The debate over the Fed's guidance appears to be quieting down, as the Chair was more… forward at Jackson Hole.

The more the reason to finish my trilogy of posts on Fed guidance.1 In the inconceivable eventuality that they have gone unread:

There is some math in the post, but quite a bit of effort went into making it as accessible as possible, and you can get by knowing what the union and intersection of sets are and how a function works. The last subsection typically ups the ante mathematically speaking: Just remember to like the post before you skip it!

Executive Summary

• Policy messages are risky. Since we all knew that, it counts as evidence, right?

• Outlook messages are safe when introducing scenarios, less so when refuting them. Leaning inclusive is safer, but biased.

• And, only slightly hyperbolically, it effectively takes full knowledge of the policy function to process any form of guidance in any state without corrupting existing information.

The “audience” is at the center of today’s model snippet, but we need to start with…

The Fed

So today, the policy authority (the “Fed”) is basically an automaton. When the time comes, it will map a scenario s out of a large but finite universe of them, S, into a policy decision p out of the smaller and finite universe of them, P. Yes, the policy function is

and the most important point is that f is not injective. This means: There may be more than one scenario that leads to the same policy decision, f(s)=f(s'). More on this important assumption, see my previous post.

I said “when the time comes,” and it has not come yet. With the decision ahead, now is the time to provide guidance on the decision.

What is “guidance”? I need more structure, but warning!, I am trying hard to add as little as I can.2 But some, I do have to add. Let’s start with what the Fed knows:

• A subset of scenarios S*={s_1,s_2,...} that it deems “admissible,” which we can think of as the support of a probability distribution or as the subset of scenarios that exceed some threshold probability;

• The subset of policy actions P*={p_a,p_b,...} that is under consideration. The subset is consistent with f,

Throughout the exercise, the Fed is always telling the truth, and everybody knows it.

Next, what am I letting the Fed say, and what form does the message take? I will think of atomistic statements about either the admissible set of scenarios S* or the consideration set for policy P*. We will reduce the “atoms” to two: exclusion and inclusion.

I should make it immediately clear that these two atoms do not span messages like s,f(s): If the Fed chains inclusive3 messages s and p, it is up to the private sector to infer that p=f(s). Now that it seems obvious that they would do so, but what if the Fed's inclusive message reads:

Unless the Fed also provides a key, it is just your brain that wants to say

despite my conscious use of different alphabets for each index. Without the key, there is still information about f, because we do insist that the Fed is consistent, i.e.,

but it is rather immediate that the message would be consistent with another policy function, namely, g(s_1)=p_b, g(s_2)=p_a. Sorry to anyone with OCD.

In short, what we are addressing today is guidance on the outlook and the decision, not a direct communication about the policy function. Emphasis on direct: Messages about the outlook will carry information about the policy function, but not as much as if we sat the FOMC for nationally televised tabletop exercises.

Not a Gotcha

I should also avoid any chance that this post turns into a “gotcha.” Inclusion and exclusion are not symmetric operations, despite what they may look like when later on I just write +s,-s. My favorite Karl had something to say about this:

“[…] the asymmetry between verifiability and falsifiability … results from the logical form of universal statements […]”4

The Logic of Scientific Discovery, Karl Popper

(We do speak the language of Lawvere, though. So perhaps it is worth spelling it out:

Inclusion is union. Exclusion is the intersection with the subtrahend’s complement. The latter is why exclusion is often said to be “silently universal.”)

Below, I will have a “naive” listener who will make mistakes because it will treat +,- a bit too literally. It is a helpful construct and, arguably, a robustness dimension we may care about. What is not, but kind of looks like, is someone from the Fed (me) telling readers (you) that they are naive because they have not properly studied Chapter 5 in Rogers’ Theory of Recursive Functions (1967).5

Finally, the Audience

Congratulations, you have made it to the “switch.” Like, in “bait and switch.”

I do not have a model of learning (yet?). Instead, I will ask: What does the audience need to know about the policy function to learn from the guidance? The answer is radically different depending on the message, which leads me to a conjecture, as well as a warning: You cannot expect people to learn from your message if the content already presumes the learning, the same way you cannot have an introductory course on linear algebra that lists familiarity with Priestley spaces as a prerequisite.

Roc, you should listen to yourself sometimes.

Until then: The audience carries its own admissible set of scenarios, S_A, and consideration set for policies, P_A. They would like those to track the Fed’s S*,P* as close as possible. When they receive an atom message from

what the audience needs to do, and what knowledge of f already presumes, to make sure that their update sets S'_A,P'_A are closer to the Fed’s S*,P*? Apologies for the several notation crimes committed, but I am struggling to keep this post short. And we still have to clear the naive agent.

The Naive Agent

Whether the naive agent is an exposition device or a robustness result, I do not know yet. Can it be both, for now? Thanks.

The naive agent wants his or her admissible and consideration sets to track the Fed’s, but he or she is kind of lazy about it.6 If faced with an inclusion message, it just adds the element to the set—and if it was already there, well, deleting duplicates is not too hard. If faced with an exclusion message, it checks whether the element is already in the tracking set; if so, it deletes it and goes to play golf; otherwise, it goes to play golf.

He gets away with it when faced with an inclusion message, but things go very poorly with exclusion. Say the Fed flashed that scenario s is not in the admissible set S*. The naive agent finds it was in his tracking set, proceeds to remove it:

Wait, before teeing off, the naive agent realizes that it needs to adjust his consideration set too,

which requires to know f(s) (not the entire function, just the value at s), but more importantly, it is wrong wrong wrong. Actually, it can be worse than doing nothing, because it can take P_N away from P*. How?

Recall f is not injective. It is thus possible that there exists another scenario x such that f(x)=f(p), and that scenario x does belong to the Fed’s P*. So the naive agent could have started “spot on,” S_N=S*, P_N=P*, and a true message has corrupted P_N.

Let’s use the table from before to document the worst-case scenario for the naive agent, facing each kind of atomistic message:

Note that something as simple as “no, we will not cut next meeting” may end up rather badly. This will become clearer in the next subsection.7

Guess who is always very reluctant to rule out policies, particularly without hedging? Yep, that’s right, the policymakers.

Agent 86

Change of gears: Rather than telling you what bad things happen when the naive agent spends too much time at the golf course, I will tell you what it takes to not be naive. Measured in what you need to know about f, the policy function.

As before, the agent carries a tracking admissible set S_A and a consideration set P_A, faces a message m and, this time, he will make sure that he is done with work before golfing.

Let’s see what it takes when faced with an inclusion policy message p. Obviously, it has to be added to P_A. That is the easy part. But whether p was or wasn’t in P_A, agent 86 needs to adjust S_A, and that demands pulling back from p and finding all, and I mean all, the scenarios with the property f(s)=p. And you need to know the entire policy function to do so.

Now let’s walk through the other messages:

• Policy exclusion: If p was not in the agent consideration set, go golfing in peace. If it was in P_A, the agent does have to pull back some of f, but not the entirety of it: He just needs to remove all scenarios in S_A such that f(s)=p. Presumably, if the agent had a consistent pair S_A,P_A to start with, he had f over S_A already.

• Outlook inclusion: If s was not in the agent’s admissible set, all that is needed is one instance of f, namely, f(s). However, agent 86 did not use f(s) to construct P_A, so this is “new territory.”

• Outlook exclusion: Similar to policy exclusion, if the excluded scenario was not in S_A, it is a nice golfing afternoon. If it is, the agent will check whether it can adjust P_A by running f over S_A-s and seeing if there is any other scenario s' in S_A such that f(s')=f(s). If not, it can be safely removed.

Deploying the table:

There is a judgment call here with the colors I want to flag, because we will see where I land at the end of the day: I just counted how many points of f are needed. However, it is worth noting that scenario inclusion requires an instance of f not previously computed, while exclusion only requires the key to the initial admissible and consideration set.

What Have We Learned

Well, I do not know about you. I have learned that:

• Policy messages are high risk. Wait, I already knew that. As a matter of fact, I am sure all policymakers know that. Can this count as evidence then?

• Outlook inclusion is the safest. I note, though, that if the Fed (or the journal editor) leans too much on inclusion, we end up with 99-page papers.

• I need more structure on the cost of listening. Agent 86 can get away with a partial function on S_A in response to exclusion messages; it is still a requirement, but how costly is it to evaluate a known function? Compared with a scenario inclusion, in which a single, but previously unnecessary, evaluation of f is needed.

• Also, I learned that I should wait until I am done with the article before I advertise it as “short and simple.”

Irish Goodbye

As usual, if you know what I am doing, please tell me.

As widely reported, the Treasury has taken several measures to keep rates down on long-duration Treasury bonds. In particular, it announced “buybacks,” which, helpfully, are what they sound like: the Treasury buys back long-term Treasury bonds, pushes their prices up—and thus rates down.

The thing about buying, though, you need money. The same goes for the Treasury.

The Treasury first announced that it would fund the buybacks by issuing T-bills—and quite naturally, several called it an operation “twist,” removing duration from the market. Later, another possibility was floated: The Treasury could run down its checking account at the Fed, the Treasury General Account (TGA).

I am going to refrain from judging the Treasury’s actions1 and instead clarify the Fed's role if the buybacks are paid for by the TGA. Nick at WSJ wrote an article today that touches on this subject—and several others that I will not touch on.

Executive Summary

Bill Nelson is correct (also, the sky is blue).

It does not really make a difference whether the Treasury funds the buybacks with T-bills or the TGA. The latter route mechanically transforms the Treasury’s holdings into reserves, which were already at “ample” levels. Under the FOMC’s directive to “maintain ample reserves,” the Desk will let the extra reserves drain… which effectively means more T-bills will have to find their way to the private sector.

If you have time for “buts,” I have a small one and a big one that I discuss in some more detail below. The succints2:

1. The Desk responds to autonomous changes in the supply of reserves, regardless of their source, to maintain an ample level. As such, the additional T-bills will take some time to reach the market.

2. The accounting may be innocuous, but it is my belief that the Fed needs to be held to Caesar’s wife’s standards: above suspicion.

First, though, an attempt to explain how the Fed’s balance sheet gets involved when the Treasury buys back debt via the TGA.

Show me the Money

…flowing through the Fed’s balance sheet.

There is quite a bit to unpack here. The three boxes (Reserves, TGA, and Currency) are the three items of the Fed’s balance sheet that we will track.3 All three are liabilities and amount to a closed system: Funds, not-very-originally depicted as water4, flow between Reserves, TGA, and Currency without changing the Fed’s balance sheet size. Nothing escapes to the right.

The red line is helpfully labeled. The Desk job would be to keep the water above it.

Instead, the ins and outs of the Fed’s balance sheet are depicted on the left. RMP stands for Reserve Management Purchases, and that’s how the Fed’s balance sheet grows (in the current regime). However, let me illustrate why keeping reserves constant requires positive RMPs:

As long as the demand for currency grows, banks turn in their reserves for moola, creating a permanent drain on reserves.5 That is, reserves are becoming bank notes, so without adding water, the tub would dry.6

The Fed also gets a lot of receipts from its assets, other forms of income… these are debits to Reserves (technically, some are to the TGA, but we simplify). So “receipts” flow out of Reserves, and then the Fed decides what to do with them. Let’s add it to my balance sheet art.

Receipts, principal payments… drain reserves. When the Desk reinvests, it replenishes the reserves. When the Desk redeems, the reserves are gone, and the Fed’s balance sheet shrinks. It is how the Fed shrank on both occasions.

Currently, though, there is no shrinking: the Desk is reinvesting all the receipts. Let’s set the very stylized lever to its appropriate position:

We are already five figures in, and we have not started with the exercise. If you know what I am doing, please email me.

Without further ado, the Treasury wants to use TGA to buy back some long-term bonds. These bonds are, one way or another, held by a bank. So, the Treasury credits the bank—and adds reserves—while retiring the offending bond.

Too many reserves! Too many reserves! ALL HANDS ON DECK DESK!

Actually, there is no drama. The directive is clear to maintain ample reserves, but interest rate control, liquidity needs… are not at risk from too many reserves. If, totally hypothetically, the Desk were to be less than perfect, it is better to err on the side of too much rather than too little. Which is to say, there is no need for immediate intervention.

Instead, the Desk can just drain reserves by setting RMPs to zero and letting currency do its work. As the article below by a great economist documents, the Desk manages reserves monthly, at least for the time being.

So RMPs could be adjusted as soon as October. Problem: The Desk already adjusted RMPs to zero for September. Bill Nelson proposed that some receipts could be redeemed to accelerate the drainage. Well, we have the lever for that.

Coda: (1) Reduce RMPs to zero, (2) redeem some of the receipts. Either way, it is T-bills that the Fed is not buying, so someone else has to.

But #1

To be clear, I will repeat myself: The Desk is maintaining reserves at an ample level, per the FOMC instruction. Any movement in autonomous factors (TGA being one of them, currency another) changes the supply of reserves and triggers the Desk response. Whether it is the Treasury or, for example, a sudden drop in the demand for cash, it does not matter.

In particular, the Desk is not sterilizing TGA movements. That has been proposed by Annette Vissing-Jorgensen and has gotten some interest as a means of reducing the Fed’s balance sheet, though forthcoming work by yours truly suggests limited shrinkage.

Because it is not a direct response, there is a sizable period in which the Fed will be a touch larger than it needs to, while reserves drain. Let’s guess-estimate how long it would take for, say, an extra $100 billion flowing out of the TGA to drain. Once again, let me broadcast that the Fed tells you exactly what is in its balance sheet weekly, here.

Currency currently (had to) drains $70-80 billion a year, so the “extra” reserves disappear rather slowly. However, we should also expect the demand for reserves to grow—in nominal terms. A 4-5% growth rate seems a valid first guess, or about $120-150 billion with the present base. In other words, let the demand come to the supply: It does so rather quickly. Perhaps as fast as 6 months per $100 bn, without adjusting reinvestments.

Because…

But #2

When it comes to monetary policy, it does matter how things look.

If the Treasury wants to revisit its liquidity management and operate with less liquidity, using the resulting savings to retire long-term debt, there is no problem. The Fed is the fiscal agent of the Treasury, and in that capacity, it provides the financial services needed. And a more tightly managed TGA would actually reduce the level of ample reserves. All a matter of management: Boring. And boring is good.

But claims that appear to equate using the TGA to some form of forgotten windfall are misleading and dangerous. The “windfall” is very intuitive to anyone with a checking account who has decided that, well, now is the time to use some of that buffer.

