operative question is not what the asset will earn but what the next subscriber will pay. The elevated price let promoters float new schemes and fill subscriptions to partly paid shares; the calls that followed drew on those contracts. As a forecast of railway cash flows, the market failed. As a device for mobilising an unprecedented share of national saving, it worked. Only one of those jobs pays the holder.

Installation rewarded access and exit; deployment rewarded operations and consolidation. Why Good Strategies Stop Working gave the general mechanism: a strategy is fit for an environment, and environments move. Perez adds a timetable: the environment does not drift at random but advances through an ordered sequence, so the competence one phase rewards becomes a liability in the next.

The Mobilisation Question

None of this yields a timing rule, nor a licence to sit installations out: investors made fortunes inside them. It changes the questions. If the technology succeeds, where does the value settle after the turning point: with surviving equity, successor owners, suppliers, landowners or users? The technology thesis and the security thesis feel identical during a frenzy and come apart at the turning point.

For institutions, the lesson is liability structure. The 1845 subscriber’s problem was not merely horizon. Many intended to hold for decades and believed, correctly, in the railway. The problem was the call: a capital structure that could demand cash at the moment cash was dearest. The financial value of what the frenzy built goes to whichever claims are still attached to the assets after the turning point. Survivorship turns less on conviction than on the structure of those claims: whether capital can be called, margined or redeemed at the trough. Patience is a property of liabilities before it is a virtue of investors.

The framework extends beyond finance. In organisations, a new technology arrives the same way: pilots install capability ahead of any use that justifies it, and the value appears when a duller consolidating phase makes the installed base earn. In careers, promotion in one phase selects the habits the next phase retires.

A scheme with four named phases and a turning point invites you to date the present, and the date is the one thing it cannot supply in real time. Knowing that frenzies finance futures does not reveal which job the surge in front of you is doing; it reveals that the question exists. Understanding the sequence does not exempt you from your position in it.

The discipline is in asking: is this a forecast or a mobilisation, and who inherits what it installs?

References & Further Reading

1. Technological Revolutions and Financial Capital: Carlota Perez

2. As Time Goes By: From the Industrial Revolutions to the Information Revolution: Chris Freeman & Francisco Louçã

3. Collective Hallucinations and Inefficient Markets: The British Railway Mania of the 1840s: Andrew Odlyzko

4. Dispelling the Myth of the Naive Investor during the British Railway Mania, 1845–46: Gareth Campbell & John D. Turner

5. Stabilizing an Unstable Economy: Hyman P. Minsky

6. Capitalism, Socialism and Democracy: Joseph A. Schumpeter

By 1987, an estimated sixty to ninety billion dollars of institutional assets was covered by portfolio insurance, a strategy built by Hayne Leland and Mark Rubinstein on the option-pricing mathematics of the 1970s. It aimed to give a portfolio the downside protection of a put option, without buying one, by selling stock index futures as prices fell and buying them back as they rose. In the models, it worked.

The protection did not rest on a pattern in past data; it rested on assumptions that held only while prices moved smoothly, trading stayed cheap, and the fund was small enough to trade without moving prices. Putting real portfolios behind it meant relying on those assumptions in a different setting: a crowded market where many funds ran the same rule at once, and the selling itself moved prices and dried up liquidity. One fund running the rule and many funds running it are not the same intervention.

On 19 October 1987 that is exactly what happened. As prices fell, the models called for heavy selling; that selling deepened the decline, and lower prices called for still more. Official investigations treated that procyclical selling as an important amplifier of the crash, though its contribution remains disputed. All three assumptions failed together, and the protection proved much weaker and less reliable than clients had expected.

What kind of problem is this?

When the question you can answer and the question you need answered sit at different levels, and ordinary language lets you slide between them without noticing, that is a specific shape of problem. One is about observation: what goes with what, the way this signal accompanied that return. The other is about action: what happens when we deploy capital, and what would have happened had we chosen otherwise. And the discipline that has spent the last three decades making those levels precise is causal inference.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

The framework comes from Judea Pearl, the computer scientist whose work on probabilistic and causal reasoning reshaped artificial intelligence and was recognised with the 2011 Turing Award.

Seeing, Doing, Imagining

Pearl sorts causal questions into three levels, or rungs on a ladder. The first rung is association, or seeing. It asks: what does observing X tell me about Y? This is the rung of correlation, prediction, and curve-fitting, where much of predictive statistics and standard machine learning operates. The data here are, in Pearl’s phrase, profoundly dumb about causes.

The second rung is intervention, or doing. It asks: what happens to Y if I set X by deliberate action? Pearl gives this its own notation, the do-operator, because doing X is not the same as seeing X. His example is the barometer: the needle and the storm move together, but turning the needle by hand does nothing to the weather, because the correlation runs through atmospheric pressure your hand never touches. To answer a doing-question you need causal assumptions or an experiment, not observation alone.

The third rung is the counterfactual, or imagining. It asks: given what happened, what would have happened had I acted otherwise? This is the most demanding rung, of regret and attribution; answering it requires a structural causal model, or structural assumptions rich enough to identify the counterfactual. “Would the patient have recovered had we withheld the drug, given that she received it and recovered?” is a harder question than “does the drug work on average?”

Pearl’s formal point, sharpened by Elias Bareinboim and colleagues in the Causal Hierarchy Theorem, is that the ladder is a ladder, not a ramp: in general, no summary of lower-rung data settles a higher-rung claim on its own. To climb, you must add causal assumptions.

Where Our Evidence Sits

Backtests do not have a rung; the claims made from them do. A regression of returns on a signal is associational. A historical replay is a model-based policy simulation, counterfactual only when it asks what this same realised history would have looked like under a rule we did not run. Either way, replaying fixed historical prices does not by itself identify what deploying the rule will cause. The allocation question is interventional: what distribution of outcomes follows if we deploy this policy at a given size, cost, and market state?

Absent experimental or live-deployment evidence, we answer that question with rung-one evidence plus causal assumptions: that the observed relationship is policy-relevant, that the mechanism or latent state it tracks remains stable, and that the relationship survives our acting on it.

A confounder is the familiar worry: a signal predicts returns, but a third, unobserved factor, such as a shared funding condition, drives both. In causal inference that is fatal to the naive inference, though sometimes identifiable by other means. For an investor it need not be fatal: conditioning on the signal still selects exposure to the latent state that drives returns. The signal is a dial, not a lever, useful while it stays informative, not a cause you operate.

Counterfactual attribution and alternative-history post-mortems belong on rung three. Arithmetic attribution is an accounting decomposition, not a causal query, and belongs on no rung. “What would our worst year have looked like had we cut duration that January?” is a counterfactual about a single realised path; the alternative never ran.

When Doing Rewrites Seeing

Textbook medical examples treat the mechanism as stable enough that a trial reveals it rather than changing it. Markets are less forgiving: the mechanism that generates returns often includes beliefs about that mechanism. Intervening on a trading rule need not freeze the rest of a structural model; prices, beliefs, liquidity, and participation may all respond downstream without any mechanism breaking.

The harder point is that a trade by one small fund and the same rule run by many large funds are different interventions, and the second may run under different market mechanisms. So the intervention must be specified precisely: size, timing, execution, publicity, and how many others run it, with the feedback it sets off modelled as a consequence, not folded into the action. Markets do not suspend the Pearlian lesson; they make it harder, because the stability an effect needs to carry across settings is the first thing trading at scale disturbs. The real question is whether an effect found at small scale survives large-scale deployment at all.

This is the same reflexivity the series has met before. It is the strategy decay of Why Good Strategies Stop Working, where a rule’s fit erodes as its environment adapts, and the active inference of The Model That Fights Back, where the system updates on your prediction. Goodhart’s law names one way this happens. Charles Goodhart, writing about monetary policy in 1975, observed that any statistical regularity tends to collapse once it is used as a target for control. The ladder does not imply this on its own: seeing and doing would differ even if a pattern survived being acted on. In practice patterns often do not survive, and Goodhart names one common reason. Once people know you are trading on a rule, they adjust to it, and the pattern you measured changes.

Labelling the Climb

The first move is to label evidence by its rung. When someone shows you a backtest, ask whether it claims the pattern held in the past, which is rung one, or that it will hold when you trade on it, which is rung two. The presentation usually blurs the two, and separating them is most of the work.

The second is to state the causal assumptions that take you from rung one to rung two. Those assumptions carry the decision. Pearl’s other tool is the causal diagram: which variables you think cause which, and which shared causes you allow for or rule out. At a minimum, write down what you think the signal is tracking, whether it causes returns or shares a hidden driver, and what must hold for the strategy to keep working once you trade on it. In a market that reacts to you, the diagram needs to include time: this period’s trade affects next period’s prices and beliefs, which a static diagram leaves out.

The third is to treat counterfactual attribution as what it is: a claim about a history that did not happen. Rather than “cutting duration would have saved four hundred basis points,” say “on our model of how the positions interact, it would have saved roughly that much.”

The fourth follows from the same point: because acting is itself an intervention, capacity, size and decay are not details to add later; they are part of specifying the intervention, and of whether an effect found in one setting transports to another. An effect measured while trading small need not survive market-moving deployment, even though both remain rung-two questions, because trading at that scale changes the system you measured it in.

The Intervention Question

When does this lens apply? Whenever you reach from an association to a claim about what acting will do, or would have done. That is most of investing, and the shape recurs far beyond it. A doctor infers that a drug speeds recovery because treated patients recovered faster, when the milder cases may have been the ones treated. A restaurateur adds live music because the busiest nights have it, when weekends bring both the crowds and the band. A club credits a new coach with its turnaround because results improved soon after the appointment, when a bad run tends to end on its own. Each reaches from what accompanied what to what an action will bring about, and the association never settles the action on its own.

Knowing the ladder does not lift you off it. What it does is more modest: it tells you when you are climbing from one rung to the next, so you can be honest about the assumptions carrying you up and size your conviction to them, not the bare correlation. And this applies to you as much as to anyone: noticing that acting changes the system does not exempt your acting from changing it.

The discipline is in asking: is this evidence about what I have seen, or a claim about what my doing will cause, and what must be true of the world for the one to license the other?

References & Further Reading

1. The Book of Why: The New Science of Cause and Effect: Judea Pearl & Dana Mackenzie

2. Causality: Models, Reasoning, and Inference: Judea Pearl

3. Causal Inference in Statistics: A Primer: Judea Pearl, Madelyn Glymour & Nicholas P. Jewell

4. The Seven Tools of Causal Inference, with Reflections on Machine Learning: Judea Pearl

5. On Pearl’s Hierarchy and the Foundations of Causal Inference: Elias Bareinboim, Juan D. Correa, Duligur Ibeling & Thomas Icard

6. Problems of Monetary Management: The U.K. Experience: Charles Goodhart

A company you follow trades at £40 a share. Rather than ask what the business is worth, you run a discounted cash flow the other way: you hold the price fixed and solve for the future it assumes. The note writes it up cleanly: at £40, the market is pricing in roughly 11 per cent annual growth for a decade. It is a satisfying sentence, and on its own close to meaningless: the 11 per cent came from holding everything else still. Let the steady-state margin settle a point lower, or the advantage fade a little faster, and the same £40 implies 14 per cent, or 9, or a different number. The price does not choose one—it is consistent with a whole family of futures, and the figure in the note is one member an analyst selected, not one the price singled out.

This is the reverse discounted cash flow, among the most useful disciplines an investor can adopt. Alfred Rappaport and Michael Mauboussin called it reading the market’s implied expectations: rather than produce a valuation and compare, you take the price as given and ask what must be true for it to make sense. The technique is sound; the difficulty is in how its result is reported, because run backwards honestly a valuation does not return a number. It returns a set, and the shape of that set is the information.

What kind of problem is this?

Working backwards from an effect to its cause is a distinct and recognisable shape of problem. The forward direction is easy: assumptions about a company, pushed through a discounted cash flow, yield a price. The backward direction is the one we want and the one that misbehaves: given the price, which assumptions produced it? Scientists call the first a forward problem and the second an inverse problem; the inverse is almost always more treacherous, because the same effect can be produced by many causes. And the discipline that has thought most rigorously about this asymmetry is the mathematics of inverse problems.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Jacques Hadamard, one of the foremost French mathematicians of his era, reached this problem not through markets but through the differential equations of physics. His essential point was that inverting a result is not the mirror image of producing it: the backward direction may have many solutions, or none, and pinning it to one means importing an assumption the result does not contain.

What The Data Cannot Settle

In 1902 Hadamard set out three conditions a problem must satisfy to be well-posed. A solution must exist, it must be unique, and it must depend continuously on the data, so that a small error in the measured effect produces only a small error in the inferred cause. Many canonical forward problems are well-posed by construction; their inverse formulations frequently violate all three. Hadamard called such problems ill-posed, a description rather than a complaint: the data alone do not determine the answer.

Each failure has its own character. Existence can fail: no cause inside your model can produce the observed effect. Uniqueness can fail, the deepest of the three: many different causes produce the identical observation, and no amount of staring at the data says which one is responsible. When it does, the solution is not a single point but a set, the collection of causes consistent with what you measured. Stability can fail: a tiny perturbation in the observation throws the inferred cause a long way, so that measurement error is amplified rather than absorbed.

If the data do not determine the answer, how does anyone proceed? The standard tool is regularisation, formalised by Andrey Tikhonov. You cannot make an ill-posed problem well-posed from the data alone, so you add a penalty that encodes structure the data cannot supply: a preference for smaller or smoother solutions, a prior about what is plausible. Regularisation replaces an unstable or underdetermined inversion with a rule for choosing a stable estimate, trading fit to the data against the penalty. It buys a unique, stable answer at a cost: the answer now depends on the assumption you imposed.

The Many Behind The One

Return to the £40 share and read the reverse discounted cash flow through Hadamard’s three conditions. The forward calculation turns a full set of assumptions into a single price: growth, margins, the discount rate, and the rest. The reverse DCF runs the other way, from the price back to those assumptions. Hadamard tells you what to expect.

Uniqueness fails first. The solution is not the growth rate the price implies. It is a set: every combination of growth, margin, fade, reinvestment and duration that produces £40. It traces out a surface in the space of assumptions, a manifold where that surface is smooth, though in practice it may be kinked, bounded, or in places empty. A company looks inexpensive on the growth-and-margin slice and fully valued once you allow a faster fade or lower return on reinvested capital. The single implied-growth figure in a research note is one point from that set: someone fixed margin, fade and duration to make it unique, and the choosing was done by assumption, not by the price. The shape of the set, which combinations the price permits and which it forbids, is the information, exactly what the point estimate discards. This is the within-a-single-model echo of Why One Model Is Never Enough: even one valuation model, run backwards, hands you a population of admissible worlds, not one.

Existence fails next. A discounted cash flow can hit almost any price if you allow unbounded growth, a vanishing discount rate, or an unending advantage, so the question is whether any combination inside a defensible range reproduces it. When none does, the price sits outside the model’s reachable range, often above its ceiling, a finding rather than a defect to be patched: the market is pricing something the model cannot represent, a longer advantage, a lower discount rate, an option it omits. The disciplined response is to measure how far outside the reachable set the price sits, not torture the inputs into one. An empty set is itself information: the distance between the price and your model’s vocabulary.

Stability fails last. Most of the value lives in the terminal period, where the price barely pins the inputs down: a small change in the discount or fade rate swings the implied growth, so the inversion is least stable in the region that dominates the answer. This echoes The Question Shapes The Answer: entropy is defined relative to a chosen description, and an inverse solution is fixed relative to the assumptions used to identify it. The figure you recover depends on what you held fixed before you began.

The One-Way Summary

Even when a sensible range of assumptions can reproduce the price, the reverse DCF returns not one answer but the whole set of them. Some backward questions are worse than that, because the forward calculation throws information away before you ever run it backwards, and the internal rate of return is the clearest everyday case. A conventional cash-flow stream can usually be summed up by a reported internal rate of return, a rate at which the investment breaks even in today’s money, but the summary loses information. Two investments can show the same rate while one hands back several times the capital invested and the other barely more than it, because the rate says nothing about how much money came back, or for how long it was at work.

So you cannot run the rate backwards to recover the wealth a stream created: too many different outcomes share it. The rate is a real feature of the cash flows, but it is not the actual rate an investor’s money grew at once cash returns along the way and has to be put to work again, and reading it as that growth adds an assumption the rate itself does not make. When the number you want cannot be recovered by inverting the rate, the answer is to measure something else, not to invert harder: how much money came back in total, over what period, and what that was worth against the alternative you actually had.

Where you can, report the whole set of answers, not just one. The set of assumptions the price is consistent with, the bounds around an estimate: these are more honest answers, and they sharpen disagreement rather than blur it. Charles Manski built a large part of modern econometrics on exactly this idea, which he calls partial identification: when the data cannot pin a quantity down to a single value, report the set of values they do support, and say plainly how wide it is. An investor who reports that set, together with the assumption used to pick one number out of it, gives a reader something to check. One who reports only the single number hides the assumption the whole conclusion rests on.

The Inversion Question

The crowd the price hides, then, is not a literal distribution of investor beliefs, nor anything readable off an order book; it is the set of admissible worlds one clearing price leaves open. Knowing the inversion is ill-posed does not place you outside it. You must still choose one growth path, one set of assumptions to hold fixed, and quoting a single implied figure imposes a prior, admit it or not. What the discipline offers is not an escape from that choice but the obligation to make it in the open: to say what you fixed and why, and how much the price leaves undetermined.

The shape recurs far beyond markets. In medicine, a diagnosis read from symptoms has the same inverse-problem shape: several conditions fit the same presentation, so a differential lists the candidates before judgement, priors, and further tests collapse them. In science, a model fitted to data meets the same underdetermination, and overfitting often appears as instability: a fit that lurches when the data shift. In any inquiry that reconstructs a cause from the traces it left, the observation constrains the cause without fixing it, and a single confident answer is a prior in the costume of a measurement.

The discipline is in asking: not what answer the data imply, but what set they leave open, and what you fixed to collapse it to one?

References & Further Reading

1. Lectures on Cauchy’s Problem in Linear Partial Differential Equations: Jacques Hadamard

2. Parameter Estimation and Inverse Problems: Richard C. Aster, Brian Borchers & Clifford H. Thurber

3. Partial Identification of Probability Distributions: Charles F. Manski

4. Expectations Investing: Alfred Rappaport & Michael J. Mauboussin

An investment team has a bad year. They reconstruct it and find real faults: position sizing too aggressive into the drawdown, stop discipline that slipped, a correlation they had trusted that moved against them. So they tighten the sizing, formalise the stops, add the correlation to the risk dashboard. Every fix is sensible. Three years later, in a different market, the same shape of loss arrives again.

The puzzle is that there is nothing wrong with the team. Yet the second loss repeats the shape of the first, the one thing the review never touched, and there is reason to think sustained success makes that pattern worse, not better.

What kind of problem is this?

It is tempting to call it a competence problem, but that does not fit the evidence: the team corrected the errors it found. It is a learning problem, of a particular kind the word usually hides. There are two quite different things you can learn from a loss. You can run the existing strategy better, fixing the mistakes that happen within its rules. Or you can ask whether the strategy’s governing assumptions still fit the world it is trading in. The question is not why a few people are careless, but why capable people so often run their learning on the first track and so rarely on the second.

