Trading-agent model reveals market stress that volatility can miss

Identical snapshots, different sequences

Volatility measures how widely returns vary, but not the order in which gains and losses arrive. Irene Aldridge’s preprint, Learning AI Agents, Market Makers, and Stress in Automated Markets, makes that distinction central to a theoretical analysis of automated trading.

The paper studies a simplified constant-product automated market maker, a pool whose reserves determine its quoted price. It defines trader archetypes through their rewards, then analyzes how their order flow feeds back into prices. Mathematical results are checked with simulations; this is not a new field experiment showing that deployed AI agents cause market instability.

In the simplest model, each return combines independent external news with trading pressure determined by the previous return’s sign. Herd-followers trade with that sign, while negative-feedback traders trade against it. The parameter β measures the balance between those forces; s sets the scale of external news.

With symmetric news and equally likely initial return signs, markets with feedback +β and −β have exactly the same one-period return distribution. Pure herding and pure countertrend populations of equal size also generate the same gross turnover. Consequently, volatility, value-at-risk and the frequency of losses below a chosen threshold cannot distinguish these paired markets within the model.

Their loss sequences nevertheless differ. The paper defines stress as a return below a specified negative threshold, and a stress episode as consecutive stressed periods. Positive feedback lengthens those episodes; negative feedback shortens them.

For Gaussian news, a threshold of zero and feedback equal in magnitude to the news scale, expected negative-return runs last about 6.3 periods under herding, versus 1.19 under negative feedback. These are model calculations, not measured crash durations. The theoretical ratio can grow without bound as feedback strengthens relative to noise.

Reading the order of returns

The proposed remedy is to examine serial statistics: measurements that preserve information about successive observations.

In the basic model, the sign of first-order return autocorrelation identifies the feedback direction. A two-period variance ratio, which compares two-period return variability with one-period variability, carries the same information when the required variance exists.

An even simpler test counts how often consecutive returns share a sign. Under Gaussian news and feedback no stronger than the news scale, the observations needed to identify the feedback sign scale as (s/|β|)² times log(1/δ), where δ is the allowed error probability. A matching lower bound shows that observing full return values cannot improve that scaling beyond constant factors. Weak feedback therefore requires more data to detect.

Why profit-seeking traders may join a herd

The paper next lets profit-seeking traders choose whether to ride an existing herd, oppose it or stay out. Positions open and close in successive batches, and traders pay a round-trip fee.

Under the assumed execution rules, their own flow’s direct contribution cancels from expected round-trip payoffs. Profits instead come from trading ahead of predictable herd flow. Without an external price anchor, opposing the herd earns a negative expected payoff after fees.

Riding becomes more attractive as more traders do it. Depending on fees and initial participation, the model can therefore tip toward no participation or full participation, with an unstable threshold between them. Full participation lengthens negative-return runs. This is a result about the specified trading game and learning dynamics, not a universal claim that rational trading destabilizes markets.

Liquidity helps, but can disappear

Adding market makers changes the incentives. These providers receive a spread for absorbing herd flow rather than paying to trade. Their activity offsets riders’ influence and shortens modeled stress episodes, although it does not alter the persistence probability once a trend is established.

Under the paper’s stated parameter conditions, rising rider capital eventually makes full liquidity provision unprofitable. Market makers withdraw, producing an abrupt increase in expected loss-run length. Recovery requires rider capital to fall below a lower threshold: a form of hysteresis, where the outcome depends on the market’s recent state.

Maker rebates extend the modeled range over which liquidity can survive or recover. Uniform fees can deter riders, but cannot make opposing the herd profitable for fee-paying traders.

The limitations are consequential. Exact equality of return distributions relies on linearized price impact; the paper reports approximate equality with exact constant-product impact. Independent news also excludes volatility clustering. A companion cryptocurrency study discussed in the preprint exhibits such clustering, so its observed stress persistence does not directly validate this mechanism. The practical implication is a hypothesis for conditional testing—not a demonstrated regulatory prescription for live markets.