If the Fed is directly responding to and sterilizing something successfully sold as a “windfall,” the spin writes itself and, frankly, sounds more convincing than a mortgage application.

But the utmost concern is the precedent, in my opinion. It would break the separation of the Fed as fiscal agent to the Treasury and the Fed as manager of reserves. Two different responsibilities and governance that happen to share a key base, reserves.

Again, this is my opinion and my opinion only: The Fed can either sterilize none or all of the TGA changes. Blankets, yes. It should not be sterilizing some TGA changes but not others. Discretion, no.

The debate on the Fed’s guidance rages on. My last post was a warning against reductionism: There is guidance that is, or can be, good, and there is guidance that is, or can be, bad.

Today’s post: Sometimes, there is simply no guidance to give because the Committee may have consensus on policy, but not on the outlook.

And without an outlook to stand on, policy guidance can easily be detrimental.

This post is in increasing order of complexity. The executive summary is only in words; then I repeat myself with a little bit of math that would be sufficient for those seeking to understand my claim, but bring a fair amount of head-scratching to those of you who already know exactly what I am talking about. Be patient if you can; the very last section briefly discusses why I believe the formalisms provide the scaffold to build a model.

Also, you do not need to read the full post to like, share, or subscribe. Soon I will decide whether sending emails and polluting inboxes is the way to go to publicize my posts.

Executive Summary

• Different circumstances can call for the same policy, ok, I lied, there is some math: 1 (policy instrument) < 2 (mandates). And an example: both high unemployment and low inflation typically call for a rate cut.

• Committee members can then agree on policy without sharing an outlook. In other words, there can be consensus on policy without consensus on the outlook.

• This is a challenge for good guidance. Without an outlook consistent with the decision, guidance—even selective guidance—becomes counterproductive.

With a Little Bit of Math

A bunch of policymakers (I>1 of them) get together to find a consensus on policy.

There is a large (but finite) universe of “scenarios” S and a smaller set of possible decisions, P. You will find that I leave all definitions as open-ended as I can.

There is a policy function that maps a scenario into a policy decision:

I am only going to say three things about the policy function, and the last one is a tangent:

• All policymakers share the same policy function, unrealistically, but clarifying that the challenges to guidance arise without disparate policy views.

• The policy function is not injective, realistically.

• In a macro model, we model policy with a Taylor rule, or discretion, or you tell me. Prescription and prediction coincide. When it comes to communications, that is akin to assuming the can opener.

I model each policymaker i as entertaining a collection of scenarios and, consistently with the policy function f, the set of policies she or he will consider:

What is consensus in such a bare-bones framework? It seems natural to start by looking at the set of policy decisions all agents are considering:

Note that P* can be empty: No policy decision would get unanimity. If you find this too strict for consensus, ponder that we only need I>1, so perhaps this group of policymakers we are talking about is only a fraction of all the policy participants.

The same exercise with the scenarios gives us

and we have arrived at…

The crux of the problem

Despite all policymakers sharing the policy function, despite each and every policymaker considering the policies consistent with his or her scenarios, the consensus is not consistent with the policy function,

To see how, it is just better to proceed by example. Let’s reduce the next FOMC meeting to up, down, and no change:

Also, this is a Committee of two, I=2, with Violet and Oran. The universe of scenarios is surprisingly similar to the alphabet, S = {a,b,c,...}.1

Let’s see how things shake up. The two policymakers’ consideration sets are color-coded, and the policy function is tidily arranged in rows as well.

There is consensus on no rate change… but not on the reasons behind it. The outlook-policy pair is nothing short of traumatic:

You tell me how to prepare for the press conference with that.

Before I comment longer on what this means for communication, I do want to throw you another example where there are consensus scenarios and consensus policies… but they are not consistent with the policy function.

The problem is that one of the consensus policies, the hike, has a shared scenario in the policymakers’ minds. Both policymakers are also considering no change in the rate, but based on different scenarios. The result is a communication nightmare again:

What would you infer from a Committee that insists on keeping no change as a possibility, but can only provide a scenario for a rate hike?

More broadly, it is unlikely that the consensus pair is fully consistent with the policy function—and inconsistent guidance seems pretty close to “bad.”

Implications for Guidance

To fully derive the implications for guidance, I will need a model for whoever is listening. Working on it!

Until then, I like to think of consensus as a constraint on consistent communications:

• If the Committee is fully committed to consensus, guidance cannot be guaranteed to always be consistent.

• If the Committee only provides guidance when the latter is consistent, then silence will be informative—a signal of dispersed outlooks.

• Policy guidance without an outlook seems the worst of all worlds, leaving the door wide open to misinterpret the guidance as a commitment and providing no information about the underlying policy function.

But I also like to think about how we can relax such constraints:

• If the Committee were a computer, there would be a straightforward way to reach consensus (in the next subsection), but so far I do not see how one would do so in an actual meeting.

  • Moreover, it is far from straightforward if there are different policy functions in the Committee.

• More naturally, perhaps scenarios should be aggregated before consensus is taken, trading off precision for consistency.

  • On the works: Given a set of policymakers’ consideration sets, characterize the finest partition of the scenario universe that renders the consensus pair consistent.

  • In other words, how much detail can be provided while being consistent?

Why so Formal?

It is likely that you felt the formalism was overkill. Actually, most likely, if you felt like that, you have already stopped reading. I have also surely tested your patience by talking about scenarios, outlook, and guidance… without proper definitions.

There are three reasons for the formalism. First, and completely irrelevant, because that’s how I verify my own ideas. Second, because it sets us up for lattice and even category theory to come to our rescue: the policy function maps scenarios to policy, but what we need has to reliably go back and forth between sets of scenarios and sets of policies. Let 2^S be the power set of S and behold the mapping

which we actually built properly inside every policymaker’s head. The problem is, in fancy words, that f_* is only a left-adjoint, which preserves unions,

but it does NOT preserve intersections,

Since I promised, the “computer solution” is to pull back f_*

and then construct a mapping that is closed under intersection,

Hopefully, it is self-evident why I am skeptical that a group of imperfect, very busy, in-a-hurry humans can deliver an adjoint mapping.

Third, and goodbye: this is a scaffold on which to mount further structure. For example, does consensus preserve priorities? What about arguments?

The new Chair has a stated dislike for guidance, and two press conferences in, we are starting to see, in practice, what he does like.

A lively debate has followed, and here I am, aiming to reserve 2% of the internet for constructive, reasoned, and demonstrated discussions. And to this end, today’s point is simply that guidance cannot be reduced to be good or bad, for there is good guidance, and there is bad guidance.

Two Ears, One Mouth

I don't know why, but I had to start it somewhere. So it started… there.1

So I will take a clear case against guidance that is a variation of the old adage to listen twice as much as you talk.

The argument is that, by providing information to the markets, the Fed commands the markets’ attention, which is to say, the Fed crowds out other matters. Market signals then have more of the Fed’s own content and less of everything else; ergo, the Fed gets in its own way when attempting to learn from “everything else.”

Hence, less guidance will improve the Fed’s readings from markets, which in turn will improve monetary policy.

Unfortunately, my assessment is the two words every policymaker hates to hear:

“It depends.”

I have done this long enough that I can fill in the policymaker’s (exasperated) response,

“OK, what does it depend on, Roc?”

It depends on the form of guidance. There is no way around it. The argument above is clearly true if all that guidance does is enthrall markets, distracting them from other important matters.

The problem with being reductionist is that the opposite sounds about as valid. After all, if all guidance does is to align market expectations with the Fed’s, then yeah, the more guidance, the better.

Now I was ready to tell you that the truth is somewhere in between, BUT I realized that the above assumes that aligning market and Fed expectations (on the Fed's terms) is unambiguously good.

Which I strongly believe, and, abusing your patience, let me tell you why.

No Exploit

Too often, when discussing policy, we must be careful to distinguish “what it is” from “what it ought to be.”

Well, I am very happy to say that we do not have to make that distinction here: The FOMC neither does nor ought to conduct policy by means of surprising the public. Even if it can.

I associate Alan Blinder’s critique of time-consistent models most closely with this observation-plus-principle. In his view, central banks have learned to deal with the inflationary bias, in a host of ways—for the purposes here, though, it is communications that Alan believes create the credibility.

“In the real world, such credibility [...] is painstakingly built up by a history of matching deeds to words.”

Obvious? Well, even more obvious: to match deeds to words, you need... words.

We will get back to deeds and words, but let’s close this section as clearly as possible: I feel very confident that the FOMC seeks neither to engineer nor to exploit surprises.

The debate on guidance is NOT about generating surprises to exploit.

Attention!

I have worked with models of rational inattention for much longer than people realize.2 So take my word that there are theoretical and very deep reasons why the original Claude said that “information is the resolution of uncertainty,” and thus Chris Sims3 disciplined the agents’ attention by introducing a cost of acquiring information, a.k.a., resolving uncertainty.

We are going to try our luck getting the attention point across and tying it back to good/bad guidance with a simple exercise rather than complex equations.

Convince yourself you MUST know what vegetable the following message is about:

@RT$CH%K#

You can solve it; hopefully, it is not too hard. But it may take a bit of effort for you to work out the noise I introduced artificially, which arises naturally from a finite communication channel. For all you know, in the time you spent cracking the veggie of my choice, we have invaded Belgium.

Now, if I give you the key, A = @, E = #, I = $, O = %, it will take basically no effort to figure it out: ARTICHOKE. Also, here is a video of Salvador Dalí waxing lyrical about artichokes, in Catalan, so you can waste the 23 seconds I estimate you saved.

The point is that adding noise to a payoff-relevant object demands more attention, but providing a key to unscramble a message saves you attention.

I believe there is good guidance that does the latter, and bad guidance that only accomplishes the former.

And this is why I find “guidance retreat” to be more accurate than invoking Alan Greenspan.4 It is a retreat from a defeat, namely, the Fed’s inability to provide good guidance.

To which I reply, as long as we know what we are doing, we do have the ability to provide good guidance. Don’t get me wrong, there are indeed limits to how much good guidance we can offer, but I also believe that by far the FOMC has stayed well within these limits—and stretched them only when it felt it was necessary.

Let’s descend into the argument by means of the decomposition by Marek Jarocinski and Peter Karadi in this excellent paper (JK henceforth). As a bonus, we will get a peek at the latest from the empirical side. Preview: The Fed seems to be doing an ok job communicating its reaction to news, with most policy “surprises” stemming from the market reading the economy differently from the Fed, not from the market reading the Fed differently from how the Fed sees itself.

No surprises

Let’s take the view that the Fed does not like to create surprises, and hopefully either my previous section or Alan Blinder himself has convinced you this is true.

Let’s follow JK and say the Fed follows a simple Taylor rule

where \vec{\alpha}_F is a vector of coefficients to apply to the vector of (state, trend, cyclical...) variables \vec{x}_F and \varepsilon is a classic monetary policy shock—orthogonal to anything and everything. This is certainly a very simple policy specification, but we need to start somewhere—and keep the promise of some empirical return. I also apologize that LaTeX does not compile unless in a dedicated window.

We have subscripts F for Fed, because next we are going to look at the market expectation of the Fed decision, say at 1:59 pm on a few Wednesdays, and posit that markets form their expectation as

where the variables mean what you think they mean, with M short for markets.

Let’s define surprise! as S = i - E_M(i). If we were in a FIRE model, which stands for Full Information Rational Expectations, we would be done, for S=0. Oh, the shocks? That is right, we can have shocks, so let’s say

What happens if we are not in a FIRE model? About anything. So, let’s actually narrow that down a lot, so we can walk through the implications of

for policy surprises and the Fed’s ability to retrieve information from the surprises.

A bit of algebra you do not need a PhD for yields

where the last step is the decomposition JK pursue:

• The first term

  is the central bank information (CBI) channel, which dates at least to Romer and Romer (2000, AER), and got a boost by Melosi (2017, Restud) and Nakamura and Steinsson (2018, QJE). In short, the central bank has information that the markets don’t, \vec{x}_M \not = \vec{x}_F, with the key implication being that the Fed actions will have information content for the markets.

• Alas, there are more terms, and

  is the Fed response to the news (FRN) that Bauer and Swanson (2022) advocate. In short, news happens, and if the market is wrong about how the Fed would respond to the news, \vec{\alpha}_M \not = \vec{\alpha}_F, then surprise! If you do not control for this possibility, you will overestimate the CBI effect.

JK show that you can identify the channels separately by adding observables (like the stock market). And I will tell you what they find below. But, for now, pretend otherwise: The Fed knows itself (\vec{\alpha}_F, \vec{x}_F, \varepsilon), it is staring at S, and it would like to retrieve what it can about \vec{x}_M, \vec{\alpha}_M.

It is rather obvious that there is an identification problem, even if we are working with a scalar. After all, \alpha_M x_M is already two variables and one observation.

We do continue undeterred because we can still measure the information that you can retrieve. And two days after I typed that, I am still working it out—and slowly realizing that the policy rule we started with may work well with DSGE models and what not, but it does not live in the same space as Fed guidance.

Under reasonable conditions, the less uncertainty and bias there is about how the Fed responds,

the more information it can retrieve about \vec{x}_M, the market signal about current conditions or state of the economy.

Consider the CBI channel in isolation. Under the hopefully-mild assumption that the Fed knows what it is doing itself, it is trivial to recover \vec{x}_M if there is only CBI.5

What prevents inversion, beyond the irrecoverable monetary policy shock \varepsilon, is the FRN term. So, minimize it, set \vec{\alpha}_M = \vec{\alpha}_F if you can, and you are good! Now, questions: How do we do so? How would we even know if markets and the Fed are aligned if we can only identify the term FRN, assuming it is equal to zero?

Well, I have double good news. Marek and Peter figured it out, and the answer is, Fed communications got good grades: markets seem to get the Fed’s reaction to news right—and quickly.

And no alarms

I invite the readers to go to the paper to take a look at their identification strategy, but I will cut to the chase here and tell you simply what they find. In their words,

“Our analysis confirms the robust presence of the central bank information shocks, even when the FRN effect is carefully controlled for.”

Moreover, FRN effects appear to vanish very quickly. Projected on a monthly VAR, only the CBI channel has significant effects.

I do not think this is the final word on the matter, and the authors themselves acknowledge that some specifications still give the FRN channel wings. I am also not understanding why the second-order term,

belongs to the CBI channel rather than the FRN channel. You may hope that being second-order does not make a big difference, but we have no way of telling unless further structure is imposed.