When a failure is a consistent pattern in which one level is examined and another is not, rather than a lack of effort, the problem has a definite shape: the architecture of learning. And the discipline that has studied it, beginning in the mid-1970s, is the theory of organisational learning.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Its central distinction comes from Chris Argyris and Donald Schön, who developed it across work culminating in their 1978 treatment of organisational learning, and spent the following decades studying what gets in its way.

The Defences That Hold

Argyris and Schön drew a line between two kinds of error-correction. Single-loop learning detects a mismatch between intention and result and corrects it by adjusting behaviour, leaving the underlying goals and assumptions intact. Their own illustration was a thermostat: it senses the room is below the set temperature and turns on the heat, without asking whether the setting is right. Double-loop learning is the second question. It examines the governing variables themselves, the goals and assumptions that define what counts as an error, and is willing to change them.

In the cases Argyris and Schön studied, the obstacle was often not ignorance but defence. They separated two things: the espoused theory, the reasons a person gives for a decision, and the theory-in-use, the one their behaviour actually reveals. The two are often different, and the gap is usually unspoken, and may be hidden even from the person themselves. Under threat, people fall back on the same theory-in-use, which they called Model I: keep control, win rather than lose, hold difficult feelings down, and stay outwardly rational. Model I breeds defensive routines, the shared habits that spare people the discomfort of questioning their own assumptions. Certain questions become undiscussable, and then the fact that they are undiscussable becomes undiscussable too. The moves themselves can be quite deliberate, concealing, controlling, explaining things away, even while the pattern behind them stays out of sight. So the second loop ends up closed without anyone deciding to close it, by people who believe they are simply being rigorous.

The alternative they set out is Model II. The aim is not endless doubt, or shutting down anyone who argues a case, but good information, real choice, and genuine commitment to what is decided. People show the reasoning behind their views so others can test it, and they make their case while honestly inviting challenge to it. So double-loop learning depends not just on which questions get asked, but on whether the reasoning behind the answers can itself be examined.

Argyris’s most uncomfortable finding followed from this, in a provocatively titled 1991 essay, “Teaching Smart People How to Learn.” The people worst at double-loop learning, he argued, are often the most successful and the most expert. They have built careers on getting things right and have rarely failed, so they have had little practice at what double-loop learning asks: treating their own mistakes as information, not as a threat to a hard-won identity. Faced with evidence that a governing assumption was wrong, the highly capable do not always get curious. They can get defensive, pin the cause on something outside themselves, and shut the inquiry down so smoothly that the defence looks like analysis.

Clever on the Inner Loop

In an investment process you can tell the two loops apart, though not as neatly as you might hope. Single-loop learning is everything you do to run a strategy better without changing its assumptions: tuning parameters, tightening limits, refining execution and sizing. Double-loop learning asks the harder question, the one that closed Why Good Strategies Stop Working, aimed at the strategy itself: what does it assume, and does that still hold? The governing variables are wider than beliefs about the market regime. They include the firm’s risk appetite, the liquidity it has promised clients, the incentives it pays, and what it treats as success. Where the boundary falls also depends on where you stand: swapping out a failing strategy can be double-loop at the desk and still leave the firm’s mandate and incentives untouched.

Long-Term Capital Management is a telling case. Official reviews found that its risk framework, and its counterparties’, underestimated joint shocks and liquidity drying up; that positions which looked diversified carried the same convergence exposures across markets; and that extreme leverage amplified the losses and made a disorderly liquidation a threat to wider markets. That risk-management lesson is well established. A double-loop reading goes further, offered as a hypothesis: that all of this was clever fine-tuning inside a view of risk that may not have been challenged effectively enough.

Investing can intensify these defences, for two reasons that feed each other. The first is success itself. Double-loop learning takes practice at treating mistakes and surprises as information rather than as threats, and a strategy that has worked for years can reduce the pressure to build that practice, while convincing its owners they do not need it. The second is public commitment. A thesis is stated openly, defended to clients, and used to raise the money now behind it, so reopening it means admitting in public that it was wrong. The defence rarely shows up in a single decision. Averaging down is not defensive in itself; it may follow a rule set in advance that responds to the evidence. The defence lies in how the decision is justified: quietly raising the bar for evidence, explaining away the bad news after the event, or treating whoever points it out as disloyal. Diagnose the reasoning, not the trade.

Built, Not Summoned

It follows, first, that exhortation does not work. Telling capable people to be more open-minded is as empty as telling them to be less biased: defensive routines are not a bad attitude that a better one would cure, but a learned, social response to threat, strongest on the questions closest to a person’s own competence and commitment. You cannot will the second loop into running, and you cannot simply install it either. Argyris treated individual and organisational defences as reinforcing each other, so structure that lowers the cost of challenge is necessary but not enough on its own.

What structure can do is lower the personal cost of challenge: separate the team that runs a strategy from the group responsible for checking whether its assumptions still hold, so that asking is someone’s job rather than an act of disloyalty; make disagreement an assigned job, through pre-mortems and red teams; write the governing assumptions down ahead of time, so a review studies a document rather than a colleague. None of this, on its own, is double-loop learning. If the strategy owners never learn to show their own data and reasoning, and to revise their own theory-in-use, the challenge becomes one more ritual: one side prosecutes, the other defends, and the decision turns into a contest over authority rather than a shared test of the reasoning.

This changes the test of whether the second loop is running. It is not whether reviews overturn premises: a real inquiry can examine a governing assumption and decide it still holds, while an organisation can overturn assumptions for show without really inquiring at all. The test is whether the premise was genuinely open to revision: whether the assumptions and failure conditions were recorded before the result, evidence against it sought rather than just received, the reasoning opened to challenge, a dissenter able to change the decision, and the learning built into rules, limits and incentives. The discipline of holding a model loosely, which Why One Model Is Never Enough set as the goal of this series, is not a temperament admirable people can summon. It is a practice, one that has to be built into how decisions are actually made.

The Undiscussable Question

A strategy that has compounded for years can become one whose premise is hardest to examine and most costly to reopen, and the edge that lasts belongs to whoever can still reopen what success has closed.

Investing is only one place this appears. Aviation offers a useful institutional analogue. International standards make prevention the sole objective of accident investigation, require the investigation to remain separate from proceedings concerned with blame or liability, and require an authority independent of bodies that could compromise its objectivity. Its purpose is not to punish, though it is not literally blame-free.

The closing note is the one Argyris insisted on and that practitioners least want to hear. Awareness is not the cure. The professionals he studied could describe defensive routines in detail and went on producing them in the same session, because the routines act on the blind spot, not the conscious mind. Understanding the pattern does not exempt you from it, but it tells you where to build the structure, and the practice, that might on a good day force the question you would otherwise decline to ask.

The discipline is in asking: which loop is this running on, and is the premise allowed to be questioned?

References & Further Reading

1. Theory in Practice: Increasing Professional Effectiveness: Chris Argyris & Donald Schön

2. Double Loop Learning in Organizations: Chris Argyris

3. Organizational Learning: A Theory of Action Perspective: Chris Argyris & Donald Schön

4. Overcoming Organizational Defenses: Facilitating Organizational Learning: Chris Argyris

5. Teaching Smart People How to Learn: Chris Argyris

6. The Reflective Practitioner: How Professionals Think in Action: Donald Schön

7. The Fifth Discipline: The Art and Practice of the Learning Organization: Peter Senge

In 1968, the American Statistical Association and the National Bureau of Economic Research began a quarterly survey of professional forecasters. Alongside the average forecast, the individual responses made disagreement measurable: not the prediction for inflation or growth, but how far apart the predictions were. The forecasters were reading the same data, using broadly the same models. They disagreed anyway, and through the inflationary 1970s they disagreed more. The question of whether high inflation was temporary or permanent had divided the profession into camps that no longer found each other’s reasoning convincing. The price regime did not settle the argument until Paul Volcker forced the matter at the turn of the 1980s. The argument itself, though, had been visible and measurable for years before the prices turned.

This reverses the usual order of attention. We watch prices and treat the beliefs behind them as private, knowable only after the fact. But beliefs leave a measurable trace of their own: disagreement, which can be counted directly.

What kind of problem is this?

A market regime is a kind of agreement. Through the long low-inflation period that followed Volcker, most participants came to share a working model: central banks could and would control inflation, and growth, not inflation, was the main risk to manage. That working model held for the better part of three decades, and through the low-and-stable-inflation years of the 2000s and 2010s it also held that government bonds would rise when equities fell. It was a settled view about how the world was arranged, and it organised everyone’s behaviour around itself.

When what matters is whether a population of interacting believers will hold together or split apart, and when the early evidence of a split may appear in the spreading of views before it is visible in slow price summaries, that is a specific shape of problem. And the discipline that has thought most rigorously about how the opinions of many interacting agents evolve is opinion dynamics, a branch of the statistical physics of social systems.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Opinion dynamics was developed from the 1970s onward by mathematicians, physicists and social scientists, built for abstract agents who each hold a number, exchange it with others, and update by simple rules. Markets do not satisfy its assumptions cleanly: beliefs are not single numbers, the views are entangled with prices and incentives, and the networks of who listens to whom are unobservable and shifting. But the structural logic transfers: whether a population of interacting believers converges on one view or fractures into stable factions depends not on how far apart they are but on whether their disagreement stays bridgeable, and that leaves a measurable trace in the spread of their views.

How A Crowd Reaches Agreement

Start with the simplest model of social learning, set out by Morris DeGroot in 1974. Imagine a group of people, each holding a number, say their estimate of some uncertain quantity. At each step, every person replaces their number with a weighted average of those held by the people they listen to. DeGroot showed that under broad conditions this process converges. The numbers come together, and the group settles on a single shared value. Repeated averaging, across a connected group, manufactures consensus.

This is a slower mechanism than common knowledge, where shared belief turns on what everyone knows that everyone else knows and can shift in an instant. Gradual averaging instead leaves a continuous, measurable trail of disagreement on the way to agreement.

It is a clean account of why agreement is the normal state. Markets are connected groups revising toward those they respect, and such a group tends to converge; a settled regime is the visible form of that convergence. DeGroot describes for the whole population what active inference describes for the single agent, belief drawn toward coherence.

In its canonical connected and ergodic form, the DeGroot model converges to one value. Real populations form lasting factions instead. To explain that, the model needs one more ingredient.

The Limits Of Listening

The missing ingredient was supplied around the year 2000 by Rainer Hegselmann and Ulrich Krause, and independently by Guillaume Deffuant and colleagues. The mechanisms differ in detail: in the Hegselmann-Krause version each agent moves toward the average of everyone within a threshold of its own view; in the Deffuant version, agents meet in random pairs and compromise only when already close. Either way, views beyond the threshold are ignored. The threshold is a confidence bound: the width of disagreement an agent will still take seriously.

In the simplest homogeneous, one-dimensional versions, the confidence bound becomes the decisive parameter. When it is wide, agents listen across the whole range of views, and the population converges to a single consensus, as in DeGroot. When it is narrow, agents listen only to those who already broadly agree with them, and the population fragments into separate clusters. What decides the outcome is not how far apart the views are but whether they stay linked through chains of acceptable disagreement: a wide but continuous spread can still converge, while a narrower distribution can split once a gap opens between camps. These clusters are stable, the competing settled states the system could occupy, defined not by the data but by which conversations remain connected.

The teaching image is a row of people standing on a line, each at the position of their estimate. At each step everyone shuffles toward the average position of the neighbours within earshot. If earshot is wide, the whole row collapses to a single point; if it is narrow, the row breaks into separate huddles that drift apart and then hold.

This is the result worth carrying into markets, because it says something the bare idea of “disagreement” does not. What governs the stability of a consensus is not the level the crowd settles on, nor even how widely its views are spread, but whether that spread is still bridged. The level is what everyone watches; bridgeable disagreement is what lets the agreement survive a shock.

What Comes Apart First

The bond-equity relationship shows this plainly. Through the low-and-stable-inflation years of the 2000s and 2010s it was reliably negative, bonds tending to rise when equities fell, but whether they do is not fixed; it depends on the prevailing belief. When the shared model is “inflation is controlled, the main risk is growth,” bonds hedge equities and the correlation is negative. When inflation or policy credibility becomes the dominant risk instead, equity selloffs are more likely to coincide with rising yields, so bonds can stop hedging equities and the correlation can turn positive. The sign is one expression of the regime the population believes it is in: regime-dependent because macro shocks, policy expectations and beliefs interact.

The opinion-dynamics framework reframes how you watch for that change. Waiting for the realised correlation to flip is a price summary, and a slow one. This framework points instead to the belief structure those prices express, which can shift before the correlation does. Two observables follow from the model.

The first is dispersion. As a consensus loses its hold, the cross-section of forecasts may widen, become more persistent, or turn multimodal before the average or realised correlation clearly moves. Forecast disagreement is measured continuously, and the work of Gregory Mankiw, Ricardo Reis and Justin Wolfers established that it varies over time, tending to be higher when inflation is high, changing rapidly, or relative prices are more variable.

The second observable is not the confidence bound itself, which is latent, but the behavioural proxies for it: the tone and connectedness of the conversation. When the discourse hardens into camps that stop citing each other, when forecasts turn bimodal rather than merely spread, and when previously marginal narratives migrate into the core, those are signs consistent with a narrowing confidence bound. This is the soft signal that Robert Shiller’s work on narrative economics points to: narrative tracking may register changes in the belief environment that the aggregate data have not yet stabilised into.

The transfer here is specific. Opinion dynamics has been applied to markets before, but the use is not to price an asset directly; it is to watch the stability of the belief structure that makes a regime possible, something readable before it resolves into a slow price summary.

The Width Question

The same reading applies wherever a shared view is held by a population that talks to itself. In an organisation, a settled strategy holds while its supporters still engage the colleagues who doubt it, and fractures when the two sides stop finding each other worth answering. In a scientific field, a dominant position can look secure while the dispersion of published views widens and rival camps stop citing one another. In a profession, an agreed standard survives on the bridgeability of disagreement, not merely on the original case for it.

Two cautions keep this honest. The first is that dispersion is only a proxy for fragmentation, not the thing itself, and treating it as an alarm is a problem in signal detection of the kind set out in When Right Still Looks Wrong. Disagreement also rises in ordinary times that resolve without any regime change, so a low threshold for concern produces false alarms, and a high one misses real transitions. Forecaster disagreement has widened many times without a regime turning.

The second caution is sharper, because it turns the framework on its user. Once participants begin watching dispersion as a regime indicator, the indicator becomes part of the dynamics it was meant to observe. A widely shared narrative that the consensus is fragmenting can itself narrow the confidence bound, as participants brace, harden and stop listening across the divide, the very condition that produces the fragmentation the narrative predicted. Understanding how a crowd comes apart does not place you outside the crowd; it can make you a faster contributor to the process you were trying to anticipate.

The discipline is in asking: not where the consensus has settled, but how wide it is, where its gaps are, and how far it still listens across its own disagreement?

References & Further Reading

1. Reaching a Consensus: Morris H. DeGroot

2. Opinion Dynamics and Bounded Confidence: Models, Analysis and Simulation: Rainer Hegselmann & Ulrich Krause

3. Mixing Beliefs Among Interacting Agents: Guillaume Deffuant, David Neau, Frédéric Amblard & Gérard Weisbuch

4. Statistical Physics of Social Dynamics: Claudio Castellano, Santo Fortunato & Vittorio Loreto

5. Disagreement about Inflation Expectations: N. Gregory Mankiw, Ricardo Reis & Justin Wolfers

6. The Great Inflation and Its Aftermath: The Past and Future of American Affluence: Robert J. Samuelson

7. Narrative Economics: How Stories Go Viral and Drive Major Economic Events: Robert J. Shiller

8. Making Sense of Chaos: A Better Economics for a Better World: J. Doyne Farmer

A merchant ship clears Bristol for Jamaica in the autumn of 1738. An underwriter at Lloyd’s coffee house takes a line on her hull and cargo for a premium of a few per cent, and then, like everyone with money at risk, he waits. Weeks pass. The freshest intelligence is six weeks old: a passing ship had spoken her off Madeira, in good order, making reasonable way. By the reckoning of the season she is now overdue. The question that matters is the only one he cannot answer: is she late, or is she lost?

These are not two shadings of one answer. Late means she will make port, the premium was fair, and the voyage closes a modest profit. Lost means the hull is on the seabed off Hispaniola, the cargo, worth several thousand pounds, is gone, and the underwriter’s contingent liability becomes real the moment the loss is confirmed. The underwriter’s information does not separate these two worlds. His last sample of reality said afloat and well; everything since is inference laid across a gap. And the reassuring reading, late not lost, is not a blurrier version of the truth. It can be a sharp and confident picture of a ship that has been at the bottom of the sea for a month.

What kind of problem is this?

When the thing you are watching can change faster than you can look at it; when the intervals between observations are long enough for the state to turn over inside them; when no amount of care applied to the glimpses you do have will reconstruct what happened between them — that is a specific shape of problem. And the discipline that has thought most rigorously about how often you must look at something to know what it is doing is the theory of sampling.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Sampling theory was worked out by communications engineers. In 1928, at Bell Laboratories, Harry Nyquist established how fast a telegraph channel of a given bandwidth could carry distinct signals; two decades later Claude Shannon gave the result its modern form. It was built for engineered signals like telegraph pulses, a voice on a line, and an image broadcast through the air, all quantities with a well-defined highest frequency. Markets are messier: a price or a valuation has no fixed bandwidth, the process does not sit still, participants adapt, and observing can change what is observed. But the structural logic — that to reconstruct a band-limited signal, you must sample it at more than twice its highest frequency component, and that observing too slowly does not blur the truth but replaces it — transfers.

The Wheel That Turns Backwards

The statement is this. If a signal contains no frequency higher than B cycles per second, samples taken faster than 2B times a second determine it completely, and under ideal reconstruction the original can be recovered exactly, with nothing left out. The number 2B, twice the highest frequency present, is the signal’s Nyquist rate; half of whatever rate you actually sample at is that sampler’s Nyquist frequency.

The quantity B is the bandwidth: the highest frequency the signal contains, how fast it can change. A slow signal can be captured with leisurely samples; a fast one demands fast ones. The floor on how often you must observe is set by the signal, not by how much you care or how carefully you measure each sample.

The instructive part is what happens when you sample too slowly, below the Nyquist rate. The reconstruction does not merely degrade; it actively misleads.

Anyone who has watched old film of a stagecoach has seen this. The wheel turns forward, but on screen the spokes seem to crawl backwards, or turn at some lazy speed the wheel never had. The camera samples the scene a few dozen times a second, the spokes pass faster than that, and the eye, handed those sparse frames, builds the smoothest motion consistent with them: a slow backward turn that was never happening. Engineers call it aliasing. A frequency above half the sampling rate does not vanish; it folds down and reappears as, or contaminates, a lower-frequency component the original need not have contained.

This is the property that makes undersampling worse than honest ignorance. An undersampled signal is not blurry. It is crisp, plausible, and wrong. It offers a slow, smooth trend assembled out of fast events you never caught, with no internal sign that anything is amiss.

There is no cure after the fact. Once a signal has been sampled too slowly, the high-frequency information is not merely lost but disguised as something else, and from those samples alone no cleverness will bring it back. The remedy has to come before the sampling, never after it.

