A trader can build this paper's long-only crypto exposure rule. The regulatory claim rests on price-taking agents whose trades cannot move the markets they are said to stress.

Three possible positions

Aldridge and her coauthors frame trading as hourly exposure control. The agent chooses sell, hold or buy, moving exposure by -0.5, 0 or +0.5 within [0,1]. Its positions are flat, half long and fully long. It cannot short.

The reward starts with exposure-weighted next-hour return and subtracts a proportional cost c for changing exposure. The resulting net figure is divided by trailing volatility. When the stress flag is on, an invested agent pays an additional penalty.

The authors train three trader types by changing that reward:

Each type gets two training methods: a Deep Q-Network (DQN) and REINFORCE with a learned value baseline. The pilot uses AAVE, ETC and SUI, about 1,750 hours per coin, or roughly 73 days. SHAP attributions identify which inputs each policy uses. For the rational agent, the top three are 7-day drawdown, cross-asset return and time of day. In trading terms, the rule cuts exposure as drawdown grows or other coins turn.

The reported findings are qualitative. Herd-followers have the most frequent stress episodes, though theirs are the shortest. Manipulators have the fewest episodes and the longest. Rational agents have the steadiest rolling Sharpe when they trade; the two active rational agents also have the highest return volatility. The authors use these results to argue for behavior-specific regulation rather than uniform volatility or turnover limits.

Several inputs needed to reproduce the rule are missing. We did not find values for c, κ_turn, κ_mom or λ_stress. The paper says the weights were set for a validation window and refers to a test period in its SHAP section, without giving dates or a length for either. In the version we read, Data and Features has a heading but no text. The stress flag uses a CRSP definition: VWRETD below its 5th percentile or a VIX jump above its 95th. A crypto pilot would need a crypto stress definition. We could not find one.

Which AAVE policy trades?

Only one rational policy trades AAVE at all.

The rational REINFORCE agent settles at zero exposure on AAVE and SUI, two of the three coins. Its Sharpe and volatility plots consequently show flat zero lines. The authors acknowledge this, calling it "a degenerate solution under a strong volatility penalty, not as evidence of rational behavior." They say the outcome motivates a multi-seed protocol. Yet the hyperparameter section says each configuration uses five seeds with averaged results, while the results section says one training run per configuration. The limitations list also says single run, which is what we take to have been done.

On AAVE, then, the active rational strategy is DQN. By the authors' account, it is also among the pilot's highest-volatility strategies. The turnover-rewarded manipulator has lower baseline volatility, although its rolling 24-hour standard deviation spikes to about 0.02 on SUI. The ordering is surprising given the rewards. A turnover-seeking agent alternating between 0 and 0.5 could simply hold less exposure on average; the paper supplies no exposure statistics to check that reading.

We could not trade what the paper trades. Our build applies the exposure rule to AAVE only, using BTC and ETH for cross-asset signals. The paper's ETC and SUI observations, including the ETC reversal losses and SUI spike, do not transfer to that run. We have no results from it yet.

What can a 7-day Sharpe show?

Every agent's rolling 7-day Sharpe moves between roughly -20 and +20 annualized. Each window contains 168 hourly returns. Annualizing a week of AAVE P&L readily gives figures of that size, with little to say about a durable edge. We found no full-sample Sharpe, return or drawdown for any agent. The text and figure captions also report no buy-and-hold Sharpe or volatility.

The authors describe the pilot as preliminary and its findings as "motivating evidence rather than formal tests." Their metrics table defines crisis persistence as "Visual run length." All four hypotheses, H1 to H4, cite the pilot as their source: the same evidence suggests the hypotheses and illustrates them. For H3, concerning variance-ratio inefficiency, we found no variance ratio computed in the pilot.

Time of day appears among the top three SHAP features for all three agents. The 73 days supply 73 intraday cycles. We would wait for the attribution to hold in a later period before treating it as seasonality.

The agent's loss is its own

The stress metric counts hours when the strategy's net return is below -2% within a rolling 24-hour window. It tracks the agent's P&L. On ETC, the herd-follower's weakest rolling Sharpe occurs between hours 300 and 800, when momentum trades meet reversals. The manipulator's stress near hours 1,450 to 1,500 describes a position held through the market's crash regime.

The authors acknowledge the boundary. Their agents are price-takers in separate environments, so the first stage "cannot show how one archetype's trades move prices." They also say a strategy's Sharpe measures how well it exploits the market. Their proposed answer is Stage 2, linking archetype intensity to market outcomes, though they concede that analysis is correlational. They promise placebo and external-validation tests to address the possibility that intensity proxies for size, value or momentum. In the discussion, every recommendation depends on "a hypothesis that the CRSP tests can confirm or reject."

We found no Stage 2 results, CRSP regressions or placebo tests in the paper. Section 7 makes every recommendation conditional on the CRSP tests. The conclusion states a firmer claim: "The findings support regulation that targets herding and manipulation while preserving adaptive risk-taking." The authors call their manipulator a stylized proxy. Its turnover bonus operates without price impact, leaving it unable to spoof or wash anything, yet the regulatory section treats its footprint as manipulation. One proposed tool is fees scaled to order-to-trade ratios. These agents choose among three exposure levels rather than posting orders. The paper assumes trades clear through automated market makers, while the environment models none and excludes market makers as agents.

A reported Stage 2 regression could change our assessment if archetype intensity predicted market crash frequency after the authors' listed size, value, momentum, volatility and turnover controls. Until then, the testable proposition is a long-only AAVE timing rule with unknown costs.