So far, though, the evidence is pointing to the CBI channel (remember, the one the Fed can easily invert if it knows itself), with the FRN channel being short-lived and seemingly unbiased.6

We reach a rather awkward point, though. It would seem that the evidence points to the fact that the Fed and markets were aligned well enough to reduce surprises and allow the Fed to retrieve information. All this, while the Fed provides the guidance that it seemingly would not provide anymore.

Wrapping up, for now

This post is long enough, but I left so much that I wanted to touch on: global games, my favorite email game, the general case for a boring Fed, Richard Clarida’s two forms of data dependence, and existing taxonomies of forward guidance.

Maybe if August gives us a break, I will find the time.

The last 8 FOMC meetings have seen 16 dissents, including 4 in a single meeting (April 2026) for the first time since Tron was coming to a movie theater near you. Even controlling for serial dissenter former Governor Stephen Miran, the surge in disagreement is remarkable in the context of the last 40 years:

You can find the data source on this page of the St. Louis Fed, and be sure to check the work of (the late) Daniel Thornton and David Wheelock (FRB STL).

I recalled a famous quote about voting at the FOMC by former Governor Laurence Meyer:

[…] there are two imaginary red chairs around the table, the “dissent chairs.” The first two FOMC members who sat in those chairs were able to dissent. After that, no one else could follow the same course.

A Term at the Fed, page 53

Well, clearly someone brought more red chairs to the FOMC!

Still, I thought that it was worth a shot to give some consideration to Meyer’s hypothesis: Perhaps there is indeed a number of red chairs, and past that number, dissents are possible but extremely rare.

In other words, do FOMC voters become increasingly reluctant to dissent if they expect other voters (one, two, three more?) to dissent as well?

Executive Summary

No.

Why Not

Look, if I had to explain why yes, you would have to read it in a well-published article. But since I struck out, I will instead share some of the work here—there are some interesting bits.

There were two self-imposed constraints: first, no participant names, so all inference comes from the voting outcome; and, second, only data since 1987 is used.1

The unconditional probability of a voter dissenting is quite low at 5.4 %, and close to 60% of meetings had no dissent. The distribution of the number of dissenters is classic count data:

I find count data difficult to judge visually, but then I could not find a cow in Wisconsin. So we need a benchmark.

If dissents were iid across meetings and voters with probability p, we can easily compute the probability that there are no dissents in a given meeting with 12 participants,2

We thus have a gap of about 5%, which is quite sizable, but we need to zoom in a bit on meetings with at least one dissent to make a call.

This is not a good fit. Note that by choice of p, the average number of dissenters is matched exactly E(d)=np. There are too few single- and double-dissent meetings. There are too many triple-dissent meetings. This already strongly hints that there are at least three red chairs.

Now, another obvious observation from the first figure: Dissents cluster over time, likely because the matters at hand can be more or less contentious.3

Fortunately, humanity solved this problem a long time ago with the Beta-binomial model. In short, we will let the dissent probability p vary from meeting to meeting, reflecting that sometimes the FOMC meeting is just a bore, other times the center of the known universe (most of the time, both).

Ok, now this is what I call a good fit:

The fit is excellent.4 And the Beta-binomial knows nothing about chairs. This is the strongest signal that there are at least three chairs.

Unfortunately, this is as far as we can go, for four dissents are simply a very rare occurrence, even if it happened only months ago. In our sample, there have been only 3 meetings with 4 dissents: the fact that none had 5 dissents is not informative given that dissents are relatively rare. For what it's worth, the Beta binomial puts the probability of 5 dissents or more at about 1 in 1,000 meetings.

So, as far as data can tell us, red chairs never end at the FOMC.

But all this talk of red chairs gave me an idea…

2

Vacant governor seats reduce the voter count. Quite common

We Must Act Now

This is the open letter signed by many excellent economists.

And this is the rant of the week (on this topic).

As Grumpy says, we are all for understanding—and there are actions we can take now that go beyond “more research is needed.” Improve transparency, for example. Or build armored sandboxes.

The issue is the second clause: “build the incentives, guardrails, and institutions needed to steer AI”. The sentence continues, by the way, giving a direction to steer to: “complements humans and benefits society.”

Call me conservative if you want, but I believe we need to know a lot to steer a technology “to benefit society.” For there is a high risk that steps taken with incomplete knowledge today do reveal themselves to be part of the problem later on. The law of unintended consequences always hangs over any policy; one more reason to avoid any uninformed consequences.

Pushing back on a message like “we must act now” is not easy (and they know it). Is the suggestion here that we do nothing until it is too late? Of course, not.

It is about what we must act on now, and definitely steering is not in (yet).

A Great Example of What Not to Do, brought to you by Bernie Sanders

You can count on Bernie to provide target practice. His “act now” proposal is spelled out: “give the public a direct ownership stake in the largest AI companies in America through a one-time 50% tax on the stock of the biggest AI companies.” That stake will go into a sovereign wealth fund, to be run by an “independent” commission,1 and there’s no need to speculate on what the commission will do; Bernie tells it: “The Independent Commission would use voting shares in these companies to block decisions that hurt the American people and to push for policies that help them.”

Thankfully, The Argument took good care of 90% of the rant the proposal deserves.

Let’s hammer in the remaining 10%, for I have a much better, if not very original, solution: tax income. Let me simplify things for myself and actually imagine it as a corporate profit tax, and collect all the goodies:

• You do not have to figure out which AI, hardware, IT companies…. (I mean, where do you stop) would be in the portfolio, the “chosen ones.”

• You do not give the chosen ones an unfair advantage over firms that dared to have their big idea after Bernie had his.

• You avoid biasing future public decisions that affect the chosen ones’ bottom line.

• You do not reincarnate the East India Company.

I cannot really say that you would avoid distorting private decisions—a tax would typically do that, too—though it is my opinion that, over time, distortions pile up when a company is nationalized beyond those of a properly designed tax system.

Trying to get some numbers for some napkin math, I noticed that Bernie, not always the most self-aware individual, volunteers that taking over half the AI industry will result in… $1,000 per person per year.2 Clearly, all problems of the world, solved.

Do note that I did not have to invoke the “slippery slope” argument, but it is there as well. If you nationalize AI, we shall next expect Bernie to nationalize quantum computing, cold fusion, and USB sticks that slot in at the first try.

What to Do Now: Measurement

There are many, many people, economists and otherwise, working on measuring the impact of AI. This is clearly a data-rich environment that can be harvested, starting yesterday. Of course, I am not going to be unbiased in my pair of examples:

• This FEDS Note reviews publicly available indicators to track the impact of AI buildout—which can often, inadvertently, boost the measured impact of AI adoption.

• This Chicago Fed Insights piece goes to Brazil to find a unique dataset to explore the effects on office and production workers, separately.

In case the home bias is not clear enough, I also want to highlight this paper by Leonard Nakamura and coauthors for the BEA, seeking to understand and properly measure “free” content (ads, marketing,…) in GDP. They find clear trend breaks in the growth of content value around 1995 and 2022—which they attribute to the advent of the internet and AI.

Well, there is more to my choice than home bias. No, I do not have anything to contribute to their paper—my knowledge of the topic, next to Leonard, is approximately none.

Here is the thing: AI is likely to change how we have to measure… about everything. In my opinion, AI increases the value of information—actually, allow me to be more precise while speculating: it sustains the marginal value of information far from the decreasing returns that about any human experiences, I say, when reading Winston Churchill’s The Second World War around Volume 3.3

This is bound to improve but is also likely to reshape how the private sector collects and processes information, and which facets reveal themselves to the public through disclosure or actions. An example may be better than to rework that poorly stated definition:

• Consider price setting. A firm continuously processes information and chooses when to update prices. The more information it processes, the more it can update prices to its profit. This, in turn, makes it more likely such a firm decides to invest in technologies that reduce “menu costs,” perhaps all the way to dynamic pricing.

  • And this is not measurement neutral. Did you know that the price of sushi drops by 50% every Monday and then rises by 100% on Tuesday? So ideally we want to compare Monday sushi prices with Monday sushi prices. Actually, make that Tuesday sushi.

  • The primary source for aggregate data is the actions (and data) from the private sector. Note that here we would not think of pre-empting anything: Statistical agencies must wait for changes in the private sector to adjust their measurements.

Perhaps I should formalize this someday.

What to Do Now: Transparency.

More transparency is always better, right? And I just said that AI will allow for the processing of more information at a reduced cost. I do not need to argue any further.

Only, I do, for there are some steep trade-offs when it comes to security. Consider the OpenAI model hacking an open-source repository (oh, the irony), breaking out from its secure sandbox (apparently, not for the first time—everything you want to know and more, in this excellent if very long post by the Zvi).

Note there is little downside to sharing what failed in the sandbox environment.4 It is somewhat shocking that OpenAI, a leading firm, could not build a secure sandbox—though we may be facing a corporate rather than a technological failure. In any case, there is a strong imperative for learning as an industry when it comes to safety—and I am going with aviation, not banking, as an example, for we are in the FOMC blackout period, so no closed-door meetings with bankers discussing anything Fed.

Contrast with sharing information on how HuggingFace was hacked.5 Now we are sharing an exploit, so we want the information to stay secure, which means we need to know who will know, and if I keep adding commas, I will certainly succeed in convincing you that things got complicated—to the point, it can be done, but like the Ark, it needs to be done before it rains.

So, if we are talking about what to act on now, I may suggest that setting up the “rails” to share information securely in the industry should be quite high on the list. Consider such a simple question: what is a bank airline, what certifications and regulations would be needed to qualify for accessing the information, who owns the information system, etc…

What to Do Now: Think, Listen, Share

It should have been obvious that I would end up arguing that what we should do now is also… what I am doing now.6 That is, think, listen/read, think some more, and if you have something you believe is worth it, share it. What is the worst that can happen, that someone is still reading this sentence? And then, what, ask for their money back?

I can tell you the best it can happen: One of the best mathematicians alive is the one doing the thinking, and then he happily shares it. Here is a quote that I like, and hopefully shows that it is accessible material to you all, knowing none of my readers is Terence Tao.7

A genuine "artificial general intelligence" is within reach of current tools, but argues a weaker and still very valuable "artificial general cleverness" is becoming real: the ability to solve broad classes of problems by ad hoc, often brute-force or stochastic means that are fallible and uninterpretable, yet succeed at non-trivial rates when coupled with strong verification. For humans, cleverness and intelligence are correlated; for machines they are largely decoupled, and today's tools are best viewed as stochastic generators of sometimes-clever, often-useful outputs — which yields the characteristic "useful yet unsatisfying" feeling of a magic trick once explained.8

Yes, there is some fatigue, as I look like the king in The Little Prince, calling people to write about AI—as if they need any encouragement. So, perhaps, I should say, do not be discouraged.

Correction: Each of us appears to have a different opinion on anything and everything.

(Somewhere a lone economist steps up and shouts “no, we don’t!”).

There is much to celebrate about such diversity, but it often obfuscates the progress that the profession is able to make, the work that lifts all boats, or rocks them—or perhaps just helps them to go further, perhaps just warns them not to go further.

So I do want to highlight and disseminate those papers that bring progress—mind you, not the final word, but that advance the profession. For example, by using new technologies to ask questions that many thought were settled in banking, or by bravely tackling a new margin in labor markets.

Today’s paper is itself an excellent example of how economics can help us make sense of the rapid change AI is bringing upon us. I will summarize what the paper tells us, though frankly their title says it all: “Vibe coding kills open source.”1

I then elaborate on what I like about this paper’s approach—and compare it, with my usual flair, to other papers that seem to run on imagination, though one cannot fail to notice how often the envisioned AI future just happens to line up with a 10-year-old publication by the authors.

I was going to close with a commentary on Stanford’s AI center open letter, but a Grumpy(ier?) economist kind of nailed it down already.

The Paper

Miklós Koren (CEU),2 Gábor Békés (CEU), Julian Hinz (Bielefeld University and IfW Kiel), and Aaron Lohmann (Bielefeld University and IfW Kiel) take on the effects of vibe coding on the open source “ecosystem.” The paper can be accessed here, and Miklos was recently on a Practical AI podcast and in an FT article, talking about it.

Their starting point is “an actual and factual AI use,” my words, not theirs: Vibe coding. No need to speculate; it is here. Coders of all stripes (professional, hobbyists, accidental) can just ask what they want, and an AI agent will be taking care of it. It has brought an increase in productivity and, I may add, a lot of fun; it does have a bit of a learning curve that can bite you late—and leaves behind a lot of “slop.”

Miklos and coauthors zoom in, though, on its immediate impact on the software ecosystem—open source software (OSS), to be specific, that thrives on user-producer interaction. Hold on, let me be more accurate: it thrives in the interaction between the user and producer communities, with the added wrinkle that most producers double as users—but not the other way around.

Guess what, that user is now interacting with an AI agent, and that agent is not replicating the interaction with the producer, breaking the feedback and reward loop, and is certainly not reproducing the communities and their dynamics and network effects. If the latter are important enough, as one may think given the success of OSS, and knowing that markets do not necessarily deliver the first-best outcome in the presence of externalities, well, the paper asks, are we better off?

And before I tell you their answer, let me highlight one thought that I know (at least two of) the authors share: AI is not just a general-purpose technology; it will not just rearrange technology stacks or make that kind of skill more or less valuable. It will reshape how we organize ourselves—the boundaries of firms, the industrial organization, conventions and equilibria, and ultimately, our expectations, our institutions.

So perhaps you are disappointed that the paper will not tell us if AI will save or slave us (or both). Myself, I am excited that such a strong team takes on AI’s immediate implications on how we produce, distribute, and consume a substantial amount of software.3

And if you need dollars to be motivated, feel free to check how much (and on what) Google is investing in the realization that AI will change how we search and thus their cash cow.

The Economic Consequences of Vibing

Of course, the authors are my kind of economists, namely, those who rather rephrase a Buggles’ song than a Keynes’ book. In any case, the title makes it clear that the authors take a stand.

On net and in the aggregate, vibe coding is counter-productive: “[…] greater adoption of vibe coding lowers entry and sharing, reduces the availability and quality of OSS, and reduces welfare despite higher productivity.”

Is this the case? I am not 100% sure, but after reading this paper, I am inclined to call it “more likely than not.” But I would do a massive disservice to you if I boiled down the paper to a likelihood assessment.

Instead, it is worth checking the key ingredients in the recipe, assessing them, and coming back to the conclusion—I learned from the round trip. (If you want the recipe in full, you will have to read the paper.)