The Overdue Ship, Again

Now the overdue ship resolves into the practical analogue of an undersampling problem. Strictly, a voyage state that can jump from afloat to lost is not a band-limited signal, and has no finite Nyquist rate. But the failure mode is the one sampling theory teaches us to distrust: when the process can change materially inside the observation interval, the observer replaces an unseen transition with a smooth, plausible interpolation. Shipping intelligence arrived once every several weeks, so the underwriter was not holding a slightly stale but faithful picture, but a benign reconstruction laid across an interval in which the true state may already have changed: she is making slow way, the trades were light, she will be along. Late, not lost, was the spokes appearing to turn backwards.

The structure has not aged. It is the daily condition of any institution holding positions whose true value moves faster than it is observed. A private asset marked once a quarter has economics that do not pause politely between marks. A counterparty reviewed once a year can fail over a single weekend. In each case the interval between observations is not a gap in resolution that leaves the broad shape intact; it is room for an alias to form.

The clearest instance is the smoothness of illiquid returns. Getmansky, Lo and Makarov showed that the reported returns of assets which trade rarely or are valued by appraisal display artificially low volatility and high serial correlation, because each figure leans partly on the last and the true price is observed too seldom. David Geltner had documented the same smoothing in appraisal-based property indices. The mechanism here is not literal spectral aliasing but its close relatives, stale pricing and the moving-average smoothing of marks; the family resemblance is what matters. The low measured volatility of such a series says little about the asset and a great deal about the sampling: it is the calm the sampling invents.

An earlier post, When Right Still Looks Wrong, examined how one sets a threshold on a noisy but genuine signal, trading false positives against false negatives. Aliasing sits upstream of that decision: it corrupts the input before any threshold sees it. You can hold a perfectly calibrated judgement and still be confidently wrong, because what you are judging is not a faithful sample of reality but a smooth fiction the sampling produced.

Before The Gap Opens

If the diagnosis is a sampling problem, the remedies are the engineer’s, and there are only two.

The first is to raise the sampling rate. Observe more often: mark to market where a market exists, monitor counterparties continuously rather than at annual review, seek intelligence within the period, not only at its boundaries. This is worth doing, and it meets a hard floor. Some exposures are illiquid precisely because they cannot be observed or traded often; the observation cadence a quarterly-valued asset forces on you cannot be wished away, and pretending otherwise reinstates the original error in a more confident form.

The second remedy is the one institutions reach for too rarely. In engineering it is to band-limit the input before it reaches the sampler. The translation is awkward, because an institution usually cannot make an illiquid asset observable, and it cannot filter away the asset’s real economic jumps. What it can do is design the exposure before the gap opens: size it, collateralise it, margin it, hold liquidity in reserve against it, attach covenants, or refuse it outright when the unseen interval can contain a loss it could not survive. None of this recovers the path between two valuations, which is gone; it makes the unseen move survivable rather than fatal. The decision belongs at the front, in exposure design, not at the back, in reconstruction from a handful of stale marks after the loss has surfaced.

A corollary cuts against how risk is usually read. The smooth series always looks more trustworthy than the jagged one: lower volatility, steadier returns, fewer alarms read as quality and control. When the smoothness is genuine, that reading is correct. When it is a smoothing artefact of a process observed too slowly, the reading is exactly backwards, and the calmest line in the book conceals the most.

The Sampling Question

Knowing the theorem raises no one’s sampling rate by itself. It gives the underwriter no faster ship and does not make an illiquid asset liquid; understanding the dynamic does not exempt anyone from living inside it. What the theorem changes is the kind of attention one pays. It replaces a vague trust in sparse data with a specific suspicion: that the reassuring reading may be a confident reconstruction rather than a faithful one, and that the smoothness of a number can say as much about how often it was sampled as about the thing it measures.

The problem extends beyond investing. In medicine, a chronic condition monitored by a test every few months can flare and subside between visits, and the calm chart records a steadiness the patient never had. In opinion polling, sentiment sampled monthly can swing and return between surveys, so the smooth trend line is partly an artefact of when the questions were asked. In industrial monitoring, a sensor read too slowly reports a steady machine while a fast vibration builds unseen. In each case the cadence of observation, not the thing observed, sets what can be known.

The discipline is in asking: not whether the latest reading is accurate, but whether we are observing often enough for the smoothness we see to belong to the thing itself, rather than to the gaps between our glimpses of it?

References & Further Reading

1. Certain Topics in Telegraph Transmission Theory: Harry Nyquist

2. Communication in the Presence of Noise: Claude Shannon

3. The Origins of the Sampling Theorem: Hans Dieter Lüke

4. An Econometric Model of Serial Correlation and Illiquidity in Hedge Fund Returns: Mila Getmansky, Andrew W. Lo & Igor Makarov

5. Smoothing in Appraisal-Based Returns: David Geltner

6. A History of Lloyd’s: Charles Wright & C. Ernest Fayle

Here is a market story that sounds simple: equities fall when rates rise.

It has the appeal of compression. A large part of the market reduces to one variable. Discount rates rise, present values fall, long-duration assets suffer, growth stocks derate.

Then reality interferes. Rates rise and equities rally.

So the story acquires a patch. Perhaps earnings expectations improved, or the rise reflected stronger nominal growth, or positioning was too bearish, or the real yield mattered, not the nominal one. Any of these may be true, but notice what has happened. The headline rule was five words; the rule that survives contact with the data now runs to a paragraph of conditions, each added after the previous version failed. The story did not explain the market; it hid the market in the exceptions.

The problem is not that these variables are irrelevant; several belong in any decent model. What makes them special pleading is timing: they were not part of the model when it made its call, but were added afterwards, one for each failure, to explain it away.

Every investor is in the compression business. The world is too big to hold in mind, to trade whole, or to explain to a client or a committee, so you reduce it to a theme, a factor, a regime, a multiple, a line on a pitch deck. The danger is not that you compress, but that you forget you have done it, and mistake a short slogan for a short explanation.

What kind of problem is this?

When an explanation looks short but runs long once you count what it leaves unexplained, when a model survives not by predicting new data but by absorbing each contradiction as a special case, the question is not “is this simple?” but “how much special pleading does it need?” A corner of information theory and statistics has formalised exactly that trade-off, in a principle called Minimum Description Length.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Minimum Description Length, introduced by the statistician Jorma Rissanen in 1978, rests on a plain intuition. The best explanation is not the one with the fewest words, nor the one with the most moving parts. It is the one that gives the shortest total description of the observed data: the cost of specifying the model, plus the cost of encoding the data given that model.

The Price of Brevity

A model is a compression scheme. To describe a set of observations you need two things: the rule, and the data it does not already capture. A short rule that leaves a large pile of exceptions yields no short description at all; a longer rule that shrinks the pile enough is worth paying for. Minimum Description Length counts both. Formally, it picks the model that minimises L(M) + L(D|M): the bits needed to describe the model, plus the bits needed to describe the data once you have it. The question is not how elegant the explanation is, but how much special pleading it takes to keep standing.

So three things that look alike should be kept apart. Simplicity is a short explanation. Compression is a short explanation that preserves the structure that will later matter. Deletion is a short explanation that throws away the part of the world that would have made the decision hard. Markets reward compression and punish deletion, and from a distance all three look identical.

A valuation multiple is compression. A price-to-earnings ratio takes an entire business, its margins, capital structure, and competitive position, and reduces it to a number that gives comparison, discipline, speed. But it also deletes. A low multiple may mean undervaluation; it may equally mean terminal decline, peak earnings, hidden leverage, or profits that will never become cash. The multiple is wrong not when it compresses, but when the variable it deletes is the one that decides the outcome.

Factors carry the same hazard. “Value works.” “Momentum persists.” Each compresses a historical regularity and carries an implicit claim about which details are safe to ignore, and the trouble begins when the label replaces the mechanism. Value is a family of mechanisms, from mean reversion to distress risk to institutional constraint, each implying different conditions under which the strategy should work, stop working, or reverse. The label compresses the mechanism into a single word. Use the word without the mechanism, and you cannot say which conditions the strategy depends on, or when it will stop working.

Backtests offer the appearance of compression in a subtler form. A strategy with enough rules can describe the past with great precision: buy this signal, exclude that sector, winsorise this variable, drop the crisis months, neutralise beta, reintroduce it when the regime changes. The result looks like a model, but it is closer to a lookup table for the past: it has not compressed the data, it has memorised it. A useful model says that, given this structure, these observations are no longer surprising; an overfit one says that, given these observations, it can build a structure that would have predicted them. The first compresses; the second only rearranges the archive.

The Narrative That Forbids Nothing

A market narrative compresses attention: it tells investors which facts matter, and a thousand details collapse into a few decisive questions. But a narrative can explain everything by forbidding nothing. If stocks rise, it explains why; if they fall, it explains why. If margins expand the thesis is confirmed; if they compress it is a better entry point. Nothing is ever disconfirming; every observation is reclassified as support.

This feels like explanatory power and is usually the opposite. A good model has exclusions: it tells you what would count against it. A narrative that absorbs all evidence is not compressing the data; it is refusing to count its errors as a cost. The thesis may still be right, but it is no longer simple, and its description length is growing. A model that explains everything predicts nothing. The same instinct hides exceptions off balance sheet: a private asset marked at par, stable until the discount rate moves or the exit market closes or no sale forces the question. The real test is not whether the story can be repeated but whether it keeps reducing surprise as new data arrive.

A model that compresses well makes new data cheaper to interpret; you know where to put the next observation. A bad one does the reverse. Each new fact makes it more expensive, until it exists mainly to protect the original conclusion: still named and defended, simple from a distance, but no longer compressing anything.

For over a thousand years, astronomers did this. When observations failed to match the geocentric model, they added circles upon circles to preserve it; it could always be made to fit, but at a description that would not stop growing. Investors do the same. A company misses guidance, but the miss is temporary; then margins fall, but that is strategic investment; then management changes the metric, then the thesis, though the position does not. At each step the explanation is plausible, but plausibility is not the test; the test is whether the total explanation is getting shorter or longer. Sometimes the right move is not to sell but to admit the compression has failed: the position may still be attractive, but it now needs a different model.

The Opposite Mistake

There is a failure in the other direction: worshipping simplicity. Some investors treat complexity itself as error and demand one variable, one chart, one clean causal line. “Money printing causes inflation.” “Deficits push bond yields up.” “Cheap stocks beat expensive ones.” Each compresses something real, and none is sufficient, because the omitted variables do the deciding: velocity, credibility, global savings, refinancing conditions, the policy reaction function. Minimum Description Length does not say the simplest model wins; it says the shortest adequate description wins, and adequacy is the operative word. A map with one road is useless if the terrain has many. The over-padded model fails for the mirror reason: a thesis promising revenue growth, margin expansion, re-rating, benign competition, and skilful capital allocation may hold five assumptions, or one assumption stated five times, that everything goes right. Pay for complexity only when it earns its cost.

Consider two explanations for a bank that fails. The first: depositors panicked, which captures the final motion but not the stored instability. The second: it funded long-duration assets with flight-prone deposits, hedged little, and met rising rates carrying unrealised losses that turned fatal once confidence broke. The second is longer, yet it compresses more, because it separates the trigger from the vulnerability and tells you where to look for the same risk elsewhere. The better compression is not the shorter sentence but the one that discards the most irrelevant detail without discarding the cause. That is the standard: not elegance, not cleverness, not how easily a thesis can be sold, but the shortest explanation that survives.

The Compression Question

The principle does not promise the model is true. It tells you what to watch. How many assumptions must hold, and are they independent or the same optimism in different cells? How many observations are filed as “temporary”, and how long has that file been growing? When a contradiction arrives, does the explanation get shorter, because the structure was revised, or longer, because a clause was added to protect it?

The test applies wherever an explanation has to survive new facts. In medicine, a diagnosis that files each contradicting result as an atypical presentation is lengthening, not compressing, while the patient stays unexplained. In engineering, a root-cause account that adds a special condition for every failure it did not predict has begun cataloguing the system rather than explaining it. In law, a doctrine that needs a fresh distinction for each adverse precedent grows the same way. The move is always the same: protect the conclusion by adding clauses, and call the result understanding.

Knowing this does not place you outside the problem. You compress too, and the urge to rescue a failing thesis with one more clause is not a temptation only other people face. Understanding why descriptions lengthen does not stop your own from lengthening; it only tells you where to look first.

The discipline is in asking: not whether the explanation is short, nor whether it is persuasive, but whether it is compressing reality, or merely hiding the part it cannot explain?

References & Further Reading

1. The Minimum Description Length Principle: Peter Grünwald

2. Modeling by Shortest Data Description: Jorma Rissanen

3. Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance: David Bailey, Jonathan Borwein, Marcos López de Prado & Qiji Zhu

In 1952, Harry Markowitz published “Portfolio Selection” in the Journal of Finance. It made a simple but important claim: that the proper unit of analysis for an investor was the portfolio rather than the individual security, and that the relevant quantities were expected return, variance, and the covariance between holdings. The mathematical framework required to define an efficient portfolio along these lines was set out in the article. The framework that would later be called modern portfolio theory was, in its essentials, complete.

The institutional investment community didn’t, broadly, adopt it. Through the 1950s and 1960s, many large institutions continued to allocate capital using methods that had little to do with mean-variance optimisation. Asset allocation was governed by tradition, by category-by-category prudence, by intuitions about what one ought to own. The portfolio-level thinking that Markowitz had formalised existed in academic finance but not in trustee meetings.

Sharpe built on the framework in 1964 in developing the capital asset pricing model. Markowitz and Sharpe were among the recipients of the 1990 Nobel Prize in Economic Sciences for foundational work in financial economics — thirty-eight years after the original paper. By the early 2000s, the language of mean-variance analysis, efficient portfolios, and optimisation had become embedded in institutional asset allocation, even where committees used constrained, liability-aware, or heuristic versions rather than textbook optimisation. The framework had finally crossed.

Half a century is a long time for an idea to traverse the gap between being formulated and being adopted in practice. The puzzle isn’t that the idea was wrong. The puzzle is what determined when, having been articulated, it finally became behaviour and practice.

What kind of problem is this?

The Markowitz story was not, fundamentally, about an idea waiting to be understood. It was about an understood idea waiting for the social conditions of its adoption. The information was available throughout. The logic was clear enough to reshape academic finance. What had not yet formed was the social structure through which the behaviour of using it could spread.

When an idea is widely known but not adopted, when the gap between awareness and behaviour cannot be closed by stating the case more clearly, when the diffusion of a practice depends not only on its merits, but also on the social structure around it — that is a specific shape of problem. And the discipline that has thought most rigorously about how behaviour spreads through networks of people is the sociology of social contagion, particularly the work of Damon Centola at the University of Pennsylvania.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Where the spread of a behaviour requires more than a single exposure, the structure of the network — not merely its connections, but the way they overlap — governs whether the behaviour can cross.

Two Kinds Of Contagion

The starting point is a distinction. Some things spread once a single exposure occurs. Others require reinforcement from multiple independent sources before adoption follows. Sociologists call the first a simple contagion and the second a complex contagion. The defining feature is structural — how many independent sources of activation must converge on a person before they adopt — rather than a property of the thing being transmitted.

A virus is the canonical simple contagion. One infected source can be sufficient to transmit it; confirmation from multiple sources isn’t structurally required. Much information behaves the same way — a single trustworthy mention is generally enough to know a fact — though gossip and credibility-sensitive claims can themselves require reinforcement from several sources. Models that describe simple contagion treat the network as a set of pipes. What matters is connectivity. Shorter average path lengths and broader reach speed transmission.

Behaviours are typically complex contagions, with adoption thresholds that exceed a single exposure. Adopting a new practice, joining a movement, changing a professional routine — these usually require seeing the behaviour from multiple sources before adoption becomes plausible. The reasons combine. Social proof: I’m more confident the behaviour is reasonable when I see several people doing it. Risk reduction: the cost of being wrong is lower if others share the choice. Legitimacy: the behaviour feels institutionally sanctioned only when it appears across the social field. Learning: the practical know-how is typically distributed and requires more than one teacher.

Why Bridges Must Be Wide

This distinction has a network consequence that’s easy to miss. Mark Granovetter’s celebrated 1973 paper, “The Strength of Weak Ties,” showed that distant, weakly connected contacts were particularly valuable for information transmission. They connect otherwise separate clusters and so accelerate the diffusion of facts and opportunities. The result is famous, and for simple contagions it holds.

For complex contagions, the weak-tie result often reverses. A single weak tie can’t, by itself, transmit a high-threshold behaviour, because adoption requires reinforcement from multiple independent sources and a single tie supplies only one. What’s needed instead is what Centola and Michael Macy, in a 2007 paper, called a wide bridge: a connection between two communities that consists of several overlapping ties, rather than one. A wide bridge can carry behaviour across because it delivers the reinforcement that adoption requires. A bridge too narrow to deliver the required reinforcement can’t carry the behaviour across, however short the path it creates.

Centola’s empirical work made this concrete. In a 2010 study published in Science, he constructed online networks with identical numbers of nodes and ties but different topologies — some random, some clustered — and observed how a particular behaviour, registration with an online health forum, spread through each. The clustered networks, with their wide bridges and redundant exposures, propagated the behaviour faster than the random networks, even though the random networks had shorter average paths. For a simple information contagion, that result would have been backwards. For behaviour, it was exactly what the theory predicted.

What Markets Reveal

Markowitz’s framework was one instance of a much broader pattern. The same distinction surfaces across markets day to day, but the unit of analysis matters. Headline-level repricing can occur quickly because it requires only marginal orders from actors already authorised to place them — a trader changing a quote, an algorithm rebalancing a book. Durable exploitation of a thesis is a different kind of behaviour. It requires capital commitments, mandate changes, the willingness to risk credibility on an unpopular position. That kind of behaviour propagates at the speed of reinforcement: often slowly until wide bridges form, and sometimes rapidly once they do.

Several familiar phenomena fall out of this framing.

An apparent inefficiency can sit visible for years if betting against it requires behaviour that the relevant network can’t easily transmit. The information is there. The behaviour doesn’t cross. This doesn’t replace the standard limits-to-arbitrage explanations — financing constraints, benchmark risk, career risk, short-sale frictions. It helps explain why the social and institutional willingness to bear those limits may itself diffuse slowly.

When the wide bridges finally form — through new intermediaries, new platforms, new regulatory frameworks, new generations of decision-makers, new vocabulary that becomes shorthand inside a profession — adoption can cascade rapidly. The window between “still lonely” and “fully crowded” can be much shorter than the window of persistent visibility before it.

Some practices become institutional defaults not because the underlying case is overwhelming but because the network achieved wide-bridge density inside the right professional clusters early. Other equally defensible practices remained marginal because they didn’t. The set of universal practices is shaped not only by analytical merit, but also by social topology.

Strategies based on widely discussed but not widely adopted ideas are the strategies whose practitioners spend long stretches looking wrong. The frustration is structural. Holding the position requires sitting on a narrow bridge while waiting for it to widen.

The Bridge Question

Centola’s framework does not give you a clock. It cannot, by itself, tell you when the bridges around an idea will widen, or whether the network involved is approaching the threshold where adoption becomes possible. What it gives you is a structural test, and the test is specific.

