The detection of "crumbling quotes"—transient erosion of bid-ask spreads driven by mechanical liquidity withdrawal rather than fundamental repricing—represents a critical challenge in market microstructure research. Unlike informational events that reflect genuine price discovery, mechanical liquidity withdrawal occurs when market makers systematically reduce order book depth, typically in response to inventory imbalances or risk management triggers. This distinction carries profound implications: algorithms designed to detect informational signals will systematically misfire during mechanical erosion events, leading to costly false positives in execution strategies and market-making models. The core difficulty lies in the fundamental observability problem—real market data provides no ground truth for the causal mechanism underlying quote changes, making supervised learning approaches historically intractable.
This work circumvents the ground truth problem through an elegant methodological choice: leveraging the ABIDES (Agent-Based Interactive Discrete Event Simulation) framework to construct synthetic market environments where the causal mechanisms generating quote erosion are fully observable and controllable. By instantiating stochastic regime-switching dynamics in simulated market makers—where liquidity withdrawal emerges from well-specified probabilistic transitions—the researchers obtain time-resolved labels for crumbling events unavailable in empirical market data. This simulation-based supervision strategy enables training of neural models on features extracted directly from limit order book microstructure.
The technical pipeline operates in two stages. First, a feature engineering layer extracts order book observables—spread dynamics, depth profiles, order flow imbalances, and temporal patterns—that correlate with mechanical versus informational quote changes. The authors then train a neural classification model to map these features to calibrated crumbling probabilities, optimizing for both discrimination (AUC maximization) and calibration quality. Notably, the framework achieves a 36% relative AUC improvement over rule-based baselines, suggesting that neural models capture non-linear patterns in liquidity erosion that simple threshold-based heuristics miss. The calibration objective is particularly important: practitioners require probability estimates that reflect true event likelihood, not merely discriminative rankings.
A critical strength of the evaluation lies in its systematic stress testing across market regimes. The authors validate their framework under normal volatility conditions, high-volatility episodes, bull markets, and bear markets—each presenting distinct microstructure signatures. This cross-regime robustness is non-trivial: many market microstructure phenomena exhibit regime-dependent characteristics that cause models trained on one market state to degrade substantially under different conditions. Ablation studies further strengthen the contribution by isolating the contribution of temporal features and testing generalization across both independent and autocorrelated liquidity withdrawal dynamics, confirming that the learned representations capture fundamental structural patterns rather than overfitting to specific simulation parameters.
Within the broader landscape of market microstructure research, this work addresses a longstanding friction between simulation-based research and empirical validation. The microstructure literature has historically bifurcated into two communities: those using agent-based models to study mechanism design and market design questions (with perfect ground truth but limited ecological validity), and those analyzing real market data (with ecological validity but fundamental unobservability of causal mechanisms). This paper occupies a productive middle ground: using simulation to obtain ground truth for training, then implicitly testing transferability by evaluating whether features learned from synthetic data generalize across diverse market conditions. The implicit assumption—that order book microstructure patterns are sufficiently stable across synthetic and real markets—deserves scrutiny, but the cross-regime validation provides some evidence for this claim.
CuraFeed Take: This research has immediate practical implications for execution algorithms, market-making systems, and market surveillance. The 36% AUC improvement suggests that neural models trained on simulation-derived labels could materially improve real-world detection of mechanical liquidity erosion, reducing false-positive signals in execution strategies. However, the critical question for practitioners is sim-to-real transfer: will models trained on ABIDES synthetic data maintain their performance when deployed against actual market microstructure? The cross-regime validation is encouraging but insufficient—real markets contain edge cases, exotic order types, and latency effects absent from simulation. The authors should prioritize empirical validation on real market data (even without ground truth, one could use proxy labels from alternative detection methods or expert annotation) to establish whether the simulation-derived training signal transfers meaningfully. For the research community, this work exemplifies how agent-based simulation can overcome fundamental observability constraints in market microstructure, potentially unlocking supervised learning approaches to other causally-defined phenomena (information leakage, predatory trading patterns, etc.) where ground truth is unavailable in real data. Watch for follow-up work attempting sim-to-real transfer and for adoption of similar simulation-based supervision strategies in other microstructure problems.