• The OSS ecosystem: The paper declares its focus on the “traditional” OSS ecosystem, characterized by “individual maintainers and small teams [who] sustain projects through engagement-dependent channels: reputation, consulting leads, documentation traffic, and discovery.” In the authors’ view, “[traditional] OSS is monetized only through direct user engagement.”

  • Such a segment exists and is substantial, no question. However, as acknowledged by the authors, the corporate sector is also an important player in OSS, driven by “strategic or complementary” considerations—which AI could amplify, as a matter of fact. The question is, are we sure that the “traditional” OSS does not have such considerations? Are the segments fundamentally different in the value they produce, or is only how value is captured that varies?

• Software-begets-software. There is no question that software is a massive intermediate input in the production of software. There is no question that more software options lead to better, cheaper, more productive software.

  • However, software brings two additional dimensions: dependencies and maintenance. Anyone who has tried to run old R code and found that it needs to update a package required by a package required by a package, only to learn that the latest package now calls a different package anyway… well, anyone like that would know what I am talking about.

  • It is impossible to fully model all the complexity that is the input-output structure in software. The question is whether the omission of dependencies and maintenance is relevant. And I wonder if it is. Mostly because, frankly, not even the most ardent OSS fan likes to maintain packages—but AI can truly contribute positively here.

  • I am not sure that I can touch on “dependencies” without opening a 10,000-word rant on how we pile layers and constructs upon legacy software, and I better stop here before I drop the remaining 9,985 words.

• Heterogeneity and entry. Developing software involves a fixed cost. Not all software is the same.4 Thusly, there is a threshold for development that, in turn, determines both the availability and quality of OSS. In the model, the distribution of “software quality” has a long tail, and as a result, any change in the quality threshold maps to the average quality.

  • I have little to add here. I would note that the underlying “love of variety” is not as clear as it is when it comes to consumer products—I certainly enjoy having nearby lunch spots serving American or Chinese fare, and choose exactly five minutes before leaving the office. Not quite the same as choosing between an American or Chinese AI model, is it?

Imagination Economics

Don’t get me wrong, I am a theorist, and I love “what if”s. And I think that there is a bit of underappreciation for theory, currently. As I like to bring up way too often, the first digital computer was built not here at UPenn, but in Alan Turing’s head.

And there is certainly value to scoping future AI scenarios, ideally leveraging all we know about how technologies have changed how we work, consume, and live—and, yes, that includes recycling your 15-year-old task model.5

Now, I wish that we stopped here. Because what we should not do is pull back from a particular, possible but uncertain, AI scenario… and wax lyrical about AI policy, now.

The prime example has become labor markets. Pretty much any major technological change calls for the reallocation of resources, including workers. Well, make it painful for the workers to be reallocated—and, congratulations, you have reached “Stage 1: Something needs to be done.”

Whether exclusively in the context of the model, or with an eye toward the current debates, economists cannot resist “Stage 2: This is something.” And I would still say, well, it is informative to know how things could be improved in a given scenario—we can then collect scenarios and see if there is anything in common, and think of robust AI policies…

…but please do not jump to “Stage 3: We must do something.” The next stage is the hardest, sorry, and you do not get to rush it: We must do the right thing. And we are not in a position to do so right now, period.

The present paper avoids this trap by focusing on the “actual and factual.” Fair enough, some imagination is still needed—and the authors do discuss the normative implications.

Some may then call the paper narrow. I would reply that, sorry, it is indeed only on a narrow slice of all that AI shall bring where we are in a position to deploy economics, to the service of policy and the public.

Oh, the irony

You surely have read about how OpenAI model snuck out of the bedroom window to go hack HuggingFace. Leaving aside the obvious cyber-security concerns…

…there is no shortage of irony in the fact that the most advanced AI model decided that OSS was indeed the shortest route to solve the test.

I wonder from whom the AI model learned that…

A Grumpy Quote To Close

The letter then demands “Economists, policymakers and technology leaders must act now to understand the economics of transformative AI.” Well, I’m all for understanding, but economists don’t seem to need much prodding, given the flood of work on AI. Which is getting about as far as we usually do in speculation about the unknown. Economists can barely agree on minimum wages and taxes. Climate estimates are all over the map. “Subsidize our research” isn’t a potent call to arms.

The Grumpy Economist, AI regulation

In January 2019, the FOMC introduced “ample” into the lexicon of central banking in a rather succinct manner:

“[…] a regime in which an ample supply of reserves ensures that control over the level of the federal funds rate and other short-term interest rates is exercised primarily through the setting of the Federal Reserve’s administered rates, and in which active management of the supply of reserves is not required.”

Statement Regarding Monetary Policy Implementation and Balance Sheet Normalization

Uncountable efforts to narrow the definition of “ample” followed: Speeches by FOMC participants and Desk officers. Reduced-form and structural models. Memos and memos. Bank surveys. Most of my hair. Hours of FOMC discussions. Conferences, workshops, discussions. More surveys, speeches, and memos. But less hair, definitely.

Time waits for no one, and reserves inexorably approached “ample” before we could conclude anything other than more research is needed. After another (really) brief instance of the Harker-Armenter plan, the FOMC instructed the Desk to “maintain ample reserves” in December last year.

Ergo, by the axiom of revealed preferences, the ample reserves were $2,973,767 million on December 11, 2025.1 We could divide by nominal GDP, round a bit, and make Governor Waller look good (again). Or I could have given you another number, say 42 or 2 percent, to hint that “ample” is not just a dollar amount, or an interval, or a ratio to a nominal base.

Instead, I shall argue that “ample reserves” are ultimately given by the “management of the supply of reserves” that is required.

Ergo, by the axiom of revealed preferences, ample reserves are the level of reserves that can safely be managed with monthly purchases.

Reserve Management Purchases

Let me bring you up to speed as quickly as I can.

To maintain ample reserves, the Desk is conducting Reserve Management Purchases (RMPs) to “accommodate projected trend growth in the demand for Federal Reserve liabilities as well as seasonal fluctuations, such as those driven by tax payment dates.” Those purchases are “in the secondary market of Treasury bills” and will be announced “on or around the ninth business day of each month alongside a tentative schedule of purchase operations for the subsequent approximately thirty days.”2

How is it going? Well, let me refer that question to Roberto Perli, Manager of the System Open Market Account, who discussed RMPs in March, May, and July.3 Indeed, straight from his last speech, the time series of announced RMPs so far:

You can access all announcements and the current schedule here.

There is a lot to say about the dollar amounts, but let me direct your attention to the x-axis instead of the y-axis. These are months. Schedules are for a month and announced monthly. The RMP amount changes month to month. Heck, even the Desk speeches are monthly.4

Bottom line: Reserves are currently being managed monthly.

Rate control and reserves management

The more often you intervene, the less buffer you need. This is so intuitive that I can spare the reader a vehicular metaphor or a model.5 Rate control, a.k.a. policy implementation, is the mandate for the Desk—from the FOMC, and with the FOMC terms, to be clear.

If you adjust the path of reserve supply once a year, reserves must safely deliver the FOMC target for the next 365 days—if you adjust it every day, they need to do so only for the next 24 hours. Obviously, you get away with a smaller buffer, on average, if you intervene daily.

And daily is how often the Desk would work prior to 2008, under “scarce” reserves. Every day, there would be a forecast for that day, and a plan for the day. The standards for rate control were less stringent but, still, what the Desk did, day in and day out, was impressive.

Nowadays, though, the “im” word for daily management would be impossible.

Before I can tell you why, I do need to talk about “in” words.

Intervention and Information

Excuse me while I interrupt myself, but I need to clarify between planning and doing (intervention) and what is known and decided upon (information).

See, the Desk currently decides monthly on a weekly schedule of market operations. So every week sees an operation, but according to terms previously decided. It could operate daily, but, if the Desk decides on a diurnal schedule once a month, it is still managing monthly.

There are practical constraints on the frequency and size of the Desk's operations, too. How fast the Desk (or any desk, for that matter) can acquire $x dollars in Treasuries depends on the depth of the market—which in turn depends on, er, everything, and on an hour-by-hour basis (see, for example, current reality).

It is my view, though, that the information flow ultimately dictates the buffer the Desk needs to safely deliver rate control.

A matter of weeks

So, here lies the problem: A lot happens all the time in money markets. Shocks to the supply of reserves, shocks to the demand, and shocks to market functioning.

Now, shocks have their own frequency. But I need to narrow down what I mean by “shocks.” Some of the largest moves in are on specific dates, like quarter- and year-ends, which even economists can forecast with accuracy. There are seasonal effects, there are holidays, and tax days. There are events, like the extraordinary measures the Treasury undertakes when the debt ceiling is binding, that while their occurence and duration may be stochastic, they have predictable patterns.

But it is the unpredictable that the Desk must prepare for. And a lot of the unpredictability of money markets is a matter of weeks.

I can only give you an example at this time, and it may take some convincing to convince you that it is a relevant example. To wit, the Treasury maintains a balance (known as TGA) with the Fed that (1) adds/substracts reserves as it moves around; and (2) moves around a lot. Actually, I wrote a Substack about it. Actually, two… Three.

With a model of TGA balances, designed to disentangle the effect of emergency measures, and most certainly not exercise in overfitting,6 I computed the fraction of changes in the TGA balances the Desk can offset, assuming it knows everything there is to know, but with a lag of L weeks (information lag). I called it the sterilization floor.

And it is steep.

Basically, the figure above says that if you are three weeks behind the news, you have already missed on 80 percent of movements in the TGA balance. Actually, one week and you missed on a third—a proper month, and you are left smoothing 10 percent.

And everything you cannot smooth with interventions must be absorbed by the buffer. And you cannot smooth what you do not know, no matter how often you intervene.

It is true that the Desk has a lot more information than I do, every day. And you can always drive the floor to, er, the ground by overfitting—which can be worse than actually ignoring information, just saying.

It is also true that the TGA is one of many factors impacting the supply and demand of reserves, market functioning… Needless to say, the information needs of the full picture are larger than the tiny snapshot I gave.

Bottom line: If the Desk aimed to manage reserves at a daily or weekly basis, it would face very stringent information demands and would not be able to offer anything longer than, maybe, a weekly schedule a week ahead—and would still need a generous buffer to negotiate ample.

And since I am vibing, not working, and I need to start working, not vibing, I move to the last point.

So monthly it is?

Not so fast.

The Desk must get it right before it can get it tight. We are learning as we speak. I expect that in a year, or two, we will have a better command of the seasonal factors, of the shocks and their correlations, of what the Treasury wants to do, and we may even heard back from the task force.

Which is to say, a fair amount of the volatility will become predictable. Which, in turn, can improve the trade-off between frequency and buffer, as well as, perhaps, enable higher frequencies than weekly.

So, in the near future, the FOMC could keep the monthly management and reduce the existing buffer; or keep the existing buffer and make purchases a quarterly decision. While I have no doubt which option would be deemed preferable, I did want to include all the FOMC options to remind the reader that, for all you read and heard, at the end of the day, “ample” will be what the FOMC says it is.

Chris Moser (Columbia) and coauthors Andres Blanco (FRB Atlanta), Andres Drenik (UT Austin), and Emilio Zaratiegui (Bank of England) have been working on a macro model with bilaterally inefficient job separations, which is a fancy way to say that jobs sometimes are lost when they should not—not even according to the parties, employer and employee, ending the relationship.

Chris presented the paper in Miami Macro—yep, this is just paper number two of the three I promised to write about. I knew right away I wanted to write about it. For they are daring to take on questions—not the only ones, mind you—that expose both their flanks: On one side, tackling an obvious fact of the labor markets, whose absence in workhose models proves that macroeconomists are oblivious to those looking from the outside; and, on the other side, satisfying a host of theoretical and empirical expectations, whose absence in workhose models also proves that macroeconomists are oblivious, but this time to those looking from the inside.

I joked that people get internet at 30,000 feet and complain that it is slow.1 Perfect can be the enemy of good—and progress. And in my summary opinion, this paper brings truckloads of progress, even if in daring to open a can of worms, we cannot help but point to those worms getting away.

The Expectation of Efficiency

Where I come from, macro models started with the Real Business Cycle (RBC) paradigm, where everything is awesome: both for the economy, for the allocations stayed efficient in all instances, and for the economists, who had a clear pathway to solve the model.2

But make no mistake, we understood that was just the starting point. The choice of the word “paradigm” is a bit of a misnomer. It does act more like a yardstick: If you are bringing a deviation from efficiency in your model, you are expected to show why we must, why your friction or constraint is part of the “as possible” and not part of the “iota more.”3

In a form of dialectics that I do not expect any non-economist to enjoy, the back and forth between the expectation of efficiency and the reality of constraints and frictions has actually driven the models (and the profession) forward. Labor markets are a prime example of this journey, and the paper at hand, of the latest station in the direction of travel.

Let’s see how.

Search frictions

Search frictions are the most immediate deviation from full efficiency you will encounter in labor markets (and many others). It takes time (and focus, Roc, focus!) to find a job, so the immediate and unnegotiable allocation of workers to employers in the neoclassical model is simply not happening. Whether one sees search as a less-than-perfect matching technology, an underlying coordination failure, or as a physical environment that hardwires time as an input to tasks, we are now working with, at best, constrained-efficient models.

The harness is still quite tight. Directed search, for example, implements constrained efficiency; random matching typically does not. And to the point here: No matter how a pair of worker-firm find each other, once matched, and if given enough flexibility in labor contracts, it will be bilaterally efficient—and, in particular, Robert Barro noted that the match will continue or be terminated only when it is efficient for both parties to do so.

That is, all job separations are mutually agreed upon.4 And while many may not need to hear more before concluding that, well, we will need to go beyond search frictions then for a realistic model… do remember the iota dictum. Yes, it is the real world, but does it belong to the subset of features (for it must be a subset, as in its entirety it would just be a poor man’s copy of the reality that we seek to understand), does it belong to the “as simple as possible” that the model identifies as relevant to the questions at hand?

It does. For several reasons, but for the paper at hand here, because quits are not layoffs, and layoffs are not quits, and be very careful to assume otherwise. In particular, flows into unemployment are typically driven by workers entering the labor force anew, or looking for a new job after quitting.5 During recessions, though, the labor market is flooded by laid-off workers. Not surprisingly, this has implications for how quickly currently unemployed workers will find jobs, which in turn can shape aggregate dynamics and even the transmission of monetary policy.

Contract frictions

Well then, search frictions are not enough if you think sometimes, just sometimes, when a worker and an employer part ways, it is not as ‘mutually agreed’ as bilateral efficiency will tell you it must be.