Picture what adoption would look like at scale — not one institution adopting a new approach, but several converging on a pattern: a portfolio framework like the Total Portfolio Approach taking hold across allocators, a new research practice for valuing private assets becoming standard, a shared methodology for assessing long-horizon risk, common analytics for factor exposure. Then trace the network through which that pattern would have to spread — who is connected to whom, which clusters reach which adopters. Then ask: how many independent sources are already moving, and how well do they reach the practitioners who would have to follow?

The same questions apply far beyond markets. In organisations, a practice held up by a single senior champion is on a narrow bridge — and it disappears with the champion. In scientific communities, an idea cited by one camp may be widely heard but nowhere adopted; an idea taken up across previously distinct sub-disciplines, each translating it into its own language, has already crossed. In professions, the moment a new approach appears at once in textbooks, conferences, regulators’ guidance, and peer practice, crossing to it no longer feels like a choice.

Recognising the structure does not give you a forecast. Bridge formation can take much longer than expected and then happen suddenly. And recognising the structure does not place you outside it: your own decisions are reinforcing some bridges and leaving others narrow. Knowing why your position is lonely does not make it less lonely; it only tells you not to mistake the loneliness for being wrong.

The discipline is in asking: are the bridges widening around this idea, or is it still resting on a single source?

References & Further Reading

1. How Behavior Spreads: The Science of Complex Contagions: Damon Centola

2. Complex Contagions and the Weakness of Long Ties: Damon Centola & Michael Macy

3. The Spread of Behavior in an Online Social Network Experiment: Damon Centola

4. The Strength of Weak Ties: Mark Granovetter

5. Portfolio Selection: Harry Markowitz

6. Capital Asset Prices: A Theory of Market Equilibrium Under Conditions of Risk: William Sharpe

7. Capital Ideas: The Improbable Origins of Modern Wall Street: Peter Bernstein

This week, OpenAI announced that an internal reasoning model had disproved Erdős’s long-standing unit-distance conjecture — a central part of the planar unit-distance problem, first posed in 1946.

The mathematicians who checked the work were direct. Noga Alon, Professor of Mathematics at Princeton, called it “an outstanding achievement.” Fields Medallist Tim Gowers, Combinatorics chair at the Collège de France, wrote that if a human had written the paper and submitted it to the Annals of Mathematics, he would have recommended acceptance without hesitation. Jacob Tsimerman of the University of Toronto put the model’s edge plainly: it can play for longer and in more treacherous waters than mathematicians without getting overwhelmed.

The model didn’t summon a new kind of cognition. It applied known tools from algebraic number theory — Golod-Shafarevich theory, infinite class field towers — with more patience than a human could sustain. Sophisticated, elegant, recognisably mathematical. And once explained, comprehensible.

Which puts the result inside an older question.

In his 1985 autobiography Enigmas of Chance, the mathematician Mark Kac reached for a distinction that has stuck. He was writing about Richard Feynman, whom he had known at Cornell.

There are two kinds of geniuses: the "ordinary" and the "magicians." An ordinary genius is a fellow that you and I would be just as good as, if we were only many times better. There is no mystery as to how his mind works. Once we understand what he has done, we feel certain that we, too, could have done it. It is different with the magicians. They are, to use mathematical jargon, in the orthogonal complement of where we are and the working of their minds is for all intents and purposes incomprehensible. Even after we understand what they have done, the process by which they have done it is completely dark. … Richard Feynman is a magician of the highest caliber.

Some are commensurable with us. Others remain unaccountable, even with the trick laid bare.

Which kind of genius is AI?

In Machines of Loving Grace, Dario Amodei sketches a scenario: a country of geniuses in a datacenter — millions of model instances smarter than a Nobel laureate, running at ten to a hundred times human speed. From it he argues that fifty to a hundred years of biomedical progress could be compressed into five to ten.

AI is an ordinary genius. Not a magician.

The scaling laws are not dark. They are the most ordinary thing in the world — more data, more compute, more capability, on predictable curves. The argument those curves invite is that no rung of cognition the model occupies is qualitatively beyond us. The gap is quantitative — speed, memory, parallelism.

Every industry was built around a shortage

Economic activity is the crystallisation of knowledge into matter and decision. A pencil embodies forestry, mining, chemistry, manufacturing. A drug embodies biology, chemistry, clinical practice, regulation.

The computation required to produce anything complex exceeds what any one mind can hold. To work around that, we built distributed systems—specialists, teams, committees, firms.

Look at any knowledge-intensive industry and you see the same architecture. Investment management has analysts, credit desks, regional teams, investment committees. Law firms have practice areas and partners. R&D organisations have departments and review panels.

These structures are workarounds for a constraint—the visible footprint of a shortage. There has never been enough ordinary genius to go around.

Ordinary genius at scale dissolves those workarounds. Three questions follow.

Where has the limit on what one mind can hold forced work into serial handoffs? AI now brings analytical capability directly to the object of work—the patient record, the legal brief, the engineering drawing—rather than routing it through serial expert review. The bottleneck dissolves.

Where has the cost of integrating across domains kept synergisation rare and meeting-bound? Where five or six domains were combined laboriously through meetings, AI can hold dozens at once. Synergy becomes the default, not the special case.

Where has the cost of analysis made coverage a triage decision rather than a quality decision? Oncology clusters around common tumour types. Investment coverage clusters around the largest few thousand securities. Legal precedent gets sampled rather than read exhaustively. When the cost of competent analysis collapses, the question shifts from what can we afford to cover? to what should we cover?

The bottleneck is magic

The reason Amodei thinks the twenty-first century can be compressed by ten times is the same reason every knowledge industry’s specialists, silos, and committees can be redesigned. What looked like a bottleneck — the patience to do the analysis, the connection-making across domains, the sheer attention required to read everything that mattered — was a shortage of ordinary genius. AI removes it.

But Kac’s distinction cuts both ways. If AI is the ordinary genius, the magician’s role does not vanish. It is left for us.

The magic is judgement, determination, empathy — being human. It is the reading of context the data cannot capture. The conviction that holds through a decade of contrary evidence because the causal model is sound. The empathy that moves people when argument alone cannot. The leap that reorganises a field from outside, made by someone who looked at the same numbers as everyone else and saw something no one had seen.

Lord Kelvin had declared heavier-than-air flight impossible. The Wright brothers, finding that accepted calculations did not match their glider tests, built their own wind tunnel and generated the aerodynamic data that made the 1903 Flyer possible. The magic was not ignoring evidence — it was refusing to treat inherited data as final. No extrapolation from what was known in 1902 would have predicted success. That took a leap.

As ordinary genius becomes abundant, the human part does not become obsolete. It becomes more visible. More valuable. More clearly the thing only people supply.

The architecture changes around it. The magic does not.

What we are actually building

The future Amodei describes does not require us to summon a god. It requires us to manufacture, at scale, the kind of mind we already understand—and to redesign around the dissolving constraint.

That is an ordinary project, in Kac’s sense.

It is also the project that frees us for the work only people can do.

The ordinary genius is what we are building. The magic is what it is for.

In the eleventh century, one strand of long-distance commerce across the medieval Mediterranean was sustained by a network of Jewish merchants known as the Maghribi traders, operating between Cairo, Sicily, Tunisia, and southern Spain. They sent goods through agents in distant ports. A merchant in Cairo who consigned cotton to an agent in Tunis could not verify the price obtained, the costs deducted, or the timing of the remittance. Formal legal enforcement was limited and costly. In Avner Greif’s account, influential but contested, it could not by itself solve the agency problem. And yet the trade sustained itself from the tenth to the twelfth century.

The same pattern recurs across markets and periods. Wholesale foreign exchange, dealer-mediated bond markets, syndicated lending, and most negotiated trade in instruments without central clearing all share this feature: cooperation that holds together more through the expectation of future interaction than through the prospect of formal enforcement.

The economics points the other way. In any single exchange, the agent who has sold the cotton has every reason to underreport the proceeds; the dealer who has agreed a price has every reason to walk away once the market moves. By one-shot logic, none of these markets should hold together.

But they do. One-shot logic treats each exchange as standalone. The exchanges are not standalone, and what connects them is what does the work. The structural question is what that connection consists of, what holds it, and what breaks it.

What kind of problem is this?

When agents interact repeatedly with identifiable counterparties who remember past behaviour, when cooperation is sustained by the prospect of future encounters rather than external enforcement, that is a specific shape of problem. And the discipline that has formalised it, beginning in the late 1970s, is repeated game theory.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Robert Axelrod’s framework, developed across two computer tournaments in 1979 and 1980 and synthesised in The Evolution of Cooperation (1984), was built for an abstract setting: two-player games with well-defined payoffs, symmetric players, and explicit move sequences. Markets and historical trading networks do not satisfy these assumptions cleanly: participants do not always face the same counterparties, payoffs are rarely symmetric, the games are nested inside larger institutional and legal structures, and players can sometimes choose not to play. But the structural logic — that conditional cooperation can emerge among self-interested actors when a small set of conditions holds, and that those conditions are empirically identifiable — transfers.

The Strategy That Won Twice

Axelrod’s tournaments asked a simple question. In a repeated Prisoner’s Dilemma — the standard two-player game in which each side chooses cooperation or defection, where defection pays better in any single round but mutual cooperation pays better across many — what strategy wins? He invited submissions from game theorists, economists, psychologists, mathematicians, sociologists, and political scientists. Each submission was a programme that received the history of past interactions with its opponent and chose its next action. The submissions played each other in a round-robin tournament. The strategy with the highest total score across all matches won. Tit-for-Tat won Axelrod’s particular tournaments; the lesson is not that it is universally optimal, but that simple, legible reciprocity performed robustly in that environment.

The first tournament received fourteen submissions, plus a Random strategy as control. The winner was the shortest programme submitted, four lines of code: Tit-for-Tat, written by the mathematical psychologist and game theorist Anatol Rapoport. The strategy was simply to cooperate on the first move, and then on every subsequent move to do whatever the opponent had just done. Cooperate if cooperated with. Defect if defected against. Forgive immediately. Never escalate.

Axelrod ran the tournament a second time the following year, with sixty-two entries — the first tournament’s results now in the public record. Many of the new submissions were sophisticated attempts to exploit Tit-for-Tat or to beat it through cleverer conditional logic. Rapoport submitted Tit-for-Tat again. It won again.

What the tournaments revealed was that four properties characterised the strategies that did well across the field of opponents. Niceness: never defect first. Provocability: retaliate against defection. Forgiveness: do not punish forever; recover cooperation when the opponent does. And clarity: be simple and consistent enough that others can recognise the pattern and adapt to it. The successful strategies signalled their conditionality clearly rather than trying to exploit it.

What The Future Polices

The analytical core of the framework is a threshold condition. For Tit-for-Tat to be stable against defection, the probability of future interaction (the shadow of the future, in Axelrod’s phrase) must be high enough that the long-run payoff to cooperation exceeds the one-shot payoff to defection. The condition has a clean form: if w is the discount factor combining the probability of future interaction with the rate at which future payoffs are valued, conditional cooperation can resist invasion by defection when w satisfies the relevant payoff thresholds. A higher w — a longer shadow of the future — means today’s choices weigh future consequences more heavily.

The condition has several non-obvious consequences. First, no prior trust is needed for cooperation to begin, provided the agents can identify one another, expect continued interaction, and use strategies capable of reciprocation. Small clusters of reciprocating strategies can establish cooperation even in a population otherwise inclined to defect. Second, cooperation is fragile to changes in the discount factor — anything that shortens the shadow of the future, even temporarily, can collapse a stable cooperative equilibrium. Third, the structure of the underlying interaction matters as much as the disposition of the participants. The same individuals can cooperate in one structural setting and defect in another. The disposition is selected for by the structure, not imposed on it.

Axelrod and the biologist W. D. Hamilton extended the framework in a 1981 Science paper to evolutionary settings, showing that conditional cooperation can spread through populations of agents whose dispositions are inherited rather than chosen. Reciprocal behaviour in animals lacking the capacity to deliberate falls under the same conditions. The framework is not specific to deliberate decision-making: it applies wherever agents interact repeatedly, identify each other, remember, and adjust.

Where Reputation Is Collateral

The financial markets in which cooperation visibly holds are precisely those that meet Axelrod’s structural conditions. Members are identifiable. Interactions are frequent. Behaviour is remembered. A bond dealer who reneges on a firm quote or agreed trade pays an immediate cost in the willingness of other dealers to quote in return; the cost is incurred not through any formal mechanism but through the accumulated adjustments of counterparties who track. The same structure underlies wholesale foreign exchange, dealer-mediated credit, and most negotiated trade outside anonymous central limit order books.

The frameworks for understanding these markets have tended to use the vocabulary of trust, relationship, and reputation. Axelrod’s contribution is to give those words a structural underpinning. In settings of repeated interaction with identifiable counterparties, much of what participants call trust functions as the equilibrium behaviour of agents who expect to interact again, can identify each other, and can adjust. Reputation is not a separate quantity that decorates the trade. It is the running tally each counterparty maintains of the others’ conditional behaviour, and the adjustments that flow from it.

The framework also explains why some markets cooperate and others do not. Anonymous order books cannot support counterparty-specific bilateral reputation through the trade itself. The interaction is structurally one-shot, since no individual interaction history is available. Central limit order books are designed precisely to suppress the counterparty-specific cooperation game, since reliance on reputation in large anonymous markets would advantage incumbents and disadvantage new entrants. The cost is that discipline shifts toward rules, central clearing, margining, surveillance, and platform governance — formal mechanisms that police what bilateral reputation would otherwise have policed.

The Network You Cannot See showed that systemic risk lives in the edges between participants. The Axelrod framework specifies one of the things those edges carry: a record of conditional behaviour that enables cooperation without central enforcement.

When The Shadow Shortens

The framework’s most useful diagnostic application is identifying when the conditions for cooperation erode. They can erode slowly, through structural change to the market, or quickly, through shock to expectations.

The slow erosions are the most diagnostic. When long-tenured dealers retire and are replaced by traders with shorter expected tenures, the shadow of the future shortens for the new participants, regardless of the market’s overall structure. When interdealer broking becomes anonymous, or name-give-up rules are removed, the recognition requirement is no longer met. When reputational systems become institution-level rather than person-level, with personnel turnover decoupling individual reputations from their employers, the policing weakens. None of these changes is necessarily bad. Each was introduced for a reason. The framework’s contribution is to make explicit what is being given up in exchange.

The fast erosions are most visible in stress. The expectation of future interaction can collapse abruptly when a counterparty’s solvency comes into question, or when participants believe the market itself may not survive the day. The cooperation game flips. The participants who were cooperative last week defect today not because their dispositions have changed but because the supporting conditions no longer hold. This is why a moral reading of cooperative breakdown often misfires. What looks like a failure of character is usually a failure of structure: the same individuals, in the same market, under different conditions, behave differently.

What the framework does not give is a prediction of when the conditions will erode for any particular market. The discount factor w is not directly observable, and the interactions between structure, personnel, and expectation are too rich to model in closed form. The framework’s contribution is to identify what to look at: not whether the cooperation has held in the past, but whether the conditions that supported it remain.

The logic extends beyond markets. In diplomatic relationships, the same conditions — identification, expected continued interaction, capacity for proportional response — determine whether sustained cooperation is feasible without external enforcement. In professional communities and academic disciplines, reputation operates as a running record of conditional behaviour that polices conduct prior to any formal sanction. In long-term partnerships of any kind, the shadow of the future shapes what each party can credibly commit to without contract. In each case, the cooperation visible at any moment is the equilibrium under the current conditions; and the conditions are themselves a variable. What we observe as durability may be a particular equilibrium holding rather than an underlying robustness.

When cooperation has held for a long time, what should we be watching: the cooperation, or the conditions sustaining it?

References & Further Reading

1. The Evolution of Cooperation: Robert Axelrod

2. The Complexity of Cooperation: Robert Axelrod

3. The Evolution of Reciprocal Altruism: Robert Trivers

4. Evolution and the Theory of Games: John Maynard Smith

5. Governing the Commons: The Evolution of Institutions for Collective Action: Elinor Ostrom

6. Institutions and the Path to the Modern Economy: Lessons from Medieval Trade: Avner Greif

In the autumn of 1993, MG Refining and Marketing — the United States energy subsidiary of the German industrial conglomerate Metallgesellschaft AG — held a position that, on paper, was a textbook hedge. The firm had sold long-term forward contracts to deliver heating oil, gasoline, and diesel to commercial customers at fixed prices, in some cases ten years out. To offset the resulting price exposure, MGRM had built a roughly equal notional hedge using near-dated NYMEX energy futures and short-dated over-the-counter swaps, rolled forward month by month — a stack-and-roll hedge.

The economics looked clean in spot-price space. If oil rose, the long-term supply contracts became more onerous, but the hedge appreciated. If oil fell, the contracts gained value, the hedge lost. Against directional spot moves, the hedge was substantially complete.

Through the second half of 1993, oil prices fell sharply. WTI crude declined from around $20 a barrel in early June to about $14 by mid-December. The hedge position, marked to market daily, generated cash margin calls on a scale that consumed MGRM’s available funding. The supply contracts also gained economic value under the spot-price logic of the hedge, but those gains were long-dated, illiquid, and contested in valuation. Margin calls demanded cash today; offsetting gains existed only on a discounted-cash-flow schedule running out to 2003.

By December 1993, the supervisory board of the German parent, alarmed at margin calls in the hundreds of millions of dollars, dismissed executive chairman Heinz Schimmelbusch and appointed a new management team that liquidated the futures positions and unwound the customer contracts. The realised loss, on the auditors’ calculation, was approximately $1.3 billion. To the extent those gains existed, they were not monetised on a timetable that funded the margin calls, and much was lost when the programme was terminated.

The puzzle is not whether the hedge was right in theory. Subsequent academic work continues to debate that, with serious arguments on both sides. The puzzle is structural. A position designed to match one projection of price exposure failed catastrophically when other control-relevant dimensions — basis, maturity mismatch, and the timing of cash flows under funding stress — became binding.

What kind of problem is this?

When a control system maps fewer dimensions than the system it is meant to control, when the regulator’s response set covers some kinds of disturbance and is silent on others, when the disturbance the regulator cannot recognise is the one that arrives — that is a specific shape of problem. And the discipline that formalised it, in the middle of the twentieth century, is cybernetics.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Cybernetics was developed in the 1940s and 1950s by Norbert Wiener, W. Ross Ashby, and others, working initially on biological control, electrical engineering, and the design of self-regulating systems — thermostats, gun-laying servos, neural feedback. Markets do not satisfy its assumptions cleanly: participants are reflexive and adaptive, regulation reshapes what is regulated, and disturbances arrive in forms no designer can enumerate in advance. But the structural logic — that any regulator must match the variety of the system it confronts, and must internally embody a model of that system on the dimensions that matter — transfers.

Where Variety Was Missing

W. Ross Ashby, a British psychiatrist who became one of the founders of cybernetics, published An Introduction to Cybernetics in 1956. Its central result, the Law of Requisite Variety, is one of the most compact statements in the formal study of control. Ashby’s slogan was: only variety can destroy variety.

