Monetary policy is going asymmetric 

The stock market was down today because the FOMC “signaled the possibility of further policy tightening,” Bloomberg reported.

Or because Chair Warsh “highlighted that the inflation risk wasn’t improving,” CNBC said.

Or something else, maybe.

We can never know for sure why the market does what it does — because “the market” isn’t even a thing, really.

Saying the market went up or down for some identifiable reason is to anthropomorphize an abstract representation of millions of individual decisions, made for a million different reasons. Only some of which we can guess at.

Sometimes it’s easy to turn that process into a story, like today. Today’s story — rates up, markets down — is a standard one.

But it’s only a story. I know this because I spent about two decades telling them. A large part of a trader’s job at an investment bank is coming up with plausible-sounding explanations for why a stock (or the market) went up or down that day.

I was very good at it! Only rarely did I resort to the last-ditch explanation of “more buyers than sellers” (or vice versa).

But AI is about to make this a lot harder.

Consider that part of the reason the market was down today may have been an expression on Chair Warsh’s face  — because even a look from the Fed chair can move markets.

Expressions of concern, stress, or unhappiness seem to have the greatest impact. “We find that investors adversely react to negative expressions revealed during the press conference,” one study found.

The researchers used facial recognition software and machine-learning algorithms to determine that, on average, the S&P 500 fell 0.53 basis points in the three minutes after a negative look from the Fed chair. Index volatility rose 3.8 basis points. 

“Market participants observing Chair's negative facial expressions during the FOMC press conference may associate similar negative feelings with the discussed topic,” the researchers concluded.

In other words, if a Fed chair says “inflation” with a sour look on his face, the market will think the situation is worse than they’re saying.

Unfortunately, the study did not consider the effect of smiles, winks, or raised eyebrows, because there wasn’t enough of them: “The amount of positive facial expressions in our sample is marginal.”

That must make the above photo from today’s meeting an exception that proves the rule — perhaps because Warsh is still new to the job: "As their tenure increases, negative expressions become more frequent, eliciting adverse market reactions," another study found.

Sad, really. For central bankers and markets both.

Whatever his current mood, Chair Warsh is probably unaware that his face muscles can move markets on their own: “I find that Fed Chairs do not strategically control their expressions,” the author concluded.

They should, though, because strategically moving markets is most of what the Fed does.

The purpose of monetary policy is to affect the real economy, but the Fed does that indirectly by affecting markets first. The FOMC communicates its policy framework (or “reaction function”), and financial markets use that information to price bonds across the yield curve.

Lower bond yields raise employment and higher yields lower inflation (in theory).

But this only works if there’s a shared understanding between the Fed and markets. If the Fed doesn’t know why markets are moving, it won’t know how to set monetary policy. If markets don’t understand why the Fed has made a monetary-policy decision, they won’t know how to price bonds. 

“To get policy right,” Chair Warsh has explained, “we also need to get the relationship right between financial markets and the central bank.” 

Anything that makes the markets’ reaction harder for the Fed to read weakens that relationship — and, in turn, the Fed’s ability to affect the real economy. 

In the case of facial expressions, the effect is measured in increments that only an academic economist could discern.

But economist Markus Brunnermeier warns that there’s much more of this to come: AI, he says, is creating an "asymmetric understanding” between the Fed and markets.

Large language models trained on the entire historical record of Fed communications — minutes, speeches, testimony — give AI agents an increasingly sophisticated understanding of what the Fed is likely to do in any given situation. 

But the Fed has little understanding of what AI agents are likely to do. The decisions they make cannot be fully translated, audited, or understood by even the engineers that developed them, let alone the economists at the Fed.

This is a problem!

In addition to making it harder for the Fed to do its job, it also makes the Fed easier to manipulate.

In normal times, monetary policy is a “common-interest coordination game.” The central bank and financial markets work together to promote economic prosperity and stability.

“But it can also turn adversarial,” Brunnermeier explains, “especially when the central bank’s balance sheet is involved.” 

He cites the canonical example of George Soros attacking the British pound because he could perfectly predict how the Bank of England would first defend and then abandon the exchange rate.

Now imagine AI agents doing something similar.

Colluding AI agents can engineer vulnerabilities and set hidden traps, Brunnermeier warns, forcing central banks to cave into demands like cutting interest rates or buying bonds to avert a manufactured crisis.

Unlike in the Soros example, though, the central bank won’t know why it’s being attacked or by whom.

Scary.

To make this nightmare scenario less likely, Brunnermeier believes central bankers will have to reverse their decades-long move to greater transparency.

“With asymmetric understanding,” the Princeton professor explains, “transparency has to be rethought as predictability arms the opponent in the financial dominance game and invites ‘moves’ that trap the public authorities. Hence, there is a case for more opacity.”

In other words, central bankers should stop explaining like Jerome Powell (an open book) and start equivocating like Alan Greenspan (“constructively ambiguous”).

But that is only a start because language models can see beyond speeches and press conferences.

“AI also limits the power of opacity,” Brunnermeier says. “Agents extract and game rules even when they are never communicated.”

He cites the Fed chair’s facial expressions as an example.

Should central bankers start wearing masks then? Work on their poker faces? Mumble in monotone?

Brunnermeier thinks they may have to consider more radical measures.

One is to become genuinely unpredictable. “AI will discover any hidden rules unless actions are truly random,” he says. He suggests central banks adopt “randomization within announced bounds.”

Imagine the FOMC setting the fed funds rate by pulling its decisions out of a hat.

Beyond that, central banks might have to consider the nuclear option: bypassing markets entirely. 

Instead of relying on financial markets, the Fed might choose to affect the real economy directly — by buying and selling bonds of all durations, for example, issuing loans directly to consumers and businesses, or offering “remunerated CBDCs” (giving it direct control over the supply of money).

“Central banks may want to move from recruiting markets to side-stepping them,” Brunnermeier concludes. 

That would be truly a last resort, however, because even a swarm of colluding AI agents could not predict what it would mean for the economy.

But it might still come to that — because asymmetric understanding may demand asymmetric risk-taking.

— Byron Gilliam

_{Brought to you by:}

Crypto's premier institutional event is arriving in Asia this October 7.

Join us for a day of top-tier conversations and networking and hear CEO & founder voices from the top protocols in the Space.