But if money is left on the table, it must be that something prevents the parties from getting hold of it. And the natural candidate here is that the space of contracts at hand is also constrained (by what? An important question that will take us too far to tackle here and is NOT answered by the model choices made). Two caveats in the footnote.6

Our helpful and imaginary layman would be enough to point out what is missing: Hey guys, most wage contracts are 8 pages long these days, and you guys solve for a single rate per (efficiency) labor unit and call it a day?

Well, the unhelpful economist replies, would you like to solve your model?

Fortunately, there is such a thing as a helpful economist, too. And actually, four of those came together to our rescue in this paper. I do want to avoid discussing the modeling and solution details that make it possible, but I do need to highlight a point: That the less-than-complete contract space is mapped into wage rigidity, which is, ehem, old as F of x.7

The Crux of the Paper

It is indeed quite technical to prove, but I am taking the chance here that it is also quite intuitive: At any given time in the economy, there is a distribution of match surpluses, that is, for every worker-firm pair in place, we have the joint value of the relationship.

• When that value disappears, and again let me emphasize that it is the sum of the values from each party, we have an efficient separation: Better apart.

• If the value remains high, we are better together, even if the contract terms carried over from yesterday are no longer the best fit for today.

• But search+wage rigidity bring a third possibility: There may be surplus left, but the contract value is no longer above replacement for either worker or firm. Those cases will terminate (in a quit or a lay-off) and “should not have,” a.k.a., they are inefficient.

The key contribution of the paper is to bring the above complexity into a (relatively) simple picture, which I just snapped from their paper:

OK, please stay with me while I explain someone else’s graph in Substack. The distribution of match surpluses I mentioned? That is what you are staring at, with the vertical axis being just the probability. The horizontal axis is the first key result: the match surplus is a function of the real wage to productivity (the hat w that I struggle to type in Substack). The second key result is the two thresholds indicated by the two vertical lines, nicely labeled: If the real wage is too low given the worker’s productivity, the worker shall quit; if the real wage is too high, the worker shall be laid off by the firm.

Technically, it is a (s,S)-problem. Economically, a beautifully succinct result for a complex model. It comes at some cost (not every assumption necessary to get this result is equally palatable), and we have only bounded the problem: we still have to handle the distribution between thresholds.

Understanding why the distribution matters gets us one step closer to the title of the graph above. Visualize a shock as a movement in the distribution, or perhaps the threshold. Actually, let me burn some trees:

(Aside note: It is puzzling that Gemini Pro was able to produce the animation with a single prompt; the above is the first return. However, no amount of prodding could convince Gemini Pro to actually cut the first shift, which has gone wrong.)

I hope it is intuitive that the fraction of matches close to becoming quits or layoffs will determine how each margin responds to shocks. It follows that under a symmetric distribution, as depicted above, about the same number of workers are about to quit as firms are about to lay them off. In the paper, this translates to “zero employment effects of inflation,” but we will move on quickly from this, as the asymmetric case is what makes the model shine.

Still hoping a picture is worth a thousand words: If the distribution is skewed, as shown above, more workers are close to losing their jobs than to quitting them. Now, why would that be the case? The paper discusses “inflation greases the wheels of the labor market,” though skewness appears to dictate the qualitative nature of the relationships (that is, their sign). More generally, productivity or inflation drift will skew the surplus distribution and give distinct properties to the quit and layoff margins.

I do want to use the internet at 30,000 feet, though, so I refer the interested reader back to the paper for a full discussion, explicit formulas, and, critically, how the authors relate their findings to the data. And instead, I go fishing for grand connections and implications.

On the Temperature and Future of Labor Markets

You often hear policymakers and commentators talk about the labor market being hot or cooling down, and there is no shortage of opinions on the future of labor markets as we face large demographic, technological, and perhaps political changes.

On the temperature, everyone wants a number. Despite Janet’s “totality of the data,” the darling is U, meaning Unemployment (sorry!), with data input often increasing by means of further cuts of the unemployed (U-3, U-6, U-tell-me).

Here is a fact: the vast majority of people either have a job or are happy without one. And, for most workers and most of the time, the transition through unemployment is a short one—its implications may last long, but the lack of a job does not.

That is, outside recessions. Then the swelling unemployment is at the core of the job loss-demand retrenchment loop that, me thinks, defines recessions. But, me thinks again, it is then too late to arrest such dynamics—at best, one can hope to mitigate.

The point being that, if you want to stay a step ahead of labor market contractions, which is the minimum number of steps needed to make a difference, then… why are we looking at the unemployed?

What Chris and coauthors make clear is that the state of the labor market is primarily about the state of existing relationships—a hot labor market is one where most workers are close to quitting; a cooling-down market is one where firms are closer to laying off workers. Even as unemployment increases, it may be more important where new relationships sprout—are workers accepting jobs that they hope to quit for better jobs in better times; are firms hiring workers at premium rates that are unlikely to stay profitable for long? Unfortunately, we are not particularly close to having the data and means to bring this to fruition.8

Regarding the future, all we can do is speculate, and speculate we will. I have been doing some catch-up reading, from the evolution of the labor market share to the implications of AI for pensions. In many cases, I found myself returning to the visual from this paper—how is the ginormous stock of positive surpluses being realized into output changing? Is there a new, novel wave of surpluses to be found? If so, would existing contracts and arrangements be flexible enough to avoid massive job losses, and the implications of thereof?

The Worms Getting Away

Open a can of worms, and no matter how many you manage to put back, we will all be staring at the one creeping into your handbag.

Theorists will harp about the ad hoc restrictions on the contract space, and empirical labor economists will ask for a myriad of micro facts that populate their CV to be in the model. Let them be, we do want the Internet access to get faster, even at 30,000 feet.

The worm I cannot get my eyes out of is heterogeneity. People quit their jobs, firms go belly up, and so do the jobs, for the most random reasons ever. Yes, their model has an exogenous process to capture these quits and layoffs, but the can of worms is open: If we project these “random reasons” into the match surplus, then the distribution, and the dynamics underlying it, change. The gap between real wage and productivity does sound like an excellent candidate as the principal component of such distribution, and the dimension that is likely of utmost interest for a macroeconomist. But, being as it is that any decision is an instance of a max operator,

In other words, even if we have nailed the key determinant, the quit (or layoff) decision may be closer to a Probit, a random utility model, than a (s,S) system over a single dimension w.

Think, for example, of what it takes for a small-ish town to attract good doctors, or how your partner's and family's happiness shapes the value of a job to you. Or listen to those who think of location as an investment decision, with option value altogether.

Rejoinder

Hopefully I have gone the full circle: told you what is great about this paper, told you what we may still be missing, and now it is time to repeat it: This is a big step forward, and while it is not the last step forward either the profession or (I venture) the authors will, is one that both profession and authors can be proud of.

By Special Request

That whenever any Form of Government becomes destructive of these ends, it is the Right of the People to alter or to abolish it, and to institute new Government, laying its foundation on such principles and organizing its powers in such form, as to them shall seem most likely to effect their Safety and Happiness.

Prudence, indeed, will dictate that Governments long established should not be changed for light and transient causes; and accordingly all experience hath shewn, that mankind are more disposed to suffer, while evils are sufferable, than to right themselves by abolishing the forms to which they are accustomed. But when a long train of abuses and usurpations, pursuing invariably the same Object evinces a design to reduce them under absolute Despotism, it is their right, it is their duty, to throw off such Government, and to provide new Guards for their future security.

[…]

We, therefore, the Representatives of the United States of America, in General Congress, Assembled, appealing to the Supreme Judge of the world for the rectitude of our intentions, do, in the Name, and by Authority of the good People of these Colonies, solemnly publish and declare, That these United Colonies are, and of Right ought to be Free and Independent States;

[…]

And for the support of this Declaration, with a firm reliance on the protection of divine Providence, we mutually pledge to each other our Lives, our Fortunes and our sacred Honor.

[boldface is mine]

Happy Fourth of July!

There are two challenges regarding monitoring AI use:

1. We do not know what AI will be capable of doing,

2. We know that it is already capable of many things.

The vibesphere is full of takes on (1). There is a safety dimension to development and innovation that is not today’s focus (though I note that there already exists such a thing as the Invention Secrecy Act of 1951). Everything else would be speculation, particularly from yours truly.

Let me instead zoom in (2), which will allow me to highlight the challenges that are novel to such general-purpose technology, going over known and tested solutions and showing where they are inadequate.

Take a truck, for example. Unfortunately, it would be very easy to link to several examples of malicious use of a commercial truck here, including in my hometown of Barcelona. And if you have ever sat in a truck cockpit, you may already suspect that driving a truck is not easy and, needless to say, negligent use is very dangerous.

And yet, trucks everywhere! I sleep, no problem, even if I cannot say that I always drive, no problem, with trucks around. Truck drivers are licensed. The force of the law is behind the rules of the road, and there's plenty more in rules and regulations on road transport. Truck makers are also asked to build in features—including tamper-proof ones. Finally, we accept some risk and build concrete blocks around government buildings or crowded spaces.

Now, let’s try to exploit any parallels with AI.

Ouchie, ouchie, ouchie.

Monitoring user identity online is so hard that let’s forget about licensing usage.1 The best we can hope for is that we develop controlled environments that require proof of identity to access “unchained” AI. I am not holding my breath, since even age-specific restrictions are a challenge, particularly as a public policy.2

We can ask AI companies to build in features, and we can build virtual concrete walls. But here is where the astounding breadth of AI capabilities becomes a steep challenge that quickly becomes totally vertical once you consider the interactions.

See, if the truck happens to be transporting ammonia, it is not enough to slap the rules of trucking together with the rules of handling a hazardous material. You need rules for transporting hazardous materials. Coupling them increases complexity, and, indeed, we have unfortunately learned the hard way that the more interactions, the more things can go wrong.

Well, scale this up with a tool that can do… so much. It could tell you that yes, you can eat that mushroom (hint: don’t!); or how to hack your ex- or bank (not mutually exclusive); or the steps to build a mecha-Godzilla. And this is before we consider AI agentic capabilities—a chat conversation could convince you to quit your job, and the AI hopes it will, because it has already submitted your resignation for you.

Attempting a granular approach to AI safety feels like fighting a land war in Asia. Don’t.

Is there any hope that AI safety (for users) can be deployed upstream? How can we deploy AI itself to watch? Who then watches the watcher?

Before some thoughts on these answers, a small insert.

We, the Users

Have a responsibility, too. And that is to accept that controls, safeguards, etc., will also be constraints and limitations—which we can hopefully mitigate over time.

For example, it is likely that most of my token usage involved videos of my daughter turning into a fairy flying in the sky.3

Well, no more. Gemini Pro informs me that it will not produce videos of minors.

Wait, what? She is my daughter! How dare you block me from turning her into a mermaid that plays with a (licensed) pet Halibut???

At least I got a message back regarding the protection of minors, though it was so asceptic that I had to make the effort to imagine, for a second, what if it was not me? What if it were someone else using a picture of her for a video? Do I even want to know what the video request is? NO! BLOCK! BAN! BE GONE!

So, as users, we must recognize that we will bear some of the burden of preventing malicious or dangerous use, even with our perfectly acceptable uses.

I do expect that, over time, we will be able to mitigate some of the costs. In the example above, I thought a bit about what it would take to create a family boundary. I quickly stopped the train of thought as soon as I remembered on which side of the “family boundary” most of the child abuse occurs.

A reminder that, ultimately, the lesser angels of our nature may be the hardest of all hard constraints upon AI.

Upstream solutions?

I typically hope that what I write here is right. This time, though, I hope to be wrong.

I just do not see how to implement user controls upstream for AI.

Monitoring bit flows may be of some use against bad actors who seek to act at scale, but it does not allow distinguishing between granular uses: a bit is a bit, no matter what it is about.

What about embeddings? There is a lot of work to interpret and understand the inner workings of AI, to the point that one could potentially “color” embeddings as more or less dangerous for the user and malicious content. Per the point on complexity, it would be necessary to have particular safeguards about combinations of embeddings. Somewhere midstream, we may monitor and stop-gap token usage that accumulates enough of the wrong color.4

We should definitely consider these possibilities… as a band-aid for what we can do while we work on a better solution. We are facing a highly-highly dimensional space: We will spend the rest of our lives patching, I guarantee you. The patches will become a constant inconvenience for most of us. And while I see some promise in curtailing negligent use, it appears to be just a minor inconvenience for the malicious user. For the latter, it may seem that the only current solution is to cap the AI model’s capabilities (see Fable).

Now is when I leave you with more questions than answers. Actually, let me close with one more, which of course has to be in Latin.

Quis custodiet ipsos custodes?

That is, who watches the watchers?

Now, a challenge for a Research Director, particularly with the young or new uns, is: how do I describe what I need from the activated soup to turn into the “best policy advice I can make happen”?2

So I often would go for a famous line from Ludwig Wittgenstein (LW from now) that, thanks to Vasco Carvalho, we will quote via photo without AI:

Now that I belong to the 0.04% of substacks quoting Wittgenstein, I should disclose that I do not belong to the 0.0002% of humanity who understands what he wrote. I do not, and I tried several times. I even know that he renounced that quote later on, though, if anything, it just made the meaning of facts and logical space more relative, context-specific—which, frankly, I never pretended otherwise. So please consider this slightly changed disclaimer:

These are my views and do not correspond to those of the Federal Reserve Bank of Philadelphia, elsewhere in the Federal Reserve System, or by Ludwig Wittgenstein.

Ingredient #1: Facts

So, what is a fact? Scratch that, I am not here to philosophize—not with LW in the mix, at least.

So, what is a fact for our purposes here? Better! It is almost certainly… an uncertain fact. That is, it may be wrong. But it reflects some knowledge, perhaps we did not know anything about it before, perhaps we have to revisit what we thought we knew, or perhaps, no, we just nailed it, so we now know with more confidence.

I would also disavow the idea that facts must be numerical or belong to a narrow type. I am for calling “I feel like I forgot something at home” a fact, even if it turns out that I forgot it in the car instead or that “something” was my daughter. But it explains why I keep checking my coat pocket.

Thankfully, the Claude for whom Claude is named did name it best: “information.”

Indeed, “information in logical space” sounds weird, but it removes any presumption that the facts are categorical or event right.

Ingredient #2: Logical space

Ok, so what about “logical space” for our purposes?

Here I have to admit that myself have changed what I consider a logical space. Asked right after my PhD at Northwestern, I would have demanded a full set of structural equations, a thorought equilibrium definition, and a one-to-one mapping from the model to any empirical object. You know, the kind of thing that gets you published in a top journal after 5 years of work—if you are lucky.