The technical content is precise. Variety is the number of distinguishable states a system can occupy. A coin has variety two; a six-sided die has variety six; a portfolio with twenty independent risk factors, each potentially in any of several regimes, has variety in the millions. In its information-theoretic form, Ashby’s law sets a floor on residual variety: the variety appearing at the essential variable cannot be reduced below the variety of the disturbance set minus the effective variety of the regulator, unless the structure of the system itself attenuates it.

The implication is structural rather than tactical. There is a minimum effective complexity below which the specified degree of control is impossible. No amount of speed, discipline, or effort substitutes for variety the regulator does not possess.

Apply this to MGRM. In spot-price space, the hedge had substantial variety. But matching variety was insufficient across basis, maturity, and funding space. A fall in spot prices generated immediate cash losses on the stacked short-dated hedge, while the offsetting gains emerged only gradually, depended on customer behaviour, and were not pledgeable as collateral. Contango was a term-structure disturbance that produced roll losses and funding stress. The regulator suppressed one channel of disturbance while leaving others open; in Ashby’s terms, residual variety had a route to the essential variable when the liquidity disturbance arrived.

The Hidden Map

In 1970, Ashby and Roger Conant published a paper with one of the more striking titles in the cybernetic literature: “Every Good Regulator of a System Must Be a Model of That System.” The result extends the law of requisite variety into a structural theorem about the regulator’s internal organisation.

The Conant-Ashby theorem, stated informally, holds that the simplest optimal regulator must contain a model of the system, in the sense that the regulator’s events map the relevant events of the system. A regulator that does not encode such a mapping, on the dimensions where regulation is required, cannot be the simplest optimal regulator there.

The consequence is practical. A risk control system that does not model funding dynamics cannot regulate funding risk optimally, however sophisticated its handling of price risk. A governance protocol that does not model the difference between mark-to-market and economic loss cannot regulate the response to drawdowns optimally, however disciplined the decision-makers.

MGRM provides two instances. The trading desk’s hedge was a model of one projection of price dynamics, and a precise one. It was not a sufficiently rich model of basis, maturity mismatch, and funding dynamics. In Conant-Ashby terms, the regulator did not map the system on the dimensions that later became decisive.

The Frankfurt supervisory board, by late 1993, faced the position through a frame in which mark-to-market losses were salient and the long-dated supply contracts’ offsetting value less so. Whether a richer frame would have led to a different choice is contested — Culp and Miller argued yes, Mello and Parsons no, on the grounds funding stress made continuation infeasible regardless. Either way, the board needed a usable model of the joint state — economic value, funding capacity, governance tolerance — and the regulator on the second floor had to map a different system from the regulator on the trading floor.

Where Compression Costs

The temptation to reduce variety is structural in capital allocation. Variety is expensive to monitor, govern, and explain. Compression makes oversight tractable, simplifies committee discussion, and lets a complex system be presented to non-specialist principals. In each of three places, the price of compression is paid in residual variety reaching the goal.

The first is risk measurement that distils a multi-dimensional system into a single number. Value-at-risk, expected shortfall, single-figure stress-test outputs — each is a useful summary and a poor standalone regulator. They compress a high-dimensional state into one measurement channel. The danger is not the scalar itself but the response rule attached to it. If the rule does not preserve the distinctions that matter for action — funding liquidity, basis, crowdedness, counterparty fragility, liquidation horizon — the missing variety reappears at the essential variable.

The second is governance escalation paths that map one dimension. A protocol that says “intervene if drawdown exceeds X” maps drawdown. The dimensions it omits — counterparty stress, headline risk, succession dynamics, the difference between permanent capital impairment and recoverable mark-to-market loss — are precisely where senior intervention shapes outcomes. The protocol’s variety is fixed; the situation’s variety is not.

The third is rule-based portfolio construction that selects a fixed response to a variable system. Static allocation grids, fixed-period rebalancing schedules, and tilt strategies with constant weights all have determined response variety. Markets do not. When the regime occupies a state the rule does not distinguish, the rule responds with whatever it has, which may be wrong in the dimension that has changed.

None of these compressions are wrong as summaries or as starting points. They become wrong when they are mistaken for regulators.

Two earlier posts intersect this framework. The Path The Maths Misses showed that a calculation correct on the ensemble average can be ruinous on the realised time-path. Variety mismatch is the cybernetic complement: a regulator whose response set spans the average-path disturbance but not the realised-path disturbance is undermatched in exactly that sense.

The Model That Fights Back described why a system under pressure intervenes harder rather than revising its model — the precision trap, the complexity cost of moving prior beliefs. Coupled with requisite variety, the trap closes mechanically: an under-varietied regulator cannot afford to acknowledge missing dimensions, because acknowledgement requires rebuilding rather than running harder. Conant-Ashby suggests why a richer model would have been required for an optimal response. The free energy framework suggests why such a model is hard to build under pressure.

The Capacity Question

When a control system is failing, the first question is not whether it is being applied with sufficient discipline but whether it has the variety to recognise and respond to the disturbance arriving. If not, more discipline cannot rescue it; only more effective variety — or a redesign that attenuates the disturbance before it reaches the controlled variable — can. Running an under-varietied regulator harder produces faster failure, not eventual success.

The diagnosis applies wherever a control system maintains something against disturbance. In organisations, a control system that maps performance against budget but not capability accumulation will under-regulate the dimension that matters most for long-run survival. In professional practice, a quality regime that maps procedure but not judgement leaves residual variety to whatever individuals improvise. In health, a regimen tracking measurable biomarkers but not subjective state under-regulates the channel through which much of the relevant disturbance arrives.

The discipline is in asking: not whether the regulator is being run well, but whether it has the structural variety to recognise and respond to what it is being asked to control?

References & Further Reading

1. An Introduction to Cybernetics: W. Ross Ashby

2. Design for a Brain: The Origin of Adaptive Behaviour: W. Ross Ashby

3. Every Good Regulator of a System Must Be a Model of That System: Roger C. Conant & W. Ross Ashby

4. Metallgesellschaft and the Economics of Synthetic Storage: Christopher L. Culp & Merton H. Miller

5. Maturity Structure of a Hedge Matters: Lessons from the Metallgesellschaft Debacle: Antonio S. Mello & John E. Parsons

6. Metallgesellschaft: A Prudent Hedger Ruined, or a Wildcatter on NYMEX?: Stephen Craig Pirrong

7. Derivatives Debacles: Case Studies of Large Losses in Derivatives Markets: Anatoli Kuprianov

8. The Cybernetic Brain: Sketches of Another Future: Andrew Pickering

On the morning of 16 September 1992, sterling was pinned to its lower bound inside the European Exchange Rate Mechanism. The framework had been clear for nearly two years. Sterling would hold within a band against the Deutsche Mark. The Bank of England would defend the band. Domestic monetary policy would be subordinated to it. In return, Britain would import the Bundesbank’s anti-inflationary credibility.

By that morning, every part of the framework was being tested simultaneously. The Bundesbank had raised rates to absorb the fiscal shock of German reunification. Britain was in recession, with rising unemployment and falling property prices. Domestic conditions called for lower rates. The framework called for keeping UK rates high enough to defend the parity. The market noticed the gap.

At around 11:00 the government announced an increase in the base rate from 10 to 12 per cent. Sterling continued to fall. An emergency announcement promised a further increase to 15 per cent the following morning. Sterling continued to fall. By 19:40, the Chancellor of the Exchequer announced that the United Kingdom was suspending its membership of the ERM. The 15 per cent rate would never take effect. Treasury papers declassified in 2005 put the estimated loss at around £3.3 billion, measured at February 1994.

The framework had been internally coherent. It bought one desirable property — anti-inflationary credibility — at the price of another, monetary autonomy. As long as those two were not in conflict, the price was unobservable. The day they came into conflict, it was paid in a single afternoon.

What kind of problem is this?

The ERM was not a bad framework that turned out to be wrong. It was a framework with a particular structure: a set of desirable properties that could not all be jointly held. Free capital movement, a fixed exchange rate, and independent monetary policy form what Robert Mundell formalised in 1963 as the impossible trinity. Any two are achievable; all three are not. Britain by 1992 had committed to the first two and was discovering that the third was not available.

When a system of desirable properties cannot all be held simultaneously, when committing to one set forces a choice about which to give up, when the choice itself is the cost of coherence — that is a specific shape of problem. And the discipline that has thought most rigorously about systems of properties that cannot all be satisfied is the foundations of quantum mechanics.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

The foundations of quantum mechanics work with entangled particles and photon pairs, physical systems whose correlations can be measured in the laboratory. Markets are messier: there is no equivalent apparatus, participants adapt, and committing to a framework shapes the portfolio. But the structural logic transfers. When a set of desirable properties is mutually inconsistent, every coherent position must surrender one of them, and the choice of which to surrender is the price of taking a position.

Three Truths, At Most Two

In 1964, the Northern Irish physicist John Stewart Bell published “On the Einstein Podolsky Rosen Paradox”. It addressed whether quantum mechanics is a complete description of reality, or whether hidden variables would restore a more classical picture in which particles have definite properties at all times.

Bell took three assumptions that any sensible classical theory ought to satisfy. The first was locality: nothing influences anything else faster than light. The second was definite outcomes: measurements yield single, classical values, and those values reflect properties the system possessed prior to being measured. The third was measurement independence: experimenters can choose what to measure freely, without their choices being correlated with the hidden state of the system.

From these assumptions, Bell derived an inequality that set a ceiling on how strongly distant measurements of entangled particles could correlate. Quantum mechanics predicts that this ceiling can be exceeded. The two predictions disagree by an amount testable in the laboratory.

In 1982, Alain Aspect’s group in Paris ran a landmark version of the experiment. Quantum mechanics won. In 2015, three independent groups closed the main experimental loopholes. Quantum mechanics won again. In 2022, the Nobel Prize in Physics was awarded to Aspect, John Clauser, and Anton Zeilinger for this work.

The implication is uncomfortable. Those three assumptions cannot all be true together. Locality, definite outcomes, or measurement independence: one must go. The empirical result establishes that. It does not tell us which.

Four Ways To Pay

The leading responses to Bell’s theorem either reinterpret the standard quantum formalism, as Many Worlds, Bohmian mechanics, and Copenhagen-type views do, or modify its assumptions, as superdeterminism does. What separates them is which of Bell’s assumptions each is prepared to surrender.

Many Worlds, originated by Hugh Everett in 1957 and developed by Bryce DeWitt and David Deutsch, keeps dynamical locality, measurement independence, and the realism of the underlying wavefunction. The price is the single-outcome component of definite outcomes: there is no unique result, but a branching structure in which every quantum-allowed outcome is realised in some branch.

The pilot-wave or Bohmian interpretation, developed by Louis de Broglie and David Bohm, keeps definite outcomes — particles have positions at all times — and measurement independence. The price is locality at the hidden-variable level: the guiding wavefunction ties distant particles together, though not in a way that permits faster-than-light signalling. Lee Smolin notes a further asymmetry: the wave influences the particle but the particle does not influence the wave. That asymmetry is part of the price.

The Copenhagen interpretation and its modern descendants keep operational locality and measurement independence. The price is the other component of definite outcomes: classical values cannot be assigned to observables outside a measurement context. The wavefunction describes potentialities; values come into being with the act of measurement. The result is a strange relationship between observer and observed that has bothered physicists for a century.

Superdeterminism, advocated by Sabine Hossenfelder and others, keeps locality and definite outcomes by giving up measurement independence. Experimenters’ choices are correlated with the hidden state of the system being measured. Critics argue that this dissolves the basis of experimental reasoning; proponents argue that the price, properly understood, is bearable.

Each of these is a coherent response to Bell’s constraint. The first three are usually presented as empirically equivalent interpretations of the standard quantum formalism; superdeterminism is more radical, modifying assumptions rather than reinterpreting them. Bell’s theorem proves that something must be surrendered. The position ruled out is the one that tries to keep all the incompatible commitments at once.

What Every Framework Surrenders

Investing under uncertainty has an analogous structural shape, even though the mathematics is different. The desirable properties that frameworks for capital allocation try to hold are abundant: high expected return, low volatility, resilience to drawdown, liquidity, regime independence, low cost, robustness to behavioural error. No portfolio holds all of them at once. The framework is a choice about which to surrender.

Naïve mean-variance optimisation gives up robustness to input error and regime change. The framework assumes returns are stationary and covariances estimable from the past. When the distribution that generated yesterday’s data is not the distribution generating tomorrow’s, the optimisation computes the wrong answer with great precision.

Risk parity gives up resilience to a world in which its diversifiers stop diversifying. The framework equalises risk contributions across asset classes by levering the lower-volatility ones, premised on diversifying behaviour that held through the 2010s. In 2022, when bonds and equities fell in concert, the leverage that had been the framework’s elegance became its liability.

Trend-following gives up the turning point. The framework profits when prices continue in their current direction and loses when they reverse. The price is paid at every reversal, when the strategy is committed to a direction the market has just abandoned.

Concentrated value gives up the resilience that comes from diversification. The framework trusts that conviction in a small number of positions outweighs the protection of being wrong about any one. When the theses fail together, the concentration carries the correlation it appeared to avoid.

Volatility selling gives up convexity. The framework collects steady option premium against the assumption that realised volatility will fall short of implied. When the assumption holds, the income is reliable; when it fails, the losses can arrive in concentrated bursts that outweigh years of accumulated premium.

Carry strategies give up resilience in risk-off regimes. The framework borrows cheaply in one currency or maturity and lends at higher yields in another, harvesting the spread. The unwind, when it arrives, can compress months of accumulated carry into days of forced selling.

Every framework names what it surrenders, even if adherents find it convenient not to look sometimes. The most expensive framework in any room is the one whose proponents claim it costs nothing. Frameworks whose costs are honestly identified can be hedged, complemented, or held with humility. Frameworks whose costs are denied compound until the world demands payment, as the ERM compounded from October 1990 until September 1992, and as LTCM’s framework compounded for four years until a New York Fed boardroom filled, in September 1998, with representatives of the institutions called upon to settle the cost.

Knowing What You Owe

When physicists weigh these interpretations against one another, the experiments cannot adjudicate among them. The debate turns on what each interpretation keeps, what it surrenders, and whether the price is acceptable. They are not free to escape Bell’s price. They are free only to choose how to pay it.

The same discipline is available to those who allocate capital. No framework escapes the price. The unknowns are too dense, the data too sparse, the regimes too non-stationary for any single framework to be both complete and free. The framework is its trade-off, not its content.

This lesson extends beyond finance. In engineering, every design surrenders something: weight reduction can cost fatigue resistance, redundancy costs efficiency. In clinical medicine, every treatment surrenders something: aggressive intervention costs tolerability, sensitive screening costs specificity. The professional who claims a design or regimen with no downside has not escaped the trade-off; they have hidden it.

What follows is not relativism. Some frameworks are better than others for a given problem. Bell’s theorem itself rules out positions that try to keep everything. But within the space of internally coherent positions, the choice is between which truth you find least painful to give up. Pretending no truth is being given up is the only position provably wrong.

The discipline is in asking: what is the price my framework is paying, and am I paying it knowingly?

References & Further Reading

1. Speakable and Unspeakable in Quantum Mechanics: John Stewart Bell

2. Einstein’s Unfinished Revolution: Lee Smolin

3. The Quantum Challenge: Modern Research on the Foundations of Quantum Mechanics: George Greenstein and Arthur Zajonc

4. International Economics: Robert Mundell

5. Manias, Panics, and Crashes: A History of Financial Crises: Charles P. Kindleberger and Robert Z. Aliber

In late October 1988, the board of RJR Nabisco received what was, on the surface, a straightforward decision. The company’s chief executive, Ross Johnson, had proposed a management-led leveraged buyout that would take the company private. Within weeks, Kohlberg Kravis Roberts had submitted a competing bid. By the time the board announced KKR’s victory on 30 November, the contest had drawn international attention and produced one of the largest leveraged buyouts in history. The eventual price, around $25 billion, would stand as a record for years.

The conventional account of the auction focuses on the price competition. The bids escalated. The financing structures grew more complex. What received less attention was the structural fact at the centre of the case. The auction was managed by a special committee of the RJR board, chaired by Charles Hugel, which the directors had constituted for that purpose. The committee defined the criteria. The committee evaluated the bids against those criteria. The committee supplied the recommendation around which the full board’s decision formed. Whatever each bidder offered in headline price, the operative question was always how the bid would land with the committee under the criteria the committee itself had set.

By 30 November, when the board announced its decision, the headline numbers had converged closely enough that the price differential was not the decisive factor. KKR had structured its bid in a way the committee judged superior — employee protections, a plan to keep more of the company intact, more credible financing, a stronger securities package for existing shareholders, and a cleaner governance path after the transaction. The management group, despite having been the architects of the original transaction and despite a competitive headline price, did not prevail because the committee’s criteria were not reducible to price alone.

The puzzle is straightforward. The headline weights — the bid prices — were not the operative quantity. The operative quantity was the committee’s structural position: no auction outcome was likely to survive the process without the committee’s affirmative recommendation, governed by criteria broader than price. Bidders who treated the auction as a price contest misread the structure of the decision.

What kind of problem is this?

When the visible quantity — claim size, vote weight, headline allocation — diverges systematically from the operative quantity, which is how often an actor is the one whose agreement an outcome cannot do without, that is a specific shape of problem. And the discipline that has formalised the conditions under which power, in any voting or coalitional system, can be measured rigorously is cooperative game theory.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Cooperative game theory was developed by John von Neumann and Oskar Morgenstern in the 1940s and refined by Lloyd Shapley in the 1950s, built for abstract games with well-defined rules, payoffs, and players. Restructurings and corporate governance do not satisfy its assumptions cleanly: actors do not always have well-defined preferences, value cannot always be freely transferred between parties, the rules themselves are sometimes negotiable. But the structural logic — the habit of asking who is pivotal rather than who is largest, and the recognition that the gap between the two is where misjudgment of power lives — transfers cleanly. The vocabulary for marginal contribution to coalitions and the axiomatic basis for power allocation transfers.

Where Power Sits

A player is pivotal, in the technical sense, when their joining is the move that converts a losing coalition into a winning one — when their addition is the threshold-crosser. Pivotality is the formal property of how often an actor occupies that position across the structure of possible coalitions.

Lloyd Shapley’s result, circulated by RAND in 1952 and published canonically in 1953, is one of the cleanest in mathematical economics. He proved that there is a unique way to allocate the total value created by a coalition across its members that satisfies four axioms.

The first is efficiency: the total value is fully distributed across the players, with nothing left over. The second is symmetry: two players who contribute identically to every coalition they could join receive equal allocations. The third is linearity: when two coalition games are combined, the allocation in the combined game is the sum of the allocations in the components. The fourth is the dummy property: a player who contributes nothing to any coalition receives nothing.

Any allocation rule satisfying all four axioms is the same rule. Shapley showed it has a specific form: each player’s allocation is their average marginal contribution across all possible orderings in which the coalition could form. In the special case of a simple voting game, this becomes a direct measure of pivotality. A player who never tips the coalition has Shapley value zero, regardless of headline weight.

The application of Shapley’s framework to formal voting bodies — committees, legislatures, weighted majority games — was developed by Shapley and Martin Shubik in 1954; the resulting Shapley-Shubik power index is the canonical measure of pivotality in voting systems. A short worked example makes the point: a three-member board with voting weights of 49, 49, and 2, and a simple-majority threshold. Any minimal winning coalition requires exactly two members. Run through the orderings: the 2 per cent member tips the coalition as often as either 49 per cent member. Shapley value: one-third each. Headline weight, in this case, misleads completely.

