Behavioral Prior Elicitation via LLM for
Heterogeneous-Agent Bitcoin Flash-Crash Simulation

Overview

Why Bitcoin Flash Crashes Are Hard to Simulate

Bitcoin perpetual futures occupy a structurally distinct position in modern financial markets. They trade continuously, carry no expiry date, support persistent leveraged positions, and exhibit order book dynamics unlike any equity or dated-futures venue. The result: concentrated liquidity near the best quotes, shallow depth beyond the top levels, and elevated price impact — all of which make these markets especially vulnerable to millisecond-scale flash crashes.

Four-Stage Pipeline

Figure 1. Overall pipeline. Empirical BTCUSDT flash-crash events feed phase-level microstructure summaries into an offline LLM (Mistral-7B). The validated priors parameterise a heterogeneous LOB simulator, and the resulting 500-run synthetic panel is analysed with DAGMA and DirectLiNGAM for mechanism verification and counterfactual sensitivity.

Agent Archetypes

Each simulation run instantiates 100 agents drawn from four archetypes. Their behavioral parameters are sampled from phase-conditional LLM-elicited priors, not fixed values.

Flash-Crash Detection (Simulation)

A simulated run is classified as a crash if price falls ≥1.93% within a 10-step (1-second) window — calibrated from the 10th percentile of the rolling-10-tick drawdown distribution across the empirical 66-event panel. This threshold is distinct from the historical detector (≥3% in 5 minutes) used to build the event corpus.

Temporal Anatomy of a Simulated Crash

The locked LLM panel does not generate isolated price jumps. Instead, crash events emerge through a coordinated mechanical sequence: directional pressure turns sharply negative, quoted trading cost rises as book balance deteriorates, trading activity accelerates, and price then falls in a concentrated burst with only partial early recovery.

Figure 2. Temporal anatomy of the locked LLM crash panel, aligned at the first detected crash. Price (top), directional pressure (OFI), quoted trading cost (spread), book balance, and trading activity co-move in the expected crash sequence. This should be interpreted as correct sequencing and co-movement — not perfect amplitude calibration against the empirical reference.

Three-Way Locked Comparison

The LLM condition is evaluated against two locked baselines on the same frozen 500-run panel: a broad uniform prior (uninformative) and a literature prior (hand-crafted values from prior simulation studies). The LLM condition is considered successful if it dominates the broad baseline and remains non-inferior to the literature baseline on crash incidence and mean short-horizon drawdown.

The LLM condition clearly improves over the broad-prior baseline on crash incidence and mean drawdown, and remains competitive with the literature baseline. It does not reproduce the full tail magnitude of the crash-conditioned BTC reference panel (a 295× kurtosis gap), nor does it perfectly match drop-phase directional flow — both expected limitations of a linear price-impact rule with static within-run sampling. The appropriate read is competitiveness, not uniform superiority.

Recovering Crash Mechanisms from the Synthetic Panel

The run-level panel Z ∈ ℝ^500×p summarises the pre-crash state for each of the 500 simulation realizations. Two complementary graph-learning methods test whether standard algorithms recover the embedded crash pathways from cross-run variation. This is a simulator-internal mechanism check, not a causal claim about real Bitcoin markets.

DAGMA retains a direct edge from pre-crash leverage into the crash label, while DirectLiNGAM assigns the most stable crash-adjacent role to wealth-level summaries. Directional pressure and book imbalance are strongly separated across crash and non-crash runs but do not recover as uniformly oriented crash edges once the panel is no longer label-conditioned by construction. The most defensible conclusion is partial structural consistency: leverage emerges as the most stable simulator-internal amplifier signal.

Which Lever Matters More — Directional Pressure or Leverage?

Using the fitted logistic model and the frozen run-level panel, we estimate local, support-preserving counterfactual sensitivity to reductions in two pre-crash channels. Each intervention reduces the target feature by one within-panel standard deviation, floored at the non-crash median so the intervention stays on the data manifold.

Figure 3. Pre-crash separation and local mitigation effects on the symmetric trailing-window LLM panel. Directional pressure is a strong warning signal — it separates crash from non-crash runs clearly — but leverage is the cleaner local amplifier: reducing it by one standard deviation produces a meaningful crash-rate reduction, while the same reduction in directional pressure has negligible effect.

Summary and Limitations

This paper presents a simulation-first framework that transforms empirical BTCUSDT flash-crash evidence into behaviorally grounded agent priors, validates them against microstructure anchors, instantiates them in a heterogeneous limit order book simulator, and studies crash-amplification mechanisms through structural verification and local counterfactual sensitivity analysis.

The central finding is not that LLM outputs are behaviourally true, but that LLM-elicited priors become useful simulator inputs when constrained by empirical anchors and evaluated against explicit baselines. The LLM condition attains a crash rate of 0.176, well inside the 0.05–0.40 target band, while the uninformative broad baseline produces only 0.008 — a 22× gap. Leverage mitigation yields a −22.83% crash-rate reduction; directional pressure mitigation yields only −0.98%. Within the simulator, leverage amplification therefore emerges as the more consequential local crash-amplification channel.

Known Limitations