The Infinity Machine, the new biography of Demis Hassabis, mentions that at one point he had a skunkworks project inside Google to apply machine intelligence to investing. Which is not a bad idea! AI research and quantitative training share a similar fundamental process, and there's some overlap between people who've done one and the other ($, FT). They're obviously using different kinds of data and different kinds of models, but in another sense they're very closely-tied: the transformer architecture is a clever way to say "if you have x, followed by y, what's the most likely z," and chain such predictions together. In both cases, clean data beats more data.

But in another sense, these domains are fundamentally different. AI training data generally consists of people trying to communicate with each other. If there's a Reddit thread from someone asking why the grass on their lawn keeps dying, there will be lots of earnest answers from people sharing some combination of their general lawn-care knowledge and their specific lawn-care experiences. Whereas the market is closer to a consortium of people who get together to misdirect all of their neighbors, since only one of them can have the greenest grass.

A trading signal is generally some indication that some other market participant is making an error, either of commission or omission. Betting on short-term mean reversion, for example, means providing liquidity to people who put on large trades in a sloppy way. Statistical arbitrage is a scalable way to make the point that a price change in one stock should, if markets are reasonably efficient, reflect changes in other assets, too, but insufficiently diligent traders might over-fixate on one company, or just not get the relative impacts right.

And every time one of these trades gets implemented, the underlying signal gets weaker. Maybe, absent systematic mean-reversion, a given attempt to liquidate some stock will push the price down by 1% over the course of a day, all else being equal. But suppose there's a trader who can identify that kind of selling, step in, and stabilize the price when it's down half a percentage point, expecting it to drift part of the way back up in the next few days. Suddenly, your data no longer shows that big liquidation! Instead, it shows a stock sinking slightly on above-average volume, and weakly outperforming soon after.

In fact, your mean-reversion trader might be thinking a few more steps ahead than that! Suppose the profit-maximizing approach is the one outlined above: when you've identified a sloppy seller, buy when half the unaffected market impact has been realized, and hold for a few days. That means the market impact is smaller, but someone carefully studying the numbers might see it—and it might catch their interest if the average move shrinks by half, at which point they might implement a similar strategy, but calibrate it to buy when the stock is down 45 basis points. They're making a bit less than the original trader, but if the strategy is still profitable, they may be satisfied with that. Our trader could avoid this by guessing what the odds of being copied are as a function of how much alpha there is each time a signal is getting used, and might conclude that this signal can, for example, produce $1m in annualized profits with a 50% chance of being copied each year, or $500k in annual profits with a 20% chance.

At that point, someone who's incredibly careful about analyzing the data might reach an unfortunate null-hypothesis result: "It used to be that if you saw a 500-share sell order every five minutes, always at the midpoint of the bid/ask spread, you knew someone was selling and could wait for the stock to drop and then start buying. But now, that kind of trade barely moves the market at all!" So, to them, equal-sized orders made at equal time increments are noise, not signal. But that's just because they're already someone else's signal. Because everyone’s behaviors obfuscate the truth, the result is like training a language model exclusively on North Korean propaganda and then asking it to write a coherent story of the twentieth century: much of what you’d want to train on is simply missing from the data.

Solving the market at any given point means identifying:

1. The set of all signals that nobody else is exploiting, and, related to that,

2. The set of all signals that arise from other people's systematic strategies.

Unfortunately, the second set of signals often shows up when a strategy matures, and goes through a cycle where too much money is flowing into it, those inflows temporarily push prices in the direction the strategy predicts, investors collectively conclude that even though the upside of the strategy is not what it once was, the variance is lower so they can still achieve a decent return by levering up—and suddenly there's a whole ecosystem of related trades that will all unwind at once. The Quant Quake of 2007 is the canonical example of this: there were funds diversified across fixed income arbitrage, equity statistical arbitrage, owning factors like value and momentum, etc. But since they were all diversified across those strategies, and they were all levered, when any one of them started shrinking its balance sheet, all of them did, and everything went down.

So the property that emerges from the interactions of lots of traders who are looking at the same datasets, using the same statistical tools, and coming to the same (correct) judgements about historical performance, is that at some point in the future, these strategies will be perfectly correlated in that they'll all lose money in sync. It's a signal you get to test exactly once.

And it implies something important: to fully solve the market, you have to infer the counterfactual market state if each of a thousand other market participants hadn't busily solved some part of the market. And then you have to implement something that implicitly replicates their strategies, and then a model to predict their suboptimal reactions when those strategies hit a rough patch. Some people have gotten rich by doubling down when they know they're right. They haven't necessarily stayed rich doing that, but any given snapshot of the rich will have people like this. Such a snapshot will also have people who followed the opposite policy: being fairly conservative most of the time, so they could be aggressively opportunistic when there were big opportunities. This is all out-of-sample. And, even more frustrating, the historical trend is that when markets are generally stable, it leads to too much leverage and they destabilize again. It'll feel the most like asset allocation has been perfectly solved just before another 1987-style break or 2008-sized collapse.

The last reason it's hard is that making billions of dollars using AI to trade is probably not the most valuable thing you can do with those skills. Big fortunes have been made and lost in finance, but if you look at the top of the Forbes 400 list, owning software companies, especially the ones that sell ads, is where the real money is. If your world model is so robust that it can simulate what a portfolio manager at AQR would do in response to a decision by a quant researcher at XTX, you can probably use that deep psychological insight to convince either of them to move to a nicer apartment, buy a faster car, take a ludicrous vacation, etc. And you can do the same for all the high-income people who don't interact in a market-facing way.¹ The people who made the most money from deeply understanding orderbooks, latency, and adverse selection did it by putting ads on news feeds and search results, not by trading.

It’s always useful to have a sense of how efficient the market is, where it’s more or less so, and why. The Diff has looked at this question from multiple angles:

• We’ve looked at how crowded strategies create accidental adversarial attacks.

• There’s more than one efficient market paradox.

• More thoughts on correlations going to 1.

• And some reflections on Rentech.

Join the discussion!

1  In an interview, Jim Simons mentioned that his family’s original name was “Sutzkever,” so there’s a decent chance that he and Ilya are distant cousins. Depending on how cofounder equity at SSI got allocated, Ilya may already be worth more than Simons was, and at a much younger age.