The Banzhaf index, proposed by John Banzhaf in 1965 in a paper analysing the Nassau County Board of Supervisors, supplies a closely related measure — the probability that a player is pivotal in a randomly drawn coalition rather than the marginal contribution averaged across orderings. It often gives different numerical answers but the same structural insight.

The point is not the specific number. A rigorous treatment of power in any voting or coalitional system can produce an allocation that bears little resemblance to the headline weights. RJR is not a literal weighted-voting game — the committee was an institutional gatekeeper, not a 2 per cent voter — but the structural lesson transfers: the actor whose assent converts a proposal into an outcome can hold power far in excess of any headline economic exposure.

Where Headline Weight Misleads

The pattern recurs wherever multiple actors with formal voting weights or structural positions must reach a threshold to act. In investing, three places matter, and each is sustained by a specific mechanism that produces the gap between pivotality and headline weight.

The first is where the threshold is set, not crossed — board committees and structured auction processes. RJR is the leading example. When a board constitutes a special committee with effective authority over the sale process, the committee becomes pivotal because it owns the definition of the threshold itself. Its criteria are the operative constraint. The economic exposures of the parties being weighed are not. Even within ordinary voting games, the threshold dominates the weights: with holdings of 45, 45, and 10 and a simple-majority quota, the 10 per cent holder has the same formal pivotality as either 45 per cent holder. Raise the quota to 90 per cent and the same holder becomes irrelevant, because the two larger blocs together meet the threshold while 45 plus 10 falls short. The effect is mechanical but not monotone, and it has to be computed. Bidders and counterparties who model the situation as a contest of headline weights miss the operative quantity entirely.

The second is where the larger blocs disagree — governance and proxy contests. When two large holders align and together clear the threshold, a smaller third holder is practically a dummy. When they oppose, and neither side can reach the threshold alone, the same small holder can become decisive. Pivotality is conditional on the alignment structure, not on the weights alone. This is where the classical indices reach the limit of their usefulness: Shapley-Shubik and Banzhaf are a priori measures that abstract from actual preferences and treat coalitions according to specified symmetry assumptions. In institutional analysis, the relevant pivotality is preference-conditional. The formal weights are the starting point; the alignment model determines which coalitions are realistic.

The third is where structural position outruns claim size — gatekeeping intermediaries with statutory or contractual rights. Trustees in bond indentures, agents in syndicated loan agreements, fiscal agents in some kinds of transaction — each can hold influence not reducible to economic exposure. The mechanism that creates this divergence is legal or contractual; the mechanism that causes others to misprice it is informational. Counterparties who fail to recognise the gatekeeper’s structural position misallocate effort, competing on the visible quantity rather than positioning for the operative one. In the RJR case, this is what the bidders treating the auction as a price contest got wrong; in restructurings, it is what creditors focused on claim seniority routinely miss.

This connects to two earlier posts. What Others Think Others Think showed how higher-order beliefs — what each actor thinks others think, recursively — can dominate first-order analysis in coordination games; the pivotality structure of a coalition is the cooperative-game-theory analogue, where pivotality depends on what each actor believes about the alignments of the others. The Network You Cannot See showed that systemic risk lives in the edges between participants. The Shapley framework supplies a complementary insight: power lives in the pivotality structure between participants, and headline weights tell you about nodes, not edges.

The Diagnostic

The three places map to three diagnostic questions. Before judging power in any coalition situation, work through them in order.

What is the threshold structure? Simple majority, supermajority, unanimity? Different thresholds produce different pivotality distributions over the same voting weights, and the mapping is not intuitive without explicit analysis.

What is the correlation of preferences among the large actors? Pivotality is conditional on this, and the small holder’s position can shift from decisive to irrelevant as alignments shift.

Where is pivotality sitting given those two answers? The answer often diverges sharply from headline weight, and the divergence is where institutional intuition needs to be retrained.

The framing applies wherever multiple actors with formal voting weights must reach a threshold to act. In public policy, legislative coalitions exhibit the same pattern: small parties hold disproportionate power under supermajority rules when the larger parties divide. In standards bodies, technical specification authors with formally equal voting rights have unequal pivotality given correlation patterns among the larger participants. In multi-employer pension plans and joint ventures, the headline structure of voting weights and the operative structure of pivotality routinely diverge.

The discipline is in asking: not who owns the largest share, but who would the coalition need to win, and how does that change when the threshold changes?

References & Further Reading

1. A Value for n-Person Games: Lloyd Shapley

2. Theory of Games and Economic Behavior: John von Neumann & Oskar Morgenstern

3. The Model Thinker: Scott E. Page

4. Weighted Voting Doesn’t Work: A Mathematical Analysis: John F. Banzhaf III

5. Barbarians at the Gate: The Fall of RJR Nabisco: Bryan Burrough & John Helyar

6. A Method for Evaluating the Distribution of Power in a Committee System: Lloyd Shapley & Martin Shubik

7. RJR Nabisco: A Case Study of a Complex Leveraged Buyout: Allen Michel & Israel Shaked

In February 1994, the Federal Reserve raised its federal funds target for the first time in five years. The move caught markets off guard — leveraged bond portfolios suffered immediate losses. But most survived. Then the Fed raised again in March. And again in April. Over the course of the year, rates rose from 3 per cent to 5.5 per cent across six increments.

The first shock was sharp but absorbable. What followed was worse. Each subsequent rate rise eroded holdings, consumed capital, and tightened leverage ratios in structures already weakened by February’s blow. The portfolio that survived the initial surprise was slightly weaker entering March. The one that survived March was weaker still entering April. By the time the fifth and sixth increases arrived, funds that had comfortably absorbed any individual move were running on depleted reserves and unable to withstand shocks they would have shrugged off nine months earlier. Several prominent funds collapsed; Orange County declared what was then the largest municipal bankruptcy in American history.

The risk models had been asked: can this portfolio survive a rate rise of this magnitude? They had answered correctly — for any single move, yes. The question they had not been asked was: what happens to the portfolio’s capacity to absorb shocks after it has already absorbed several in succession?

What kind of problem is this?

When a system appears robust under any individual stress but fails after repeated exposure; when failure originates not at the point of greatest average load but at features that concentrate stress locally; when damage accumulates invisibly and then propagates suddenly — that is a specific shape of problem.

And the discipline that has spent a century studying how flaws concentrate stress is fracture mechanics.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Fracture mechanics was built for metals, ceramics, and composites — materials with well-characterised properties under controlled loading. Markets are messier: participants adapt, structures evolve, and measuring risk can change the system being measured. But the structural logic — the habit of asking where stress concentrates and how damage accumulates under repeated loading — transfers.

Strength and Its Flaws

The discipline begins with a puzzle that troubled engineers for decades: materials fail at loads far below their theoretical strength. In 1921, the British engineer A.A. Griffith resolved this by showing that fracture is not about the average stress in a material but about the energy balance at the tip of a crack. A crack will grow when the strain energy released by its extension exceeds the energy required to create new surfaces. Below that threshold, the crack sits dormant. Above it, it propagates — and in brittle materials under the right loading conditions, almost instantaneously.

Griffith’s insight was that strength is not a property of the material alone. It is a property of the material and its flaws. A sheet of glass is theoretically strong enough to support enormous loads. But a tiny scratch concentrates stress at the crack tip to levels thousands of times higher than the nominal load, and the glass shatters. The scratch does not weaken the material; it reorganises how force flows through it, creating a local intensity that the average measurement misses entirely.

The discipline made this concrete in 1954, when two de Havilland Comet airliners — the world’s first commercial jets — disintegrated in flight. Investigators found that the square corners of the passenger windows concentrated stress far beyond what the overall fuselage load would suggest. Pressurisation cycles drove microscopic cracks at these corners — damage too small to detect on any individual flight, but accumulating irreversibly until a crack reached critical length and the structure failed catastrophically. The fuselage had been strong enough for any single pressurisation. It was not strong enough for thousands.

Two concepts from the discipline are particularly useful for our purposes. The first is the stress intensity factor — a measure of how much stress is amplified at the tip of a flaw given its geometry and the applied load. Two structures carrying identical average loads can have vastly different stress intensities at their critical points. The structure with the sharper flaw is closer to failure, even though nothing about the average load distinguishes them.

The second is fatigue: the process by which repeated loading below the failure threshold drives incremental crack growth. The Paris law — an empirical relationship describing crack growth in its stable mid-range regime — reveals that the process is self-reinforcing: as the crack grows, the stress intensity increases, which accelerates the growth rate, which lengthens the crack further. Long periods of apparently stable operation are followed by rapid acceleration and sudden failure.

Where Stress Concentrates

What does this framework reveal when applied back to portfolios?

Start with stress concentration. The standard tools of portfolio risk — volatility, correlation, value-at-risk — are aggregate measures. They describe the overall stress field. They do not identify the features that concentrate stress locally, the points where the experienced load far exceeds the nominal one.

In a portfolio, stress concentrations take several forms. Illiquid positions concentrate stress because they cannot shed load when the overall portfolio is under pressure. A portfolio with 5 per cent in an illiquid asset does not have 5 per cent exposure in a drawdown; it has whatever-is-left-after-selling-the-liquid-assets exposure, which can be far higher. The illiquidity is the sharp corner in the window — a geometric feature that reorganises how stress flows through the structure.

Leverage is another concentrator. A leveraged fund does not simply have more exposure; it has a feedback mechanism where losses trigger margin calls, which force sales, which create further losses. The leverage concentrates stress at the point of forced selling, creating local intensities that exceed what the nominal exposure would suggest. In the 1994 bond rout, this was the mechanism that turned a sequence of modest rate rises into fund-ending events: each move consumed capital, which tightened leverage constraints, which forced selling, which locked in losses that further consumed capital.

Funding mismatches concentrate stress similarly. A portfolio of long-dated assets financed with short-term borrowing has a stress concentration at the rollover point. Under normal conditions, the mismatch is invisible — the funding renews routinely. Under stress, it becomes the point of failure. This was the structure of Northern Rock, of the conduits and SIVs in 2007-08, and of many carry trades before and since. The ergodicity problem we explored in The Path The Maths Misses applies here directly: the ensemble average across many possible funding states looks manageable, but the time-average path through a funding crisis is not.

Invisible Damage

Now consider fatigue. Portfolios do not experience a single load and either hold or fail. They experience repeated loading cycles — drawdowns, recoveries, periods of underperformance, liquidity squeezes that come and go. Each cycle can leave invisible damage.

What does damage look like in a portfolio? It is the erosion of the buffers that allow the portfolio to absorb future stress. Capital consumed by losses. Liquidity reserves drawn down. Risk budgets exhausted. Organisational patience depleted. Each drawdown-and-recovery cycle that looks like the system returning to normal may in fact be the system returning to a slightly weaker version of normal — the crack a little longer, the stress intensity a little higher, the margin of safety a little thinner.

This connects to the stability framework we developed in What Makes Stable Things Break. There, we described how a system can appear stable while the basin of attraction around it is shrinking — the ball still sitting at the bottom of the bowl while the bowl gets shallower. Fatigue is one mechanism by which the basin shrinks. Each loading cycle erodes the structural reserves that define the basin’s depth, even as the system returns to what looks like the same resting state.

The Paris law dynamic has a portfolio analogue. As buffers erode, the portfolio becomes more sensitive to the next stress event — remaining positions more concentrated, liquidity cushion thinner, organisational willingness to hold through pain diminished. The damage from one cycle feeds into the severity of the next.

This is why portfolios can survive ten drawdowns and fail on the eleventh, even when the eleventh is no larger than the ones before it. The question is not how large the current stress is but how much damage has already accumulated. It was not the final pressurisation cycle that destroyed the Comet. It was every cycle before it.

The stress concentrations — illiquidity, leverage, funding mismatches, counterparty dependencies — are often in the least visible parts of the structure. In The Network You Cannot See, we explored how hidden connections create links between apparently independent positions. Those hidden links are precisely where fatigue damage accumulates undetected.

The standard risk report is a static strength test. It asks: can the portfolio survive this scenario? But it does not ask: how has repeated stress changed the portfolio’s capacity to survive the next scenario? It does not track the sub-critical damage.

The Fragility Question

What might a fracture-mechanics-informed approach look like? Not prediction — materials science does not predict when a specific component will fail. But it can change what you monitor. Map the stress concentrations — illiquid positions, leveraged positions, funding or counterparty concentration — and give them scrutiny disproportionate to their weight. Track cumulative damage, not just current state: is capital being rebuilt after drawdowns, or is each recovery incomplete? And distinguish between damage-tolerant structures — diversified, liquid, modestly leveraged — and damage-intolerant ones — concentrated, leveraged, dependent on a single thesis or a single source of funding. The Comet disasters accelerated a shift in aircraft design toward damage-tolerant philosophy: assume cracks will form, design so that they grow slowly enough to be detected, and build in redundant load paths so that no single crack can bring down the structure. The difference matters enormously under repeated loading, even if both structures look adequate under a single static test.

The framework extends beyond finance. In organisations, repeated restructurings can fatigue institutional capacity even when each individual restructuring appears successful. The formal structure recovers; the informal networks accumulate damage that is not repaired by the next reorganisation. In infrastructure, repeated minor flooding events can degrade foundations, embankments, and drainage systems in ways that each post-event inspection clears as sound. The recovery looks complete from the outside. The damage is internal and cumulative.

The discipline is in asking: not whether something can survive this shock, but what has every previous shock already left behind?

References & Further Reading

1. Fracture Mechanics: Fundamentals and Applications: T.L. Anderson

2. Structures: Or Why Things Don’t Fall Down: J.E. Gordon

3. The Stress Analysis of Cracks Handbook: Hiroshi Tada, Paul Paris & George Irwin

4. Against the Gods: The Remarkable Story of Risk: Peter Bernstein

5. Big Bets Gone Bad: Derivatives and Bankruptcy in Orange County: Philippe Jorion

6. The (Mis)Behaviour of Markets: Benoit Mandelbrot & Richard Hudson

Beginning in October 1979, Paul Volcker’s Federal Reserve adopted a tightening regime that would push the federal funds rate above twenty per cent by early 1981. The objective was to break an inflationary spiral that had persisted for more than a decade. The cost was enormous: a deep recession, unemployment approaching eleven per cent, a wave of corporate and agricultural bankruptcies. The intervention was deliberate and blunt. Inflation was not declining because the economy had found a new equilibrium. It was declining because the Federal Reserve was suppressing demand with sufficient force to override the prevailing inflationary dynamics.

By any reasonable measure, the stability was imposed. The Fed was holding the economy in a state it would not have reached on its own, at a cost that was visible, escalating, and politically unsustainable. If the intervention had been withdrawn in 1980 or 1981, inflation would likely have re-accelerated — credibility had not yet been secured and the underlying dynamics had not changed.

But something happened over the following decade. Inflation expectations anchored. Wage-setting behaviour adapted to a low-inflation environment. Firms began pricing on the assumption that inflation would stay low, and their pricing decisions helped make it so. Central bank credibility — earned through the severity of the initial intervention — became a structural feature of the economy rather than an ongoing act of will. By the early 1990s, the Fed did not need to suppress the economy to keep inflation low. The system’s own dynamics were doing it.

The stability had been imposed. And then it took root.

For roughly three decades, this self-sustaining low-inflation equilibrium helped underpin the bond rally and, over time, the negative stock-bond correlation on which much multi-asset portfolio construction relied. It required far less effort to maintain than it had to establish — anchored expectations, credible institutions, and self-reinforcing behaviour did most of the work. Perturbations were absorbed. Oil shocks came and went. Recessions came and went. Each time, the system returned to low inflation without requiring Volcker-scale intervention.

What kind of problem is this?

When the question is not whether an equilibrium holds today but whether genuine stability has formed — when you need to distinguish between a system that is self-sustaining and one that is still being held in place by force that could be withdrawn — that is a specific shape of problem.

And the mathematical discipline that has formalised the conditions under which an equilibrium is self-restoring after disturbance is Lyapunov stability theory.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Lyapunov theory was built for mechanical and electrical systems — pendulums, circuits, control loops. Markets do not satisfy its assumptions cleanly: participants adapt, dynamics are reflexive, intervention is endogenous. But the structural logic — the habit of asking whether disturbances are being endogenously damped under the current regime — is close enough to our problem to borrow. The vocabulary for energy dissipation, equilibrium conditions, and the distinction between stability that is earned and stability that is purchased transfers.

The Function That Tells the Truth

Aleksandr Lyapunov, a Russian mathematician working in the 1890s, asked a deceptively simple question: how can you determine whether a system will return to equilibrium after a disturbance, without having to solve the full equations of motion? His answer was elegant. Find a scalar function — think of it as a measure of the system’s “energy” — that is positive away from equilibrium and that strictly decreases along the system’s trajectories. If such a function exists, the system is not merely stable but asymptotically stable: perturbations do not just stay bounded — they actively decay. The equilibrium attracts.

The power of the approach is that you do not need to trace every possible trajectory. You need to find the right function — the right measure of how much the system has been disturbed — and show that it is shrinking over time. If the function is shrinking, the disturbance is decaying. If it begins to grow, the system is moving away from equilibrium even if the state variables themselves appear calm. If it flatlines — neither growing nor shrinking — the situation is ambiguous: the system may be on the edge of stability or may need additional structure to settle.

The analogy to physical energy is deliberate. A ball in a bowl is at a stable equilibrium because its total mechanical energy — kinetic plus potential, the natural Lyapunov function for a mechanical system — decreases as friction dissipates its motion. The ball overshoots, oscillates, but each swing is smaller. The energy shrinks monotonically. The ball settles. If you remove the friction, the ball oscillates indefinitely — total energy is conserved, never decreasing, and the ball never settles, even though it keeps returning near the bottom. If you invert the bowl, the ball rolls away: the energy framework tells you immediately that the equilibrium is unstable.

From Imposed to Formed

The Volcker disinflation illustrates the most important transition the Lyapunov lens can diagnose: stability that begins as imposed and becomes genuine.

In 1979–82, the system’s internal disequilibrium was at its maximum — inflation expectations were deeply unanchored. Every perturbation — an oil price shock, a fiscal expansion, a wage negotiation — had to be met with further tightening. The displacement from equilibrium was not decreasing through the system’s own dynamics. It was being forced down by external intervention. This is the signature of imposed stability: the system requires continuous, costly effort to remain in its current state.

Over the following decade, the picture reversed. Inflation expectations had become self-reinforcing. The system’s internal disequilibrium — the degree to which expectations were unanchored, a useful proxy for displacement from equilibrium — was dissipating naturally. Consequently, the scale of external intervention required to maintain stability genuinely decreased after each shock. The 1990–91 recession, the 1997 Asian crisis, the 2001 dotcom bust: each produced economic stress, and each time inflation stayed anchored without extraordinary monetary force. The system was dissipating perturbations through its own dynamics. The stability had formed.

The practical question for an investor in the 2010s was not “is inflation low?” — it visibly was. The question was: has the low-inflation equilibrium genuinely formed, or is it still being imposed? The Lyapunov lens gave a clear answer: formed. Expectations were anchored. The self-reinforcing mechanism was operational. This mattered because it meant the equilibrium could be relied upon — not forever, but for as long as the structural conditions that produced it persisted.