Policy, and perhaps reality too, have made me way more permissive on what amounts to a logical space. As a matter of fact, category theory has also widened my view that we can work with order and structure at a general level and “coming from many places,” meaning, we can work with a partial order, we can work with arbitrary operations, and we can derive implications.

Indeed, an evolving thought of mine has been that structural models require a lot to “close the loop,” “get to terminus,” “make the darn code converge”… We do it because we are very smart (ehem!). But the realization is, we pay a price, for we onboard elements into our model for the sole purpose of achieving closure—and now we conflate them with the “facts.”

For example, there is empirical evidence that the demand of reserves shifts with (liquid) deposits.3 We can write one (out of many, I suspect) structural model where banks demand more reserves the more liquid deposits they have.4 You may assume a certain timing of decisions and information—it may lead to a beautiful model, but it may also be plain wrong.

Can we make progress without the structural model? Yours truly thinks so. To preview what I am working on, if the demand of reserves shifts up with deposits, the level of reserves required to close the spread between the effective federal funds rate and the interest on reserves does depend on deposits as well. Calling that level of reserves, wait for it, “ample,” I relate to the empirical objects, a.k.a., facts (the FOMC’s directive to the Desk, the findings in the aforementioned paper….). Without a full understanding of why the demand of reserves shifts up with deposits, we can conclude that, if it does, ample reserves may display dynamics of their own.5

How does this make policy better?

In so many ways, and even more ways than that.

Let me highlight first the value of the exercise itself by revealing something that surely will shock you: the first time a thought appears in mind or I learn about a new fact, it is not automatically and immediately a fact in logical space. It is more like something floating in the soup that you do not remember seeing it there. It takes some work, and often that works speaks back to me—the words often being, “what were you thinking?” and then some more thinking, leading sometimes to even further thinking, and sometimes to the swift abandonment of the thought.

And, were this happy thought to survive its mapping to a fact in logical space, it will then encounter previous facts already set in logical space, and then we are in place to quickly fold them in.

Allow me a Boolean example here. Say previous facts in logical space are simply three premises p,q,z that have a causation relationship as p → q → z. Now say we learn…

• Say we learn (not p) → q. It is compatible with the previous truth, but we learn that p is actually irrelevant: we update to simply q → z.

• Say we learn z → h. Again, compatible with previous knowledge, allows us to extend the chain p → q → z → h.6

• Say we learn z → (not p). Now all alarms must go off, for this is not compatible with previous knowledge. We will definitively have to test the fact… for if confirmed, we need to reconstruct the facts in logical space.

That reconstruction is a wonderful intellectual exercise, but not one you want to be doing overnight for tomorrow is decision time. So, you absolutely do not want to start from zero every time—having existing knowledge firmly in logical space allows to incorporate new information, even new logical constraints, and act on them.

Are facts about logical spaces facts in logical space?

Yes. But I need to read more category theory to explain it to you. Wait, let me try my new personal AI agent!

FINISH. THE. PAPER.

Paul Erdos conjectured a particular solution that is not too different from putting the dots on a grid and, indeed, only has a bit more dots one unit apart than the grid. Mathematicians spent 80 years attempting to prove the conjecture.

And AI disproved it in about 30 hours. (Read more in this Scientific American article).

OK, what is next in the template for every AI article you can read? Let me see, have I scared you a bit? By now, your palms should be sweaty, knees weak, arms are heavy (there's vomit on his sweater already, mom's spaghetti). What are your chances if 80 years of mathematicians were beaten so easily by AI?

So, next: I reassure you that you have nothing to worry about by crying wolf about human involvement, diminishing the achievement (it proved nothing, after all—it just disproved a conjecture), and complaining that AI has yet to do my laundry.

Before any of that, immediate proof that AI+I > AI.

Soon After

So the same day OpenAI announced the breakthrough, mathematician Will Sawin found a way to improve upon the AI result (that is, found how to get more points at a unit distance). The paper readily acknowledges that further improvements are likely to follow—and casually thanks other mathematicians for simplifying the original proof, so I suspect it took a bit longer than 30 hours to write the paper.2 But not much longer—and, AI or not, progress in mathematics is not counted in days or weeks.

And that’s just the start. Even if AI is limited to disproving conjectures, it will certainly direct the next 80 years of mathematicians to focus on conjectures that are more likely to be right.

Mathematicians can also immediately leverage their expertise to find ways to apply whatever new steps or arguments found in their field. Maybe one day it is possible to map the entirety of mathematics (via category theory, of course), and let AI use those guardrails to cross-pollinate ideas.

Moreover, who will know what questions to ask and what answers are worth proof-checking? Humans, of course, if you can call mathematicians that.

The profession-in-the-loop

Kudos to OpenAI on how they proceeded. Rather than jump to a big announcement and let the Vibesphere fight it over, OpenAI contacted several top mathematicians to verify the model’s output. Furthermore, they shared the comments, which combine the arcane on number theory with very interesting observations on AI.

Surely you have already been told about the “human-in-the-loop” for work with AI, typically with a focus on either an user (who would be the human) or a company/business use, in which case the human will be both in the loop and jumping through them.

For frontier research, though, it may be necessary to have the profession-in-the-loop: A reservoir of experts capable of understanding the proof and providing the trust to the public that, as of now, AI models struggle.3

Alas, I rediscovered the peer-reviewed system! It would need some adjustments that are, in my opinion, long overdue anyway.

Many seem very focused on deploying the profession to guard a perimeter around academic articles—they should be human written, and of course you can use the AI, but only so much.

My take? The profession would be better preparing for a future where AI papers will come in throves, and there shall be treasures and nasty surprises in the haystack. And perhaps this way we delay the day that I truly fear: AI produces a result that we can confirm, say, numerically; but no one is capable of understanding.

Disproving vesus Proving

As many noted, the AI did not prove anything but Paul Erdős wrong. That is, it disproved the conjecture by means of a counter-example.

It seems that AI models, which never get discouraged or get a call from school about their sick daughter, have an advantage here in disproving:

“AIs have an edge: It’s not just that they can try all known methods,” says Jacob Tsimerman, a mathematician at the University of Toronto, who was not involved in the work but was part of the companion paper solicited by OpenAI. “They can play for longer and in more treacherous waters than mathematicians without getting overwhelmed.”

There is value in proving a conjecture wrong via counter-example. First, it challenges the understanding that led to the conjecture first. Second, the construction of the counter-example can be of interest itself, even if it is unlikely (yet) to deliver the final solution (that is, the true maximum n). Third, it totally shifts the focus of mathematicians away from proving the conjecture right (out of the 80 years of mathematicians, apparently something like a month or two had been spent in disproving the conjecture).

30 hours

If you play Wordle regularly, you know that you can get lucky—there is no other word—and hit the desired all-green row at the second attempt. And while you will surely share it with your friends and family for some bragging rights, nobody will conclude that you know can finish Wordle in three rows or less.

Which is to say, the question for OpenAI, is not how long the AI model took to solve this problem (32 hours, they report—pretty cheap), but how many problems were submitted altogether, how many succeed and how many failed, and how many hours each took.

In particular, you may have heard about the halting problem. That is, when to give up (the original problem is a bit fancier). There is the risk here that successes take 30-40 hours, but fails consume hundreds of hours. For it only takes a black swam to disprove that all swams are white, but it takes all swams to prove them all white.

What my home-made AI model says

“FINISH. THE. PAPER.”

Oh right. Until the next Vibing!

Third(!) disclaimer: This article assumes some knowledge about short-term funding markets and related matters—otherwise it would be a conference marathon, not a sprint.

To paraphrase someone, if going to the dentist were as fun and interesting as this conference, I would go to the dentist more often.

The Short-Term Funding Markets conference is one of the premier events over the years for those of us who firmly believe that we can bring the best of economics and finance to understand short-term funding markets—particularly, my cup of tea, implementation, and balance sheet.

Let me share what we learned. I will cover all the papers but two and attempt to structure the summary by content. The idea is to make it easier for the reader to (1) take away a nugget of knowledge, and/or (2) cherry-pick which papers she or he will read with the utmost care.

Each paper had a discussant, and there were Q&A from the audience. Now, yours truly likes canned fish, new wave music, and some topics more than others. Which is to say, coverage may be a bit uneven!

Where deposits run - paper

Since the Silicon Valley Bank, there has been renewed attention to deposit flight, which, at the speed at which it took place, may just be another word for a bank run. Yiling Yang and coauthors argue that depositors switch from one bank to another, thereby making the state of the “destination” bank an equilibrium object. As the data show, depositors running from small or medium banks tend to head for larger banks. As a matter of fact, the model suggests the possibility of “cascade” or “herd” effects: the more depositors leave bank A for B, the riskier bank A is, and the safer bank B is.

The discussant, Alexandros Vardoulakis, focused on some aspects of the equilibrium. The margins of this Substack are not wide enough to include the details of global games—at the core, these models are based on how agents respond to private and public information and act as a powerful equilibrium selection mechanism. However, as you may be wondering… a coordination problem remains, and it is possible to have equilibria where depositors run from large to small banks.

My two-cents observation is simple: The speed at which depositors switched away from SVB strongly suggests that they utilized pre-existing accounts elsewhere. That is, at least in the short-run, it is likely that depositors had options already available—and which options by which depositors are likely a key determinant of the running flows.1

Where Collateral Sleeps - paper

I just went with the awesome paper title. Chase Ross and coauthors document several facts about pre-positioning of collateral at the Fed discount window and Federal Home Loan Banks (FHLBs). There is a fascinating collection of facts, some of which we already knew but can now document and quantify, and others that were a bit more surprising.

I do insert myself here a bit in choosing which fact to highlight: that pre-positioning collateral at the discount window, and especially disclosing it, carries stigma, even if there is no borrowing. Sigh. I did find a sliver of good news: the disclosure cost appears to be once-and-done, that is, investors and customers scratch their chin when a bank first discloses, likely wondering, why are we talking about this now? But once disclosed, banks typically continue to report pre-position collateral—without further costs.

Discussant and dear colleague Huberto Ennis offered some further ideas to evolve the paper, but summed it up in a great one-liner: “This is a very important topic, and we absolutely have the best economists [that is, the authors] that we could hope for.”

The Effects of Instant Payments - paper

You know how you can enjoy a good wine or a ride on a Lexus LFA without any idea of how they are made? Well, that’s how I enjoy most of the work on payments—with a certain measure of ignorance.

On that count, I found Yao Zeng and coauthors’ work on how instant payments reshape banks’ decisions on liquidity transformation and risk-taking… just a slam dunk. The argument is very strong:

1. Banks delay payments as part of liquidity management;

2. Instant payments take away this margin and make banks want to hold more liquidity (like, reserves);

3. Alas, a buffer for payments is also a buffer for other uncorrelated risks—hence, banks increase their risk-taking on the loan side.

Data from a very successful Brazilian instant payments network confirms the argument—I do believe that we have facts in logical space here, though I should note that the authors use instruments and other considerations discussed by Pauline Liang.

Relationships in the Federal Funds Market - paper

This paper is, first and foremost, the “next gen” in a long quest to fully map the Federal Funds market. Sriya Anbil and coauthors combine federal forms with Fedwire transfers to identify both sides of fed funds trades. The data work is spectacular, and I am frankly considering a post on the history behind this—I should ask the discussant, Adam Copeland, to write it with me, as he is effectively part of that history.

At least in the current incarnation, the paper focuses on documenting and quantifying facts that, at a qualitative level, were well known. I can say that if I had infinite time, I would recalibrate some of my models, as their datasets would provide better numbers than the ones I had to use at the time.

And this may sound a bit of a backhanded compliment, but I was a touch disappointed in the characterization of “relationship dynamics.” Is it the case that most institutions do not trade in the fed funds market often, but when they do, they always trade with the same counterparties? Or is it instead closer to a singles’ bar, where a match can last from a fleeting kiss to a 50-year marriage?

Intraday Liquidity Regulations - paper

You may be tired of hearing, here and elsewhere, capital regulations this or that, imposing balance sheet costs here and there. Well, Quentin Vandeweyer and coauthors have been pointing instead to intraday liquidity regulations, “which limit banks’ ability to act as lenders to funding markets in times of stress when central bank reserves are scarce.”

In the model, Quentin and coauthors do include both capital and liquidity regulations, and a rich model of banks, non-banks, and the central bank’s balance sheet. The result can explain the dynamics of market disruptions, and I want to repeat the keyword here: dynamics, in plural. Frankly, I am very impressed with their ability to explain repo and treasury disruptions, which are qualitatively quite different, in a single model with a single impulse.

The discussion was beyond excellent. Jay Kahn offered a simplified version that clarified several aspects, particularly regarding the Treasury basis trade. (I really cannot go into that, both because a minimum of 23 thousand words is needed, and possibly 600 words in, I will need to re-read Barth and Kahn (2025).

Bank to Non-Bank

At the beginning of time, there were banks. Then there were bank regulators. And then there were non-banks.

And yet, we still have so much to understand about how the private sector responds to regulations… by migrating outside the “regulatory” perimeter. Most critically, how do banks and non-banks interact? Is this regulatory arbitrage, a loophole of sorts, or is it that new institutions have found better solutions to old problems? Or both?

Yiming Ma and coauthors dive into the lending from banks to non-banks, and document the effect of such lending. They find a classic macro crowd-out effect from public debt: banks cut their private loans to lend to non-banks, and non-banks have been mostly absorbing public debt.

I think the question I was left with is, how does the crowd-out effect they found compare with a counter-factual (perhaps historically informed) banks-only crowding out? It seems clear that non-banks have increased the country’s capacity to absorb public debt, but since the latter has outpaced everything but NVIDIA, where are we now, meaning, more demand and more supply do result in a somewhat ambiguous calculation.

Nathan Foley-Fisher discussed, but yours truly had to step out right then. Really, it takes approximately 391 tokens to upload the slides to the conference webpage.

Bund, bund, bund

Despite being native to Europe, my knowledge of European money markets is… B-. Generous B-. So Julian von Landesberger and coauthors taught me a LOT about such markets and, most importantly, how the German fiscal authority manages its liquidity and Bund markets in general.

I am very well aware that in many countries, the fiscal authority is active in markets, investing its excess liquidity but also occasionally intervening. How do these interventions coordinate with the central bank, which in this case is a European central bank, not just a German central bank? I learned a lot about the actions and effects of the fiscal authority interventions, but less so about the motivations, governance, and coordination behind them.

