Using an offline language model to build Bitcoin flash-crash simulations

Turning market observations into trader behavior

Studying a flash crash requires more than reproducing a sudden price drop. A simulator must also represent how trading pressure, available liquidity and different trading strategies interact. Choosing those behavioral settings is difficult: hand-crafted assumptions may be poorly matched to the market, while detailed historical data can be difficult to obtain.

In an arXiv preprint, Tran Phuong Linh Pham and colleagues investigate an intermediate approach. They use a language model to propose distributions of trader parameters, constrain those proposals with empirical measurements, and then test the resulting simulated markets. The model supplies inputs before simulation; it does not trade or respond to events during a run.

The findings concern this constructed environment. They do not establish what causes real Bitcoin crashes or demonstrate an effective real-world risk intervention.

Historical anchors, not unrestricted role-play

The researchers assemble 66 crash-like episodes in BTCUSDT perpetual futures from June 2020 through December 2024, using a 100-millisecond grid. Perpetual futures have no expiry date and allow leveraged positions.

The historical catalog is not composed entirely of events satisfying the paper’s full flash-crash definition. That definition requires a decline of at least 3% within five minutes followed by recovery of at least half the drawdown. Of the 66 episodes, 43 meet the recovery criterion; 23 steep trend drops remain for behavioral baseline coverage.

The team summarizes market conditions across pre-crash, drop, recovery and post-event phases. These measurements include signed trading flow, volatility, bid–ask spread and a price-impact estimate.

An offline Mistral-7B model receives these empirical anchors and proposes parameters for four trader types: momentum traders, contrarians, market makers and noise traders. Proposed settings cover behaviors such as order aggressiveness, cancellation and order size. Market makers also receive thresholds governing withdrawal during volatility. The noise-trader arrival rate comes directly from empirical trading intensity rather than language-model elicitation.

Responses that fail predefined validity checks are discarded. The remaining proposals are fitted into parameter distributions and frozen before simulation begins. This makes the language model a source of constrained behavioral assumptions, not evidence that actual traders behave that way.

More crashes does not mean complete realism

Each simulation contains 100 agents operating on a 100-millisecond clock. Momentum traders follow trends, contrarians buy against sufficiently deep declines, market makers provide passive quotes, and noise traders trade randomly. The simulator explicitly builds in market-maker withdrawal during elevated volatility and recovery driven by contrarian buying.

In the reported 500-run language-model condition, 88 runs qualify as crashes: a rate of 17.6%. The broad uniform-prior baseline produces a rate of 0.8%, while informed, hand-crafted literature priors produce 15.2%.

These simulated crash labels use a different detector from the historical catalog: a decline of at least 1.93% within one second. The reported incidence therefore should not be read as an estimate of Bitcoin’s real-world crash probability. Evaluation uses a target incidence band of 5%–40%, not a single historical frequency.

The language-model condition also produces larger mean short-horizon drawdowns than the broad baseline. However, it does not dominate the literature-based alternative. Its return-tail statistic is about 29, versus about 77 for literature priors and roughly 8,582 for the historical reference. That historical benchmark is itself conditioned on crash episodes, whereas the simulated statistic pools crashing and non-crashing runs. The comparison is therefore asymmetric, but the remaining tail-fidelity gap is substantial. Simulated directional flow during the drop is also too strongly negative.

A leverage-proxy result, not a policy experiment

The researchers apply DAGMA and DirectLiNGAM to summaries of the simulated runs to examine whether graph-learning methods recover embedded mechanisms. Recovery is partial: DAGMA retains a leverage-to-crash edge, while DirectLiNGAM places greater emphasis on wealth summaries.

Here, “leverage” is a proxy constructed from price dynamics, not a direct measurement of traders’ borrowed exposure or liquidation activity.

A fitted logistic regression provides the mitigation comparison. The researchers shift one predictor downward by one within-panel standard deviation, with a floor at the non-crash median, while keeping the fitted model and other predictors fixed. Shifting the leverage proxy lowers the average predicted crash rate from 17.60% to 13.58%; shifting directional pressure lowers it only to 17.43%.

Those are local sensitivity calculations, not new simulation runs or interventions propagated through the full causal graph.

The approach remains limited by simplified agents, proxy measurements and elicitation with one model at one temperature. Important robustness checks remain unfinished, and direct liquidation feeds and structured news inputs are absent. The supported conclusion is narrower than forecasting crashes: empirically constrained language-model priors can improve this simulator over a broad random baseline, without establishing behavioral truth or real-market causality.