The Floor That Never Took Root

In September 2011, the Swiss National Bank announced a floor on the EUR/CHF exchange rate at 1.20. The SNB declared it would enforce the floor “with the utmost determination” and was prepared to buy foreign currency “in unlimited quantities.” Capital flowed in on the assumption that the floor would hold.

For three years it did. But the Lyapunov diagnostic told a different story from the Volcker case. The cost of maintaining the floor was not decreasing. It was growing. The SNB accumulated foreign reserves from roughly 260 billion francs at end-2011 to over 510 billion by end-2014, and by early 2015 the required pace of intervention had become, in the SNB’s own later description, rapidly increasing. The system was not learning to sustain the equilibrium on its own. Market participants were not anchoring on EUR/CHF 1.20 in the way that firms had anchored on low inflation — they were leaning on the SNB’s commitment, and the leaning was getting heavier.

The stability was imposed. And it never took root.

On 15 January 2015, the SNB abandoned the floor without warning. EUR/CHF repriced by fifteen to thirty per cent intraday, depending on the venue — the biggest one-day fall of the euro against the franc in the pair’s history. Several foreign exchange brokerages became insolvent or required rescue. Funds that had treated the floor as an equilibrium suffered severe losses.

The contrast with the Volcker case is the diagnostic. Both began as imposed stability. In one case, the system’s own dynamics gradually took over — expectations anchored, behaviour adapted, the intervention could be withdrawn without the equilibrium collapsing. In the other, the system never developed self-sustaining dynamics. The intervention was the equilibrium. Remove it, and there was nothing underneath.

The structural logic suggests a useful extension: what we might call a stability budget. An imposed equilibrium consumes a budget continuously — the SNB’s budget was its institutional tolerance for balance sheet expansion, and that tolerance was being consumed at an accelerating rate. A genuinely formed equilibrium is not costless, but it is low-maintenance — the low-inflation regime after 1995 still depended on the implicit threat of future tightening, but it did not require Volcker-scale force to persist.

The transition from imposed to formed is, in these terms, the moment when the stability budget shifts from escalating to modest and episodic. The system moves from requiring large external force to dissipating perturbation energy largely on its own. Identifying that transition — or its absence — is the practical contribution of the lens.

The Formation Question

Lyapunov theory says nothing about timing. It cannot tell you when the SNB’s tolerance would run out, or when Volcker’s credibility would take hold. What it gives you is a structural test, and the test is specific.

Find the quantity that measures displacement from equilibrium. Not price. Not volatility. The internal state: how unanchored are expectations, how large is the intervention required to absorb a perturbation, how far is the system from a state it could sustain on its own. Then ask: is that quantity shrinking after each disturbance, or growing?

If it is shrinking through the system’s own dynamics, without increasing external force, stability is forming. You can rely on it, provisionally, for as long as the conditions that produce the self-correction persist. If it is shrinking only because something is paying to push it down, and the payments are getting larger, stability is imposed. You are relying not on the system but on the willingness of the imposer. And willingness is a decision, not a process. It can be withdrawn in an afternoon.

The test applies wherever stability is claimed. In professional standards, a norm that practitioners internalise and self-enforce is formed; a standard that holds only because of continuous external audit is imposed, and the distinction becomes visible the moment the auditor looks away. In organisations, a culture that self-reinforces through hiring and norms is formed; a culture that depends on a single leader’s force of personality is imposed, and the departure of that leader is the withdrawal of the floor.

The discipline is in asking: is the system’s displacement from equilibrium shrinking on its own, or is something paying to force it down?

References & Further Reading

1. The General Problem of the Stability of Motion: Aleksandr Lyapunov

2. Nonlinear Systems: Hassan Khalil

3. Stochastic Stability of Differential Equations: Rafail Khasminskii

4. Volcker: The Triumph of Persistence: William Silber

5. Thinking in Systems: A Primer: Donella Meadows

6. A History of Interest Rates: Sidney Homer & Richard Sylla

On October 15, 2014, the yield on the US 10-year Treasury note fell 16 basis points in six minutes, then nearly retraced over the next six. The day’s intraday range was 37 basis points — extraordinary for what is normally one of the most stable instruments in the world. No policy had changed. No institution had failed. After a modestly weak retail sales release, one-sided order flow met unusually thin depth and produced a sharp round-trip in the world’s deepest government bond market.

The Joint Staff Report published afterward found no single cause. One-sided order flow had overwhelmed the displayed depth. Principal trading firms sharply increased and then reversed their activity; bank-dealers intermittently withdrew. For a few minutes, the market’s capacity to absorb directional flow fell below the demand for immediate execution.

This was not a crisis. No one defaulted, no bank wobbled. But it revealed something important about liquidity. The standard mental model is that liquidity is a pool — a reservoir of available capital, deep or shallow. The pool metaphor is not wrong. But it is incomplete. Depth had already thinned before the event, but the pool had not emptied — capital had not fled the system. What had failed was the market’s ability to absorb one-sided flow at the rate it was arriving.

What kind of problem is this?

When a system functions smoothly under normal load but degrades sharply when demand spikes — not because resources have been exhausted but because absorptive capacity has been overwhelmed — that is a specific shape of problem. It is about the rate at which capacity can be delivered, not how much exists.

And the discipline that has spent a century studying how congestion emerges when variable demand meets finite capacity is queuing theory.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Queuing theory was built for telephone exchanges, motorways, and computer networks — systems with well-defined servers and orderly arrivals. A limit-order book is messier: a network of price-time priority queues with strategic participants who adapt in real time. But the structural vocabulary — for congestion, utilisation, and non-linear degradation — transfers.

The Mathematics of Waiting

Queuing theory began with a practical problem. In 1909, Agner Krarup Erlang, a Danish mathematician at the Copenhagen Telephone Company, needed to figure out how many circuits were required to handle the city’s call traffic.

Erlang’s insight was that average demand alone could not answer this. What mattered was the relationship between the arrival rate and the service rate — and crucially, the variability of both. Even if average capacity exceeded average demand, random clustering of arrivals could overwhelm the system in bursts.

This is the most counterintuitive thing about queuing theory. If ten calls arrive evenly spaced across ten minutes, no queue forms. If the same ten arrive in a burst during the first two minutes, the system is overwhelmed — even though average demand is identical. It is the clustering, not the volume, that creates the queue.

The simplest queuing model — the canonical M/M/1 queue, meaning random arrivals, random service times, one server — produces a result that is elementary and profound. Define the utilisation ratio, ρ, as the arrival rate divided by the service rate. When ρ is low, the system runs smoothly. As ρ approaches 1, average time in the system does not increase linearly. It bends upward sharply, approaching infinity.

At 50% utilisation, the system is comfortable. At 80%, time in the system has more than doubled. At 90%, it doubles again. At 95%, again. The relationship is convex — the last few percentage points of capacity matter enormously more than the first few. The result is robust across a wide family of single-server queuing models, though the exact curve depends on architecture.

The core intuition — that congestion grows convexly as utilisation rises — applies wherever variable demand competes for finite capacity: motorways, emergency departments, computer networks, and financial markets. The exact curve depends on architecture, but the non-linearity is general.

Where the Queue Forms

Now consider a financial market. Liquidity-demanding order flow, such as market orders and aggressive limit orders, arrives at some rate. The market’s ability to absorb it near prevailing prices depends on the displayed depth of the book and how quickly providers replenish it after trades occur. In a telephone exchange, the cost of congestion is time — you wait on hold. In a limit-order book, impatient flow converts waiting time into price impact.

In normal conditions, the service rate comfortably exceeds the arrival rate, and the system feels liquid. But arrival rates are not constant. They cluster. A macroeconomic release, a large rebalance, a sudden shift in sentiment — any of these can spike the arrival rate.

The problem comes when the spike in arrival rate coincides with a drop in service rate. And this is precisely what tends to happen, because the two are coupled. When order flow surges, market makers face adverse selection risk — the incoming flow may be informed. The rational response is to widen spreads or pull quotes, reducing absorptive capacity at the exact moment demand for it is highest.

This is the queuing insight applied to markets: they do not just experience higher arrival rates in stress. They experience higher arrival rates and lower service rates simultaneously.

Kyle’s 1985 model formalised why order flow moves price: market makers infer information from aggregate flow, so each unit of imbalance carries a price impact cost. That is not a queuing service rate, but it explains why the price of immediacy rises when flow clusters. Cont, Stoikov, and Talreja then modelled a stylised limit-order book as a queuing system, making the connection between microstructure and queuing theory direct.

Bouchaud, Farmer, and Lillo documented persistent dynamics in order flow — long memory in order signs, clustering in trade sizes, fluctuating depth — that make a queuing interpretation natural. Microstructure already has a name for this: resiliency, the speed at which depth replenishes and prices stabilise after a shock. The queuing framework makes that concept precise.

The Steep Part of the Curve

The queuing framework explains something that the pool metaphor alone cannot: why liquidity does not drain gradually. If liquidity were only a pool, you would expect it to deplete smoothly. Instead, what markets repeatedly demonstrate is an abrupt, non-linear shift: liquid conditions one moment, gridlock the next, with very little in between.

The October 2014 Treasury event showed this. So did the equity market on August 24, 2015, when impaired price discovery and reduced displayed depth produced severe dislocations in some exchange-traded funds — nearly a fifth fell 20% or more, far exceeding contemporaneous moves in their underlying baskets.

In each case, the system was not out of liquidity in the pool sense. What was missing was depth and replenishment. The queue had entered the steep part of its curve.

This connects to the ergodicity problem from The Path The Maths Misses. The ensemble average — “on average, markets are liquid” — may be true and yet irrelevant to the investor who needs to execute during the minutes when they are not. Edge, as we defined it in The Question Shapes The Answer, is conditional information — but information you cannot act on is not edge at all. The queue is the bottleneck between knowing and doing.

It also connects to The Network You Cannot See. In a network of queues, congestion propagates. When one venue hits its capacity limit, order flow diverts to related markets, potentially overwhelming their absorptive capacity in turn. The network determines where congestion starts; the queuing dynamics determine how fast it escalates.

Slow In, Fast Out

There is a deeper asymmetry. In most markets, the agents providing absorptive capacity — market makers, dealers — have few binding obligations to do so. Some venues have designated market makers, but those duties are typically weak and waivable under stress. In practice, liquidity is provided when profitable and withdrawn when not. The service rate is endogenous.

In Erlang’s telephone exchange, the circuits did not unplug themselves when call volume spiked. In financial markets, the service rate is a strategic variable. When conditions deteriorate, the rational response for a liquidity provider is to reduce capacity — exactly the response that pushes the utilisation ratio toward the critical zone.

This creates a structural asymmetry. Liquidity accumulates gradually as market makers compete for flow, narrowing spreads and deepening books over weeks and months. But it can withdraw in seconds as providers simultaneously pull back. The buildup is slow; the collapse is fast.

The same asymmetry appears wherever service capacity is voluntarily provided. In The Room That Runs Out, we saw how carrying capacity constrains populations. The queuing lens adds a mechanism: the room shrinks precisely when the crowd grows.

The Capacity Question

Queuing theory does not predict when a liquidity event will occur. What it does is add a question. The pool question — “how much liquidity is there?” — is necessary but not sufficient. The queuing question asks: what is the relationship between the demand for liquidity and the capacity to supply it, and how does that relationship change under stress?

The framework asks specific things. How close is the demand for liquidity provision to the supply of it? What would cause the arrival rate to spike — concentrated positioning, correlated rebalancing flows? What would cause the service rate to drop — adverse selection, balance sheet constraints, concentration of provision in a few firms? And critically: are the two likely to move in the same direction under stress?

These questions apply beyond financial markets. In healthcare, emergency departments operate on a version of the same curve — operational research suggests that hospitals above roughly 85% bed occupancy begin losing surge capacity, and performance degrades sharply as occupancy climbs toward the mid-nineties. In organisations, decision-making bottlenecks follow the same logic: an approvals process that works at moderate load can seize when demand clusters.

The discipline is in asking: not how much capacity exists but can it be delivered at the rate you need it and when you need it most?

References & Further Reading

1. The Theory of Probabilities and Telephone Conversations: Agner Krarup Erlang

2. Continuous Auctions and Insider Trading: Albert S. Kyle

3. A Stochastic Model for Order Book Dynamics: Ramy Cont, Sasha Stoikov & Rishi Talreja

4. How Markets Slowly Digest Changes in Supply and Demand: Jean-Philippe Bouchaud, J. Doyne Farmer & Fabrizio Lillo

5. The Joint Staff Report: The U.S. Treasury Market on October 15, 2014: US Treasury, Federal Reserve, SEC, CFTC

6. Market Liquidity: Theory, Evidence, and Policy: Thierry Foucault, Marco Pagano & Ailsa Röell

In the mid-1980s, a small number of trading firms discovered that the price of S&P 500 futures frequently diverged from the value of the underlying basket of stocks. The gap could be wide — sometimes several index points — and the trade was mechanical: buy the cheap side, sell the expensive side, wait for convergence. The maths was straightforward, the convergence was guaranteed at expiry, and the early practitioners earned extraordinary returns for what was, in principle, a riskless arbitrage.

Within a few years, the niche had transformed. Dozens of firms built the infrastructure to execute the same trade — direct exchange connections, faster execution systems, dedicated capital. By the early 1990s, the spreads had compressed from several index points to fractions of a basis point. The trade still existed. The convergence was still guaranteed. But the returns had collapsed to levels that could barely cover the cost of the technology required to capture them.

The index arbitrageurs had not been defeated by a change in market structure or a shift in the rules. The futures still diverged from the basket; the basket still converged at settlement. What had changed was the population. Too many firms were chasing the same spread, and the spread could not sustain them all.

What kind of problem is this?

The core opportunity has not disappeared. The spread still exists. The problem is that too many participants are competing in the same territory, and the ecosystem cannot support them all. When returns compress not because the opportunity has moved but because the capital chasing it has multiplied, that is a population dynamics problem. And the discipline that has spent decades studying what happens when populations exceed their resource base is ecology.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Ecology studies organisms that do not read about themselves or change strategy in response to being studied. Markets are faster, noisier, and more reflexive. But the structural vocabulary — for competition, saturation, and differentiation — is worth importing.

The Hypervolume

A niche, in ecology, is not just a place. It is a set of conditions and resources that allow a species to persist. The ecologist G. Evelyn Hutchinson formalised this in 1957 as the “n-dimensional hypervolume” — the full set of environmental conditions within which a species can survive and reproduce. The fundamental niche is the range of conditions a species could exploit in the absence of competitors. The realised niche is the smaller range it actually occupies once competition and other constraints are accounted for.

Carrying capacity — typically denoted K — is the maximum population size that an environment can sustain given available resources. It is not a fixed ceiling; K shifts as conditions change. But the logistic growth model captures the core dynamic: growth is fastest when numbers are low relative to resources, and decelerates as the population approaches K. The resource does not disappear. It simply cannot support unlimited extraction.

Competitive exclusion — Gause’s principle — states that two species competing for exactly the same niche cannot coexist indefinitely. The resolution is typically extinction or niche partitioning: selective pressures drive the species to specialise on slightly different resources, reducing direct competition. Spatial and temporal variation can sustain coexistence even with significant overlap, but where stabilising differences are weak, exclusion is the limiting case.

What Alpha Depletes

Consider the niche first. An investment strategy occupies a niche defined by the conditions that make it profitable: a particular type of mispricing, a specific market structure, a set of behavioural regularities that generate exploitable patterns. Merger arbitrage occupies the niche of announced-deal spreads. Statistical arbitrage occupies the niche of short-term mean reversion in correlated securities. Distressed debt occupies the niche created by forced sellers and information asymmetry in bankruptcy.

Each niche has a carrying capacity — not a fixed number, but a moving threshold set by the amount of capital it can absorb before returns degrade. In ecology, the relevant population measure is biomass; in finance, the equivalent is deployed capital. The threshold depends on the depth and frequency of the opportunities, the transaction costs, and the degree to which capital inflows themselves alter the dynamics. But it is finite.

This is where the ecology lens adds something. “Alpha decay” is conventionally framed as a property of the strategy — as though strategies age and weaken like radioactive isotopes. But the ecological view suggests it is not primarily about the strategy. It is about the relationship between the strategy and the population exploiting it. A strategy does not decay in isolation. It decays when the niche fills up.

The information-theoretic framework from The Question Shapes the Answer reinforces this. Edge is relational — it exists between an observer and a market, not as an intrinsic property of the observer. The ecological lens adds a further dimension: edge is also population-dependent. A signal that is highly informative when you are the only one trading it becomes noise when a thousand funds do the same thing.

The logistic growth model maps directly. When a new strategy is discovered, early entrants earn high returns. Capital flows in. Each new entrant captures a share of the finite return pool. Returns compress. Some funds close. Others persist with diminished returns. The opportunity still exists — but the population has reached the level the resource can sustain.

Unlike mineral deposits, mispricings are a renewable resource — continuously regenerated by hedgers, index rebalancers, and liquidity-seeking flows. But carrying capacity is set by the rate of renewal, not the existence of the resource. When extraction by arbitrageurs exceeds the rate at which the market regenerates the mispricing, returns compress even though the opportunity never disappears.

Surviving Past Zero

Competitive exclusion applies with a twist. In biology, species competing for identical resources cannot coexist — one will eventually dominate. In finance, funds competing for identical alpha cannot all earn excess returns — but they can all survive at zero net alpha if subsidised by management fees. The niche can remain overcrowded long after carrying capacity has been exceeded, because the population is decoupled from the resource it nominally exploits. Fees, not returns, sustain it.

This has implications for how we evaluate strategies. The question is not simply “does this strategy have alpha?” but “what is the carrying capacity of this niche, and how much of it is already occupied?” A strategy can be logically sound, historically validated, and currently unprofitable — not because it is wrong, but because the niche is full.

Why Good Strategies Stop Working described how strategies decay through environmental mismatch — the ice melting beneath the polar bear. The ecology lens identifies a second mechanism: the niche filling up with polar bears. Both produce declining returns but demand different responses. Environmental mismatch calls for adaptation. Niche saturation calls for differentiation. Confusing the two leads to the wrong intervention.

Beaks and Seeds

Competitive exclusion sounds terminal, but ecology offers a resolution. When two species face exclusion from the same niche, they can survive by partitioning it — specialising to exploit a subset of the resource that the other does not. Darwin’s finches are the canonical example. Thirteen species coexist on the Galapagos Islands, all descended from a common ancestor, by evolving beaks suited to different food sources. One cracks hard seeds. Another probes for insects. A third feeds on cactus flowers. They occupy the same archipelago but different ecological niches. The resource is shared; the means of extraction are distinct.

Niche partitioning is the mechanism by which ecosystems sustain diversity. Without it, competitive exclusion would drive every crowded habitat toward a single dominant species. With it, multiple species coexist by carving the resource space into territories narrow enough that direct competition is reduced. The narrower the specialisation, the more species the ecosystem can support.