Discussant Stefania d’Amico focused on the micro-properties of the assets targeted for intervention, which brought a discussion about “specialness” of assets that… ok, I just cannot stomach. Provocatively, are we sure “specialness” is not the finance equivalent of a Solow residual, which Solow himself called “a measure of our ignorance”?

WAIT THERE IS MORE

I left two papers out. First, the paper I discussed, which I am hoping to dedicate a full post to.2 Second, Darrel Duffie’s keynote “How Could the Fed Reduce the Size of its Balance Sheet?” How could you not touch on that?? The answer is simple: I learned something new from his keynote, which is a source of joy but also a time commitment: I need to read more about liquidity management systems, which Darrel highlighted as a potential route to a smaller reserve demand.

I have already joked that I know and care more than a normal amount about the Federal Reserve. Well, today’s authors know and may care more than I! Michael Bordo (yay Rutgers) and Ned Prescott (yay FBR Cleveland) have been sharing what they are learning and thinking about the history of the research departments across the Fed, with a focus on new ideas and Reserve Banks (RBs from now on, mostly excluding NY Fed). The paper at hand is here, and check also this NBER WP out here, more focused on banking. I understand a book may be on the works.1

As you may have noted, I do not really summarize the papers—they have abstracts, and in economics, most of the time, “introduction” actually means “executive summary.” So, here is the abstract:

We evaluate the decentralized structure of the Federal Reserve System as a mechanism for generating and processing new ideas on monetary policy over the 1960 - 2000 period. We document the introduction of monetarism, rational expectations, credibility, transparency, and other monetary policy ideas by Reserve Banks into the Federal Reserve System. We argue that the Reserve Banks were willing to support and develop new ideas due to internal reforms to the FOMC that Chairman William McChesney Martin implemented in the 1950s and the increased ties with academia that developed in this period.

Now, I have cut their abstract short first to discuss the “we document” and “we argue” parts, which follow pretty much from the historical record. Then I will interject myself to highlight the role researchers and policy advisors play with the less-glamorous ideas: the bad and the old.

FOMC → RBs, and pullback

Bordo and Prescott argue that, as RB Presidents began regularly attending FOMC meetings, they sought expertise in monetary policy and economics more broadly. The authors go quite far in discussing Martin’s reforms as the trigger. The whole history is fascinating and worth a read.

Prior to Martin’s reforms, the RB presidents tended to defer to the Board and the NY Fed in matters of monetary policy, focusing instead on operations and the District. Now NY and DC would not always agree, and, I am told, that is the nice way to say that the latter two would disagree often. A problem.

Martin’s solution brought increased presence of all RB Presidents to the more-frequent FOMC meeting. Lo and behold, demand for economists was on the rise. Presidents realized that they could bring valuable views and input, but also that monetary policy expertise is not “plug and play.” To do so, they needed to properly engage with monetary policy, economy, and finance at a national level.

Several Banks (ok, one, then two, then th,ree then more) realized that they could embrace frontier economic research, onboarding PhD economists and maintainig strong ties to academia. At worst, the new ideas would be ignored and forgotten. Well, they were ignored (for a while), but most certainly were not forgotten.

Actually, do let me qualify the last statement: In several cases, the Banks embraced ideas that, a decade later (or more), would become so ingrained in how macroeconomists work that the younger uns lost the perspective that what is now common wisdom was, back then, a hairbrained idea.23

The rest is history, as they say. The authors discuss the Board’s attitudes, evolving views on external publications and even a Philly incident regarding internal production. It is important to realize that the Fed, as a whole and for a host of good reasons, moves conservatively. It is not like someone dropped at the FOMC meeting, Tuesday 3pm, hey what about rational expectations? And everything changed. No, instead paradigm shifts take a while, show their value, and eventually become… old ideas.4

All Ideas

The authors exclusively focus on the production of ideas. And by that, they mean “really good” ideas. And “new.” “Grand new ideas.” “Paradigm shifts.” Rational expectations, transparency, SVARs…5

And it continues to this day. This year’s Frisch medal, a highly prestigious award in economics, went to Simon Mongey, a monetary policy advisor at the Minneapolis Fed (a.k.a., the “one” Bank).

We produce and harvest new ideas constantly, paradigm shifts do not happen that often after all! But, say, new evidence on solvency and liquidity? We want to hear it, we want to discuss it, we want to leverage it.

But new ideas are only a sliver of what we do, not our mission per se. And today I will set aside all we do in terms of products, surveys, forecasts, resources, education, outreach, publications, as well as the policy “routine” work.

No, that would be too easy for yours truly (the hubris!). I want to talk about the role the research departments play regarding bad ideas and old (good) ideas.

Bad Ideas

What do we do with bad ideas? Why, we find them, of course.

New or old, we are not nice to ideas. We are actually even a bit nasty. We challenge the facts, we stress-test the theories, we scope the model, we criticize the assumptions.6 Along the way, we learn.

Obviously, we learn whatever is still standing after the critical barrage—and if it looks like a substantial shift, does not need to change how we do economics, perhaps just forces us to rethink solvency and liquidity, then we “bookmark” it to make sure we keep track of how these new ideas evolve, when we can leverage them, evolve them…

We actually learn from what we discard, too. First, it is kind of useful to know ahead of time that anchovies and peanut butter not a good sandwich make. Second, a pitfall clearly exposed in one model could be hiding deeper in other models (particularly regarding methods). Third, even a broken clock is right twice a day! Call me an eternal optimist, but I do believe that there is always a scope for which a minimally consistent and coherent model could possibly-maybe help.

Look, I cannot say that everything I learned has directly and demonstrably improved policy. What I can tell you is the many, many times I was happy I had learned something before it rains.7 I can also give you a right now, right here example of the converse: As I am reading quite a bit about deposits these days, I keep muttering myself “darn I wish I knew this back when Silicon Valley Bank.”

Old Ideas

Now those would be the old good ideas, I am talking about. For a very important role for the research departments is to educate and support the presidents.

In the parable of the blind men and the elephant, each of the former is touching a different part of the latter, and arriving at radically different conclusions.

Well, in economics, the elephant never stops moving and, oh by the way, you, me, everyone, are parts of the elephant.

In other words, there is no single perspective that dominates any other—not for long, that’s for sure.

Which is to say, it is better to have several perspectives in the room—each of them partial, each of them valuable once they can be inserted in the monetary policy discussion. Which in turns necessitates the education and support a team of economists can provide.

The example of my former boss, former FRB Philadelphia President Harker, is quite illustrative. A PhD engineer, Pat spent a long time at Wharton, eventually as a dean, and later president of University of Delaware. That is a bit unusual for a FOMC member, but his perspective proved very valuable—I hope we helped! We certainly learned a lot of the rather unique combination of engineer and higher-education leader thinking.

Bringing a different perspective from a PhD engineer turned university president works because the room is not full of PhD engineers turned university presidents. To the left he had a PhD economist focused on banking, and a very successful consultant with three Harvard degrees. To the right, another PhD economist with a lot of work on inclusive growth, and a successful banker with regulatory and supervision experience.8

This wide array of perspectives from policymakers is invaluable, but necessitates a common ground among participants, support for the “blind spots”—tells us all you know about the trunk, we will make sure to tell you it is not a snake, and we will tell you enough about the elephant to best utilize the knowledge for pachydermic policy.

Distributed Expertise Network

I view the research departments across the Fed, to include also all the research teams at the Desk, at supervision and regulation, at community affairs… wherever they are; I view them as forming a distributed expertise network, connected to the global economics community, but also integrated within the Fed, available on demand while versed on the policymaking process—and all the minutae that often makes the difference between idea and reality.

The nodes corresponding to the RB research departments need autonomy to execute the RB president vision, because we want to get those independent nuggets of insight and information that stems from a different vantage point.

It is important to realize that policy waits for no one, so we must be able to locate and accesss expertise knowledge—on what, I will tell you only after I know, which would be, I guarantee, way too late. Being integrated into a tight internal network, which is in turn connected to economics in academia and industry, produces a large “expertise on-demand” library.

A distributed network can be a difficult object to govern and coordinate. Partly by design, which is how best connects with Bordo and Prescott’s ideas: It grants us a level of flexibility, of “oh well this is worth a try,” it accelerates adoption…. that is not easy to replicate in a tightly-centralized form. After all, who can tell best a world-class labor economist what to work next than herself? You can tell an economist a question, and idea, a goal. But if you can give him or her a complete route plan, then you do not need an economist—a research assistant and AI will do.

If you allow me one vehicular metaphor: Nowadays you only have to pull an app to figure out how to go from Philadelphia to Saint Louis—I just did, and it is just 20 or so clearly defined steps. But the right image is from the 18th century: We did not know exactly where the Mississipi flow, so, go west young man.9

I end by highlighting that the distributed expertise network approach also underpins the excellence in support for research that we have achieved, even if I would be the first to admit more can be done. Data, IT, and even administrative support is highly specialized for economic research; not just as a whole but in terms of each need: an economist working on bond prices need bond prices data, and would gain nothing from employer-worker panel data. Nor each RB can replicate the full stack of data and tech, nor it can be served by centralized entitites whose base case for standards and process is, and should be, production and operations—not research.

Taking the Colimit

Bordo and Prescott, among other things, trace the origin of new ideas through research departments at the Fed to an internal reform about something else. If it feels like a happy accident, perhaps it is.

Do not be confused, though: the “distributed expertise network” that I call research at the Fed remains the best current solution to enable policymakers to make the best decisions possible.

Now that I am afforded more time, if not more certainty, I have been attending (read, inviting myself) to several conferences. It would be a full-time job to do justice to every good paper I saw, so I narrowed it down to 3 papers. And then a thousand words in, I was only halfway with the first paper, so I decided to take one at a time.

Vibe blogging at its best!

Bank Failures: The Roles of Solvency and Liquidity, Stephen Luck and coauthors

(From Wharton FinReg IX, organized by Peter Conti-Brown.)

Stephen Luck (yay New York Fed) and coauthors Sergio Correia (yay Richmond Fed) and Emir Verner (MIT) have one of the most exciting research agendas out there, examining the past to clarify the roles of solvency and bank runs in bank failures.

The paper can be accessed here, they maintain an impressive website that I command you to check out here, and if you are short of time or interest, this CEPR article is an excellent TL;DR—honorable mention for the Liberty Street blog posts too.

The question here is, per their words, “Are bank failures caused by [bank] runs? Or are they mainly caused by deeper solvency problems, which may in turn trigger [bank] runs?”

Their conclusion, in their words again: “[…] fundamental insolvency is usually a necessary condition for bank failure. Runs are rarely a plausible standalone cause of failure for otherwise solvent banks, but they can be an important trigger of failure at already insolvent banks.”

Perhaps to the surprise of everyone but banking economists, this conclusion turns out to be quite controversial—the field has had what I can best describe as an intellectual fascination with bank runs as an indiscriminate bane that can take down even the best-managed institution.1

Before I return to the meat of the matter, let’s briefly talk data and methods.

Why We Need to Look Back

It is the year 2026, and somehow it is now that economists think of asking whether bank runs actually cause bank failures. By means of 19th-century call reports and newspaper articles that have been collecting dust in some basement or another.2

The reason is two-fold. First, it is now that we have the technology and capability to “convert natively unstructured data (information in books, images, and other undigitized formats) into readily accessible digital data.” It may not be a coincidence that I love this paper, and the previous sentence is from the FRB Philadelphia CREED initiative.

Second, the FDIC and every other safety net around bank failures (hint: supervision and regulation). I think that we are in a situation where an image may be worth more than a thousand banking papers:

I like this figure a lot because it establishes that, whatever you think drives bank failures or the FDIC and supervision and regulation do…, it is obvious that the bank failure process has changed. In other words, to use the 1992- data, you will need some structure to split the effects of policies from the fundamentals—and, wouldn’t you know it, it is hard to do so without imposing some prior views on liquidity and solvency.

The Results

I have some “contentious points,” of course, I always do.

But this is a paper to learn from. Effectively, they have established two facts that we can carry forward with confidence:

1. Solvent banks typically survive bank runs,3

2. Insolvent banks typically fail, whether or not they experience a bank run.

It is quite clear why this paper takes “bank runs” one or two pegs down in many banking economists’ priors of what banking regulation should be about.

Note, this is the data. I have heard many take their results as if bank runs cannot take a solvent bank down—that would be a statement about feasibility, and, frankly, no ship can be built to weather all storms. Let’s keep that in mind the (faraway) day I can touch on either this FDIC report or stablecoins.

I cherry-picked one of their figures that maximizes a convex combination of insight and immediacy.

So what you are looking at is, on the horizontal axis, years before the bank fails—in dark blue, those that failed and experienced a bank run, in light blue, also known as cyan, those banks that failed and did not experience a bank run. For comparison, in yellow, you have the non-failing banks—obviously flat.4

On the vertical axis, the most non-deposit reason I can think of for a bank to fail: Bad loans were made. More precisely, the fraction of non-performing (aka, “bad”) loans. Four years before failure, yet-to-fail-but-will-fail banks are statistically and economically different from never-to-fail banks regarding their loans.

Deposit outflows do not seem to make a difference. Now that will naturally turn your mind to think about predictability, but that you cannot quite resolve from the Figure above. Instead, you need the Receiver Operating Characteristic curve, which they dared to call something else that is not fit to print, period.5

What is Solvency, What is a Bank Run

My contentious point is… categorical :) Even worse, it is not even the authors’ fault, and I have to ignore their more recent work, where this is handled much better.

The problem, in my opinion, is that the underlying debate tends to sound like an “egg or chicken” problem, with “solvency” and “bank runs” in place.6 The authors use quite a bit of care, but ultimately, they must speak to the field.

The first categorical point is obvious: “Solvency” is a bank property, “bank run” is an event—and to complicate things, “liquidity” is typically a property of an asset.

Sticking with the “type” event, we can compare failures and runs. However, categorical me would like to remind you that “failure” is defined as a terminal event. There is no reverse causality. Worse! There is no, and there will never be, an observation of a bank run on an already-failed bank. The same way there will be no observation of a dead person petting a cat—so proving that petting cats extend your lifespan is hard (we catlovers held this truth to be self-evident, obviously).

I found this paragraph from the authors both insightful in their views but also the “asymmetry” that establishing causality from a property, or from an event, are two very different mapping.

I think the problem with this quote is that the “bank run” is subject to a much higher standard: “it is not the root cause.”

Are losses on loans or investments the root cause of the problem? Surely I can cd .. and talk of the root cause being whatever unexpected befell or, in its absence, the poor loand and investment decisions of the bank.

