PREreview del Agentic AI in Automated Markets

The paper asks a good question. Which trader behaviours stabilise automated markets, and which destabilise them? Aggregate data cannot answer it. The authors propose a method, not a measurement. Encode each behaviour as a reward function, let reinforcement learning find the policy, and the behavioural assumptions become explicit and auditable.

Strengths

1. Behaviour as a reward function - interesting idea. Put the assumption in the objective instead of in hand-coded rules, and a reader can audit the behavioural claim without reading the implementation.

1. Interesting approach with usage of the dual-algorithm comparison.

2. The framework is portable. Price, volume and a liquidity proxy is a minimal input set, and it could travel across venues and asset classes.

3. The paper is candid about its status. Section 5.1 notes that a strategy-level Sharpe measures how well an agent exploits the market, not how efficient the market is.

Weaknesses

W1. Could not fnd H1–H4 estimates. No regression coefficient, standard error, sample size or variance-ratio statistic appears anywhere. The Results section is prose plus three figures. Would be good to view a results table, or relabel the section as a design proposal until the estimates exist.

W2. The identification is circular. The policy is trained on a series and then applied to the same series to produce the intensity measure. No out-of-sample or walk-forward rule is specified. The manuscript also refers twice to a "Stage 2" that links intensity to market outcomes, and no such analysis is presented. Specify a walk-forward protocol (fit on window t, measure on window t+1) and either report Stage 2 or reframe the paper as a design proposal.

Questions

1. What are the values of κ_turn, κ_mom, λ_stress, c and the trailing-volatility window? They would allow me sanity-check the results.

2. Would you release the environment and the three processed series? A small, well-documented archetype environment would make easier to reproduce.

Closing note

I would be glad to review a revision. The idea is worth the work, the hypothesis table gives the revision a clear target, and most of what I ask for is reporting, not new modelling. I would read the revision with interest.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they used generative AI to come up with new ideas for their review.

Text format and structure