Streetwise Professor

Bloomberg published a long article by an Icelandic hydrologist on AI trading bots. In some respects it is a contribution to the AI doom literature. It worries that many developers using the same AI platform (Claude Code being explicitly mentioned) and the same training data will cause their bots to enter into highly correlated positions that are subject to destabilizing unwinds in the event of a market shock. (Hedge fund basis trade unwinds are an example of such an unwind.)

At a superficial level this makes sense, but the analysis fails to think through the equilibrium effect of AI trading bots. Indeed, the low entry cost that Darri Eythorsson correctly identifies as the distinguishing feature of AI trading applications mitigates the effects he frets about.

In essence, AI bots can be viewed as privately informed traders who extract predictive signals from masses of highly dispersed data. Moreover, the low entry cost means that there can be–and will be, if there are not already–large numbers of such informed traders. These will be competing intensely.

Market microstructure theory predicts the effect of intense competition between many informed traders. Specifically, the model of Holden and Subrahmanyam (J. of Fin. 1992) predicts that these informed traders “compete aggressively and cause most of their common private information to be revealed very rapidly.” I italicize “common private information” because that is exactly what Eythorsson posits that AI trading bots will produce (because they are developed using the same tool–Claude Code–and the same data). Thus, the trading bot scenario fits the HS model almost precisely. Moreover, HS state that in a continuous trading environment (which characterizes virtually all equity and derivative exchanges) “market depth becomes infinite and all private information is revealed immediately” with a large number of entities trading on the same information.

This is in contrast to the situation (as in the canonical Kyle model) where there is one (or a small number) of informed traders, in which case information is incorporated only gradually in prices and market depth is finite.

The increase in depth is arguably a very salutary effect, which should be considered when weighing the benefits and costs of a trading bot army–including the costs of position correlation.

There are other implications. Prices that rapidly and fully reflect private information limits the incentive of the bots to accumulate large positions. Moreover, the competition between a large number of bots means that although their positions may be correlated, they will not be concentrated. A small number of concentrated positions (“whales”) pose, in my opinion, a greater systemic risk than a large number of correlated small ones.

The intense competition also drives the economic profit of bots towards zero. This is a self-limiting feature that constrains the aggregate scale of bot operator positions.

It should also be noted that truly smart bots will make trades based on information about other bots. As an example of this “reflexivity,” bots will attempt to predict the positions held by other bots, and will have considerable data to allow them to do so. This information is relevant to assessing the risk of cascading liquidations of correlated positions, which would reduce the amount of piling on by bots that (as they should) take both risk and expected return into account.

Of course, big hedge funds and quant firms hate this development. It essentially commodifies their expertise and human capital. This should make you treat any scare stories about trading bots with a healthy dose of skepticism, and indeed cynicism.

The main legitimate concern about AI trading bots is one that is common to all privately informed trading. Specifically, what social value does it produce? Is investment in private information for trading purposes primarily rent seeking that generates private profit (at someone else’s expense) but not commensurate welfare gains?

Privately informed trading extracts wealth from the uninformed, who trade for portfolio adjustment/risk management reasons (or for speculative thrills). This is purely a transfer, and one that requires investment in real resources to acquire and utilize the private information. In the case of AI bots, these real resources are the cost of developing and operating the bots. This is costly rent seeking, pure and simple.

Whence does the social benefit come? It is an unintended consequence of the rent seeking: more informative price signals that lead to better allocations of resources (e.g., better capital allocation or production decisions).

Financial economists have tended to trumpet the beneficial effects of more informative prices and “price discovery” without quantifying them, and discussed the rent seeking aspects of informed trading sotto voce, if they discuss it at all. Thus, a true cost-benefit analysis has not been undertaken.

In some respects this is understandable, because calculating the benefit of more informative prices is effectively impossible. (Maybe AI can someday make this calculation!) As is the benefit of potential speed-ups of the revelation of private information by a few minutes or seconds. But color me skeptical that it does.

So what it comes down to is a comparison of the real resource costs of privately informed trading by AI bots vs. the current set of privately informed traders, hedge funds, quant traders, investment banks, and the like. Here, the very cheapness (and hence ease of entry) of AI-based trading tools strongly suggests that rent seeking costs will decline relative to those in the current system.

As an extreme case, if creation and use of AI bots is free (which it isn’t, but consideration of this case is informative), we would get more informative prices at zero cost. Indeed, there would be significant cost savings due to the elimination of costly hedge funds, quant firms, etc. Indeed, this would represent a real world refutation of the Grossman-Stiglitz proposition that fully informative prices (“strong form market efficiency”) is impossible. Yes, there are costs, but if costs are low we can approach fully informative prices.

The main lesson is that if AI is a cheaper way of acquiring and trading on private information than the incumbent system, their displacement of incumbents would represent a welfare gain (though not a Pareto improvement!). This gain would come in two forms: lower rent seeking costs and more informative prices.

But that’s where the debate about the benefits and costs of AI trading should focus. The welfare effects of privately informed trading are ambiguous, at best, and such trading has a strong rent seeking component. If AI trading bots reduce rent seeking costs, their development is salutary. If not, it isn’t.

The very feature that Eythorsson highlights–the ease with which anybody can create a trading bot–suggests that AI will reduce rent seeking costs, and microstructure theory suggests it will improve price efficiency. Lower costs and higher benefits than the incumbent system represent a decided improvement.

Combine that with the likelihood that, as argued earlier, the equilibrium effects of competition among AI bots is likely to have little effect on systemic risk (and potentially may reduce systemic risk by reducing position concentration), well, don’t worry: be happy. Unless you are an incumbent privately informed trader that is. In which case, the trading bot reaper is coming for you.