In investing, this explains why “the same strategy” can have very different outcomes for different practitioners. Two quantitative equity funds may both call themselves “momentum” investors, but one trades weekly signals in large-cap equities while the other trades monthly signals in emerging markets. They occupy different niches despite sharing a label. Successful long-lived investment firms tend to be those that have found niches with structural barriers to entry — capacity constraints that limit competition, informational advantages that are costly to replicate, or time horizons that most capital cannot tolerate. The ergodicity problem from The Path the Maths Misses creates one such barrier: strategies that require tolerance for deep drawdowns and long recovery periods are inaccessible to capital with short evaluation horizons, even if the long-run returns are attractive. The inability to survive the path is itself a carrying-capacity constraint on the competition.

In a world of rapid information dissemination and freely available data, easily replicable strategies reach carrying capacity faster than ever. Strategies that require patient capital, deep expertise, or tolerance for illiquidity face less competition precisely because the barriers to entry are higher.

The Ecosystem Question

When returns are compressing and the debate is “does this strategy still work?”, the answer may be “yes, but the niche is full.”

The framework asks specific questions. What niche does this strategy occupy? What is the carrying capacity — how much capital can it absorb before returns degrade? How occupied is the niche currently? And is there a path to partitioning — a related but distinct opportunity that the crowd has not reached?

The ecology lens does not tell you when to enter or exit a strategy. It does not predict which niches will fill or when new ones will open. What it does is provide a structural vocabulary for familiar phenomena. “Alpha decay” becomes carrying capacity. “Crowding” becomes competitive exclusion in progress. “Finding a new edge” becomes niche partitioning.

The pattern extends well beyond investing. In organisational strategy, it applies to business models that once occupied white space and now face dozens of competitors extracting the same value. In scientific research, new subfields follow the same arc: early papers have high impact, grant funding flows in, more researchers enter the field, and the marginal contribution per paper declines even as the total output grows. Wherever success attracts imitation, and imitation compresses the reward for the original insight, the carrying capacity question applies. The room runs out in every domain.

Understanding the ecology of your position does not make you immune to competition. You are not the ecologist observing the ecosystem from outside. You are one of the organisms in it, subject to the same dynamics of crowding and exclusion that the framework describes. The finches that understood carrying capacity would still have to eat.

The discipline is in asking: where do I fit in this ecosystem?

References & Further Reading

1. Concluding Remarks (The Niche Concept): G. Evelyn Hutchinson

2. The Struggle for Existence: G. F. Gause

3. Adaptive Markets: Financial Evolution at the Speed of Thought: Andrew Lo

4. Frontiers of Finance: Evolution and Efficient Markets: Farmer & Lo

5. Mutual Fund Flows and Performance in Rational Markets: Berk & Green

6. The Rate of Return on Everything, 1870-2015: Jorda, Knoll, Kuvshinov, Schularick & Taylor

An investment committee reviews one hundred ideas a year and approves twenty. Of those twenty, twelve deliver returns. The committee reports a 60% hit rate. Not bad.

But is it good? You cannot answer that question with the information given. Suppose the pipeline contained eighty genuinely attractive ideas and twenty poor ones. The committee approved twelve that delivered and eight that did not — but also rejected sixty-eight that would have delivered. Its 60% hit rate conceals a significant miss rate. Alternatively, suppose only fifteen of the hundred ideas were genuinely attractive, and the committee found twelve of them. Now the same 60% hit rate reflects considerable skill. In practice, rejected ideas vanish — you rarely learn whether they would have succeeded. But shadow portfolios and post-hoc tracking of passed opportunities suggest the problem is real, even when the exact numbers are unknowable.

The hit rate isn’t a measure of skill. It’s a tangled composite of two completely different things: how well the committee can distinguish good ideas from bad ones, and how willing it is to say yes. These aren’t the same thing.

What kind of problem is this?

When you are trying to detect something real against a background of noise, and your detection system is imperfect, that’s a classification problem — not a forecasting problem or an optimisation problem. Is this thing I’m looking at real, or is it noise? And the discipline that has formalised the tradeoffs, the errors, and the hidden role of base rates is signal detection theory.

So let’s borrow.

This is the multi-model move: recognise the shape of a problem, find the discipline that has thought rigorously about that shape, and import its frameworks deliberately rather than reinventing them from scratch.

Signal detection theory (SDT) emerged from a collaboration between psychophysics and electrical engineering in the 1950s, codified in Green and Swets’ foundational 1966 text. The original context was literal: radar operators in the Second World War, staring at screens, trying to decide whether each blip was an enemy aircraft or atmospheric noise. Real signals and random noise produced overlapping patterns, and no amount of training could eliminate the ambiguity.

Sensitivity and Criterion

SDT formalises this by imagining two overlapping probability distributions — one for noise alone, one for signal-plus-noise. The overlap is the entire problem. The framework’s central insight is that detection performance decomposes into two independent dimensions.

The first is sensitivity — how far apart the two distributions are. In the technical language, this is d-prime (d’), the standardised distance between the means of the noise and signal distributions. High sensitivity means the world looks genuinely different when a signal is present versus when it isn’t. Low sensitivity means the two states are nearly indistinguishable. Sensitivity is a property of the detection system. You improve it by getting better information, better models, better analytical tools.

The second is the criterion — the threshold at which the observer decides to say “signal.” This is a choice, not a measurement. Slide the criterion leftward (become more liberal) and you catch more genuine signals but also flag more noise. Slide it rightward (become more conservative) and you reduce false alarms but miss more genuine signals. At any given level of sensitivity, this tradeoff is inescapable. The only way to get more hits without more false alarms is to improve sensitivity itself.

Every detection decision produces one of four outcomes: a hit (signal present, correctly identified), a miss (signal present, overlooked), a false alarm (noise mistaken for signal), or a correct rejection (noise correctly dismissed). Note that SDT’s “hit rate” — the percentage of genuine signals you caught — is not the same as the usage from our opening, where “hit rate” meant the percentage of approvals that worked out. These are different numbers, and confusing them is itself a diagnostic error. Plot the true hit rate against the false alarm rate as you sweep the criterion, and you trace the ROC curve — the Receiver Operating Characteristic. A system with high sensitivity bows toward the upper left; one with zero sensitivity produces a diagonal line. The area under the curve is a threshold-invariant measure of discrimination ability, untangled from any particular criterion setting.

This decomposition is SDT’s deepest contribution. In most domains, performance is evaluated as a single number — accuracy, hit rate, the percentage of investments that worked. But a single number conflates two things that should be assessed separately. A conservative investor who approves few ideas and gets a high percentage right might be cautious, not skilled. A liberal investor who approves many and gets a lower percentage right might have genuine sensitivity combined with a willingness to act. You can’t distinguish these cases without pulling the two dimensions apart.

The Base Rate Trap

Suppose genuine investment opportunities — ones that would generate meaningful risk-adjusted returns — constitute 5% of the ideas crossing an investment committee’s desk. The committee has good detection: when a genuinely good idea appears, they recognise it 80% of the time, and they correctly reject 90% of bad ideas. Strong numbers. Skilled committee.

Now do the arithmetic. Out of 1,000 ideas, 50 are genuinely good. The committee identifies 40 of them. Of the 950 bad ideas, it incorrectly approves 95. Total approvals: 135, of which 40 are good. The positive predictive value — the probability that an approved idea is actually good — is roughly 30%.

A 30% success rate from a highly skilled committee. Not because the committee is bad, but because the base rate is low. When genuine signals are rare, even excellent detectors generate mostly false positives. This isn’t an argument against trying. It’s an argument for understanding what the numbers actually mean.

The Payoff Matrix

In The Question Shapes The Answer, we explored Adami’s framework for information: edge requires specifying a target, an observer, and a baseline. SDT adds a complementary layer. Even when you have genuine conditional information — even when your sensitivity is real — the decision to act on that information involves a separate, strategic choice about where to set your threshold. And that choice should depend on the payoff structure and the base rate of genuine opportunity, not on the signal itself.

Consider the asymmetry. For a long-term investor with a diversified portfolio, the cost structure might look like this: missing a genuinely great investment is expensive because opportunity cost compounds over decades, while an underperforming investment drags returns but doesn’t threaten long-term viability. This cost structure argues for a liberal criterion — approve more, accept that many will underperform, because the cost of missing strong investments exceeds the cost of including weaker ones.

For a concentrated, leveraged fund, the cost structure inverts. A false alarm — a position that looked like a signal but was noise — can be catastrophic. The ergodicity problem from The Path The Maths Misses applies: a single bad position can destroy the ability to compound. This argues for a conservative criterion — demand compelling evidence before acting, accept that you will miss many genuine opportunities, because survival dominates.

Both are rational. Neither is “better.” They reflect different payoff matrices applied to the same underlying detection problem. SDT suggests that framing high-conviction and diversified approaches as inherently superior or inferior is a category error. The question isn’t how much conviction to have. It’s what the costs of your errors are, and whether your criterion is calibrated to those costs.

The Invisible Threshold

This connects to something the evolutionary lens from Why Good Strategies Stop Working revealed: investment strategies carry assumptions about their environment. SDT sharpens this. Every organisation carries criterion — a collective threshold for what counts as a sufficient signal to act. It is often implicit, embedded in how many approvals the committee gives, how much evidence is required, what the burden of proof feels like.

Some organisations are miss-averse and set liberal criteria. Others are false-alarm-averse and set conservative criteria. These aren’t personality quirks. They are strategic positions in the sensitivity-criterion space. But they are often invisible — embedded in norms rather than articulated as choices.

The only way to improve both simultaneously is to improve sensitivity — to actually get better at distinguishing signal from noise. If you decompose performance into sensitivity and criterion, you can ask the right question: is our problem that we are poorly calibrated (criterion in the wrong place), or that we genuinely cannot tell signal from noise (low sensitivity)? These require completely different interventions. Recalibrating a criterion is a decision. Improving sensitivity is a capability-building exercise — better data, better models, better analytical infrastructure, the kind of investment in measurement apparatus that The Question Shapes The Answer described.

Post-mortems often confuse the two. When an investment fails, the typical response is to raise the bar — demand more evidence, add another approval layer. This is a criterion shift. If the problem was low sensitivity — if the system genuinely could not distinguish signal from noise — then raising the bar simply means you will miss more signals while still getting fooled by noise that exceeds your new threshold. You have made yourself more cautious without making yourself more skilled. The ROC curve has not moved; you have merely walked along it.

The Detection Question

The discipline SDT offers is in the decomposition. Before asking “should we approve more or fewer ideas,” ask: can we actually tell the difference? Before tightening standards after a loss, ask: was the failure a criterion problem or a sensitivity problem? Before celebrating a high hit rate, ask: what was the base rate, and how many signals did we miss?

In markets, this decomposition interacts with competition. When everyone is trying to detect the same signals, the base rate of genuinely private information drops — more observers means more information gets priced faster. The more efficient the collective detection system, the lower the base rate for any individual. Improving aggregate sensitivity reduces each participant’s base rate, which means even skilled detectors will find that most of their positive identifications are false alarms. Understanding signal detection theory doesn’t make you immune to its dynamics. You will still face ambiguous signals, still set your criterion imperfectly, still be surprised by how often confident calls turn out to be noise.

The decomposition applies wherever imperfect detectors face ambiguous signals. In medicine, a screening test with excellent sensitivity can still produce mostly false positives when the disease is rare — and the response of ordering more tests is a criterion shift, not a sensitivity improvement. In security, tightening protocols after an incident can flood the system with false alarms while doing nothing to improve the ability to detect genuine threats. Any domain where someone is trying to sort signal from noise — and where the cost of errors is asymmetric — is a domain where SDT’s decomposition reveals something that a single accuracy number conceals.

But the framework changes the questions you ask. Not “was I right?” but “can I actually distinguish signal from noise in this domain?” Not “how many decisions worked?” but “what was the base rate, and does my sensitivity exceed chance?”

The discipline is in asking: is this a problem of calibration or a problem of capability?

References & Further Reading

1. Signal Detection Theory and Psychophysics: David M. Green and John A. Swets

2. The Signal and the Noise: The Art and Science of Prediction: Nate Silver

3. Noise: A Flaw in Human Judgment: Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein

4. Thinking, Fast and Slow: Daniel Kahneman

5. The Base Rate Book: Integrating the Past to Better Anticipate the Future: Michael J. Mauboussin

6. What is Information?: Christoph Adami

In 1847, Ignaz Semmelweis noticed that women in the Vienna maternity ward staffed by doctors died of puerperal fever at several times the rate of those attended by midwives. He could have counted deaths per ward and stopped. Instead, he asked what specific practice distinguished the two groups. Doctors performed autopsies before delivering babies; midwives did not. When Semmelweis introduced handwashing with chlorinated lime, mortality in the doctors’ ward fell to under two per cent. The measurement was chosen to discriminate between hypotheses, not merely to accumulate data.

This is the problem the forecasting engine must solve at every cycle. After hypotheses have been committed — the subject of the previous post — the system must choose what to measure. That choice is upstream of every conclusion it reaches.

Not all data is evidence

In stable domains the question is less acute. A physicist can repeat her measurements until precision converges; the regularity does not shift between experiments. In unstable domains, the distribution may be changing as she measures it, and the budget of observations is limited. Every data point consumed on an uninformative observable is one not spent on an observable that could have moved the posterior.

Christoph Adami’s definition of information, which we examined in Information, Causation and Persistence, explains why this is not a mere efficiency concern but a conceptual one. Information is the reduction of uncertainty about a target variable. But which variables you attend to, and which dependencies you estimate, depends crucially on the hypotheses you hold. There is no neutral measurement grid. The 2007 analysts from the previous post — balance-sheet, reflexivity, network — were uncertain about different things, and the data that would have resolved their disagreements was different data. House price levels mattered to all three. But the derivative of origination volume discriminated sharply between the balance-sheet thesis (cash flows about to meet obligations) and the reflexivity thesis (credit loop decelerating). A measurement strategy that treated all available macro data as equally relevant would have drowned the discriminating signal in confirming noise.

The system does not gather evidence to confirm what it already suspects. It gathers evidence to separate hypotheses that disagree.

What would settle it

The engine approximates this through a discriminability ranking — a heuristic inspired by what experimental design calls expected information gain. Before any search begins, a deterministic planning module reads all committed microstates across all genotypes — the competing analytical perspectives from the previous post — and asks: for each observable in the emergent ontology, how many hypothesis pairs would be separated by learning its value?

Consider a simple case. Suppose a balance-sheet microstate predicts that consumer credit delinquencies will rise, while a reflexivity microstate predicts they will fall (because self-reinforcing optimism is sustaining repayment). The delinquency rate separates these two hypotheses — observing it moves the posterior regardless of which direction the data points. Now suppose both microstates agree that the ISM manufacturing PMI will soften. The PMI reading, however it lands, has much less power to discriminate between them.

The planning module ranks observables by their discriminating power across the full microstate set, prioritising those that separate hypotheses across genotypes rather than merely within a single perspective. An observable that distinguishes what balance-sheet economics predicts from what reflexivity theory predicts addresses the deeper uncertainty — which decomposition of the economy applies — rather than merely adjudicating between variants within one lens. This is where the system’s value over a single-perspective analyst is sharpest: it knows what it is uncertain about at the structural level, and it measures accordingly.

This is where the emergent ontology from the previous post earns its keep. Because the measurement space was constructed from the hypotheses themselves, the system knows exactly which observables separate which pairs. A system that had started with a fixed data catalogue would have no principled basis for prioritising among indicators. The emergent ontology makes the prioritisation computational rather than editorial.

The engine then enforces a depth-first protocol: for the top-priority observables, gather at least two independent, high-quality evidence items before moving to the next. This resists the natural tendency of language-model search agents toward breadth-first sweeps — one weak datum per observable across the entire ontology, which feels productive but leaves every hypothesis equally underdetermined. The system trades coverage for conviction.

Independent witnesses

Even with the right observables and sufficient depth, a subtler problem remains. If a hypothesis only makes predictions about financial data, and all the evidence gathered is financial data, then the hypothesis is being tested against its own kind. Confirmation within a single measurement system is not the same as confirmation across independent ones.

The engine classifies every observable into one of five evidence interfaces: accounting data (balance sheets, credit spreads), physical activity (payrolls, housing starts), belief indicators (surveys, sentiment), institutional signals (policy rates, lending standards), and epistemic markers (nowcasts, forecast dispersion). These represent largely independent measurement systems — different instruments pointed at overlapping but distinct aspects of the economy, generated by different processes and subject to different errors.

Lateral falsification is the principle that a theory should be tested against evidence from outside its native interface. A balance-sheet hypothesis that predicts rising delinquencies (accounting) should also predict something about hiring (physical) and consumer confidence (belief). If it makes predictions only within one interface, any systematic bias in that measurement system propagates unchecked. The engine flags hypotheses whose predictions cluster within a single interface and demands cross-interface exposure. This connects directly to test environment expansion from the early posts — the five interfaces function as different test environments for the same structural claim, and lateral falsification is test environment expansion applied to evidence rather than to the belief itself.

Earning the right to update

Once the engine has selected discriminating observables and diversified them across interfaces, it still needs a conservative rule for turning observations into posterior movement. Evidence arrives as qualitative assessments, not raw numbers — grounded in what Built to Predict Not Know and Sounding Fluent Misses What is True established about where current language models tend to be more reliable (semantic interpretation) and where they tend to be less so (calibrated numerical reasoning). The system asks for a judgement per microstate: confirm, contradict, or mixed, at what strength. These scores then cross the firewall into deterministic code. A fixed lookup table maps each to an approximate Bayes factor — strong confirmation yields 2.5, strong contradiction 0.4, mixed evidence 1.0. Every individual factor is hard-clamped: no single evidence item can shift the posterior by more than threefold. Conviction must be earned cumulatively rather than seized from a single compelling narrative.

But naive accumulation has its own failure mode. Many observables that appear independent are driven by shared latent causes — retail sales and credit card delinquencies both reflect household balance-sheet health. Multiplying their Bayes factors double-counts the information. The system uses the structural commitments themselves to correct for this: if two observables share a latent parent in the committed causal graph, the engine tempers their combined evidence, raising each factor to the power of one over the cluster size. Across independent clusters, factors compound at full strength.

This has a consequence that connects back to the diversity architecture. Different genotypes impose different dependence structures on the same evidence. The same consumer credit datum carries different weight under the balance-sheet hypothesis and the reflexivity hypothesis — because each posits different latent parents. The evidence is the same; the structural context is not.

After all evidence is scored and tempered, the cumulative factor for any single microstate in one cycle is itself clamped to a range of one-tenth to ten. Conviction accumulates across cycles, not within them.

What only live data can answer

The measurement architecture converts theoretical principles into working constraints. Adami’s insight becomes a computational ranking. Test environment expansion becomes lateral falsification. The configuration problem manifests in every bound and protocol that determines how aggressively the system updates.

Whether these constraints are well-calibrated is not something architecture can answer. The conservative bounds prevent overconfidence but may prevent the system from moving fast enough when regime shifts produce unambiguous signals. The depth-first protocol builds conviction on key observables but may miss regimes that announce themselves through diffuse signals across many indicators simultaneously. Contact with live data will tell. Semmelweis counted the right deaths. The engine measures the right disagreements. The next post examines how the system composes a forecast from competing, weighted hypotheses, what its uncertainty decomposition reveals, and what the first hindcast results showed when the architecture met historical reality.