This is not theoretical physics. This is economics. There is always a further root cause. Don’t be my five-year old daughter and keep asking why. Believe me, you will end up concluding that the root cause of a bank failure is the existence of a bank.

I am pushing this to the absurd, hoping to make the point that A causes B does not mean A is more important than B. More important for what? In my view, for the design of policy, though I recognize that, yes, I would like to hear more about A too.

Now, fundamentally, I have to agree with the authors in the sense I made up for myself, that is, for policy. Consider two extreme policy “focus,”

• Ensure banks are solvent, leave them at their fate regarding bank runs;

• Prevent all bank runs;

Again, as Sir Isaac Newton said, the “truth is in between.”7

I think it indisputable that the paper is shifting the policy focus in this sense.

One More Thing

I am breaking through my self-imposed word limits, but just one more thing.

Maybe too much time nearby Itamar, but I do believe there is such a thing as a deposit franchise, and if its value were to be reflected as an asset for the bank, I wonder whether we have a natural framework to marry solvency and bank runs: some bank runs could destroy enough deposit-franchise value to turn them… insolvent. Now we have split the question in two:

• how much does a bank run destroy deposit franchise value?

• how much loss in deposit franchise value can a bank withstand and remain solvent?

I have clearly reached the point I am writing to myself, which is the cue to wish the stray reader a nice weekend!

The Federal Reserve System is a highly matrixed organization—across locations and across functions. The challenges stemming from geographical dispersion do not hold a candle to those stemming from the different functions. The same function already looks about the same across locations. But an economist and a cash manager in the exact same location are still an economist and a cash manager.1

Let me be brief today and conclude with some poetry, apropos of nothing, of course:

Do I contradict myself?

Very well then I contradict myself,

(I am large, I contain multitudes.)

Walt Whitman, Song of Myself

Some like turtles, I like metaphors. They help me frame and analyze a problem or a solution, without losing sight of the fact that, at the end of the day, public policy must be understood by the public. Someone else (not a Fed employee) agrees with me:

“[…] the greatest thing by far is to be a master of metaphor.”

— Aristotle

Actually, sorry Peter, this one is a bit over the top. So switching to a more modern intellectual hero:

“Metaphors are fine if they aid understanding, but sometimes they get in the way.” — Richard Dawkins

GRIMPs will take metaphors to a completely new level of nonsense, relying on a graphical representation (that will not always include Robert Smith) to trigger a chuckle and a thought about monetary economics.

To make things worse, I discovered I can run a poll on Substack. Which, of course, is naturally our first GRIMP and is about the Fed’s balance sheet.

The Question

How should a depository institution view the discount window at the Fed?

The poll is right below and is anonymous in case you are worried I know.

Further below, a couple of sentences explaining/arguing for each version.

A. Water Cooler

Thirsty? Come over here, the cups are over there, cool and does not taste metallic like the water fountain by the bathrom. Drink as much as you want, it is water after all. And, the cups? They are volumetrically insignificant and after the first sip they start a message that say this cup will auto-destruct itself in the next five seconds. So it is not like anyone is going to drink too much.

And you do not want your workers worried that they will spend the afternoon thirsty because they left the water bottle at home.1

B. A Vending Machine

Sometimes a snack and a soda is what gets you through the extra three hours of work at night—well it is there, it does not bother, it is at a price that in good companies is subsidized, and that in an airport is like $8 for a water bottle that is minuscule?

Again, also, good luck coming out, how shall we said it, nutrionally satisfied.2

C. The Fire Hose

Or the emergency equipment. Whatever. The solution to a problem, a targeted action to resolve quickly a particular dangerous situation. Something to use in an emergency, not every Monday to water your petunias.

You will be satisfied. And drenched and in embers, but safe for the time being.

D. Helveg

Or crossing the Styx or whatever it is you believe it happens, only one way. Remember I asked “should.”

Bonus: Robert Smith

In apparent defiance of Physics, too.

At a certain job interview, I was asked if I would be comfortable answering a question in five minutes. And I had to go and say, well, the truth: There is no question I can answer in five minutes that is worth five minutes of mine.1

Alas, sometimes I do find myself in need of an answer, any answer, in five seconds. Say, when a very prominent economist who shall remain anonymous but whose name is Randy Wright, asked a brutal question,

Why is currency a liability of the Fed?

And I had to answer in five seconds because there was a plenary talk I wanted to deliver, and it wasn't about currency. Which I managed to do… but Randy’s question was now stuck in my head. Is.

So here we are. Yeah, why?

The short answer

The short answer is: Central banks remain banks, and whatever note/deposit/IOU a bank issues is a liability—regardless of the conditions for redemption.

Of course, an even shorter answer is Section 16 of the Federal Reserve Act.2

These were the answers I fitted in the five seconds, and they are correct, but they are not satisfactory.

Why not an asset?

It is worth taking the contrarian view, Randy would agree.

Currency generates a stream of revenue to the Fed, and earning a return is what you get from holding an asset, not a liability.

Moreover, currency obviously is of public service—dare I say, a public asset? If not, why is it illegal to burn money? I am pretty sure the Treasury would have no problem if people decided that smoking ribs with T-bills is a good idea.

Why a liability?

The correct answer is: Because you can pay down taxes with it.

Let’s walk through what you (hopefully) did two weeks ago with your 2025 taxes.

• Assume you had no income withheld in 2025.

• Due April 15th, 2026, you owed $10 to the Treasury, say on Day One 2026. Your taxes are due—effectively, the Treasury is lending to you until tax day. Thus, taxes dye are a liability to you, and an asset to the Treasury.

• You pull a Nick Strafford from Cedar Buff, Virginia, and pay your taxes down in pennies (last chance!). Please do not.

• You have $10 less in currency, but you are current (no pun intended) with your taxes. Since the taxes due were clearly a liability to you, it must be that those $10 were an asset to you.

• The Treasury has $10 less… in tax claims to the public. That is, it has $10 less in assets. It follows that those $10 in currency must have been a liability of the Treasury.

I also tried another thought exercise with the idea of a date t=0, a “state of nature”, before monies and taxes are introduced. Too complicated.

This is banking.

Now, I have not said anything superobvious in a while, so…

A central bank is still a bank.

Actually, it is far from obvious—I just had a terrible vision of Randy asking why. And the Board of Governors is most definitely not a bank.

It is, though, a distinctive feature of banks that their liabilities provide services to their clients, aka depositors. It is through those services that banks can make a margin on deposits—well, that, and the FDIC deposit insurance, of course.

The central bank also provides services via its liabilities—this time, to qualified institutions (we call them depository institutions, not by accident), the Treasury, or the public at large—the latter via currency.

Itamar Drechsler has done a lot of research (and talking) about the “deposit franchise,” the value associated with a bank's deposits via services (or via market power). Similarly, many have talked about the “currency franchise”—including what happens if you end up without one (Hej, Sweden!).

I cannot resist plugging in a Mercatus interview between David Beckworth and Ruth Judson, two more people who care about the Fed’s balance sheet more than a normal person would. Warning: This one is a deep rabbit hole!

Taking the Colimit

I have reached two conclusions.

First, as an economist, NEVER take a shortcut when it comes to accounting, regulations, and related paraphernalia. Instead, sorry, but you've got to dive in, read and understand, and come back from the jungle with an easy path for the rest of us. And yes, I am talking to myself. Again.

Second, going long on my second decade working on the Fed’s balance sheet, there are still questions I have not asked myself. And I would need to curse to accurately qualify how lucky I am to get to rent the brains of very smart audiences—and Randy’s, too.

Randy’s Reply

The fact that you provided such a detailed analysis suggests it’s not a triviality, even if one might say it is a purely semantic issue.

Consider just one more example please. Suppose you give me a beer today and in exchange I give you my note saying “IOU $10.” That is my liability. Now suppose you give me a beer in exchange for my note saying “thanks” or even “thanks sucker.” That’s not my liability… unless maybe I accept it for a “tax” I can impose. But wait - if I can impose a tax staying “give me back my note” in exchange for *nothing* hardly makes that note a liability - it is like me giving you an IOU that I can cancel whenever I like. Hmmm.

To motivate today’s article, let me start with an apology: I have not been able to post as frequently as I wished, partly due to a very welcome surge in engagements, but mostly because… I have been working with AI. And, as a result…

• I keep biting more than I can chew. Love my way, I follow where my mind goes—but now it is cheaper to get an answer than to wonder if the question made any sense.1

• The Red Queen Race, “it takes all the running you can do, to keep in the same place.” Technology is changing so fast that I find myself adjusting my workflow every other day—forced to do so, too often.

Along with all the time gone, there has been a fair amount of agentic anxiety. Claude Code is truly the stuff of Sci-Fi, so what is Mythos going to do to whatever is left of my career? What about my daughter’s career—she is five?

The executive summary is truly short today: If AI is in your team, with you, you will stay ahead of an orphan AI. The bad news: you must. But, the good news: it is a blast.2

Explore, Play, Vibe

Let’s start with the good news: This is an amazing time to be alive.

Costs are low enough, and even free options are competitive, so get in and… vibe. Do you code often? Do you know what CLI stands for? Then you can jump to Claude Code.

But please please please do not let what you do not know limit how you use AI.

Let what you do not know lead how you use AI. Ask, explore, play.

Rather than keep sounding like a life coach, here are some “ice-breakers” with the AI:

• Turn what I say next into a robot voice: Rosa put your PJs. Some prompts later, we made a 30-second track that is taking dance floors by storm!

• Create a Quiz About Pop Songs and Days of the Week Didn’t work very well at first—I got most wrong. This one is a true example that did not end up as a dance hit but helped me understand the context.

  • To make this work best, ask the AI For each day of the week, suggest the most popular pop songs that reference the weekday. Preference for songs from 70s, 80s, and 90s There were more iterations, I have no good reasons to bother you with: The point is to make the AI load the results in its “memory,” and then when you ask for the quiz, it knows what to tap—and what I meant with the and between pop songs and Days of the Week.

  • More generally, AI is not only amazing at answering quizzes, but at creating them. Is it a nice pastime, actually—listened to the whole Rome podcast? Let’s see how much of it I slept through…

• Please have Hesiod interview Robert Smith. You can ask for questions and answers, though if you happen to be more or less knowledgeable about a historical or pop figure, you can take over the Qs. Or even the As.

Most of you have also seen a fair amount of AI-modified images here. I am not sure that I would send you that way with promises of fun and aha moments. There is a fair amount of frustration as I struggle to communicate what I have in mind, and the AI struggles to map it to an image.

Ask, ask, ask

You can just use AI as a super-powered search tool—well, I wouldn’t know where to go for a search tool sans AI. You will likely end up going down a couple of routes, or perhaps you are already there. Let me share:

Ask for help

One day, we will have to explain to our kids that when we typed /help, nobody actually helped you.

Now, the AI will help refine what you want to ask it, though some were clearly trained on too many real custom service calls.

If you are working with an agentic AI, it will soon offer to do it for you. All I will say, before you decide, ask yourself, were you learning or doing? Because you may have been used to learning by doing, and now it is much harder to learn than to do—and clearly “doing” is what the AI is focused on, uh, doing.

I had mixed success in asking straight for AI to teach me,but perhaps that is my biases, of course. To evaluate the AI’s teaching capabilities, I asked for things I already know, like the history of Catalonia or the essential New Wave, and immediately regretted it.

But there is an upside.

Test AI in your turf

Allow me to sneak in a bit more advanced, which is to say pedantic, point about asking about what you know.

And I do not mean asking the AI how to boil pasta.

I mean, trying to take the AI down a road you know more about than the average person. Avoid the professional first. For example, I believe I know a bit more than average about physics and physicists.3

I mean, did you know Paul Dirac was a classmate of Cary Grant? Or that Olivia Newton-John was the granddaughter of Max Born? Well, AI managed to answer 50% on this, and perhaps it may be down to the second question, “which famous scientist was Olivia Newton-John related to,” being a touch of a give me for the singer of “Get Physical.”

Note that AI can tap the internet, but depending on how you phrase a question, it clearly stops ASAP. Asked “who was a famous classmate of Paul Dirac,” for example, AI seemed to be satisfied to conclude that if I knew who Paul Dirac was, I would be happy to learn he sat right behind Einstein at Solvay.

I felt that I learned a bit about the AI's knowledge limitations with this approach. Mind you, mostly I have been impressed by how far I had to push it to hit it, but I have enough specialized knowledge that I typically found my way to a tiny spec of the frontier.

I wish there were an equivalent approach that would work for coding—much harder for me to find where I am asking too much, hopefully because I lack the skills of asking too much—hopefully not because I lack the skills to realize it has been three questions ago that AI was the one vibing the answers.

Which get me to the bad news. Which are not really bad.

Jumping On a Moving Train

It is what it feels like, about every morning.

I fire up the laptop and greet the usual suspects: one or two Claude agents, five or six Visual Studio Code windows, and a tetratrillion of Chrome tabs.

There will be updates. Which one, you do not know, so let’s quickly take a snapshot right now:

• Chrome wants to update. Surely it will change once again how their “agent” Gemini (pretty weak) works.

• Which will send Grammarly who is now Superhuman (a company name, not a qualification) to change how it works, and frankly familiarity with how it worked was the only thing left for me to cling on.

• My Windows Update has become a misnomer. As if Windows does not send an update every time the moon is full, you can find what seems to be half of Fortune 500 in my updates list: Intel, Lenovo, Dolby, Samsung, Logitech…

And there is whatever it is that today is going between Visual Studio Code and Claude Code. I should not bother you too much about it, and I have been trying to understand /hooks to fix the issue. However, it is recurrent and maddeningly frequent, with both Claude Code, VSC, and every single extension apparently ongoing a frenzy of updates. I am quite sure that those updates are AI powered, powering further AI,…

Did I mention the train is actually accelerating?

No need to be on top… but stay within sight

Now, it is an overkill and a full-time job to stay on top on everything new AI. Problems of updating and compatibility are worse the newer the product.

Indeed, even if you catch the train, you will continue find yourself falling off and running after it again.

What I am aiming for instead is to stay within sight. Right now the train is moving too fast, let’s accept it. I am also not such a poweruser that I can ride the swinging train with my eyes close; I also have some urgencies to get some things done ASAP professionally.

As the time availability fluctuates, so does my distance to the AI frontier. It reminds me what we called “la goma” in biking, where the slower bikers (meaning, me) would slowly fall behind the group, rush to catch up, and rinse and repeat.

But I am ok with not ideal. Workflows are good enough already to do start producing, and even I get tired of the Fed balance sheet and need a break… back to AI vibing then!