When Quotes Crumble: Detecting Transient Mechanical Liquidity Erosion in Limit Order Books

Abstract: We study the detection of transient liquidity erosion ("crumbling quotes") in electronic limit order books, where observable quote deterioration may reflect either mechanical liquidity withdrawal or informational repricing. Using the ABIDES agent-based simulator, we construct a multi-agent environment in which crumbling emerges from stochastic regime switches in a market maker, providing time-resolved ground truth unavailable in real market data. We develop a detection pipeline that identifies mechanically driven quote erosion using order book features, and train a neural model to produce calibrated crumbling probabilities. Experiments demonstrate that the proposed framework reliably identifies crumbling events against agent-level ground truth, with the neural model achieving +36% AUC improvement over rule-based baselines and robust performance across normal, high-volatility, bull, and bear market conditions. Ablation studies on temporal features and varying the dependence structure of the ground-truth mechanism confirm that the framework generalizes across both independent and autocorrelated liquidity withdrawal dynamics.

• The paper introduces a novel simulation-based framework to label and detect mechanical liquidity erosion in limit order books.
• It employs a mechanics-gated neural MLP augmented with temporal context features to achieve higher ROC-AUC compared to traditional methods.
• The methodology offers practical insights for adaptive execution algorithms and market microstructure research in volatile trading environments.

Detecting Transient Mechanical Liquidity Erosion in Limit Order Books

Problem Formulation and Motivation

The paper "When Quotes Crumble: Detecting Transient Mechanical Liquidity Erosion in Limit Order Books" (2604.21993) addresses the fundamental ambiguity in interpreting limit order book (LOB) quote deterioration. In fragmented and fast electronic markets, displayed depth can vanish through distinct mechanisms: either via transient, mechanical liquidity withdrawal (referred to as "crumbling quotes") or via persistent, information-driven repricing. Both scenarios result in the removal of liquidity at the best quotes, complicating the differentiation between adverse conditions for execution algorithms and typical price discovery activity.

Traditional LOB data and message feeds do not disclose the underlying cause of quote changes, and thus LOB-only detection and labeling rules are vulnerable to significant error rates that are difficult to audit. The absence of mechanism-level ground truth in market data impedes robust statistical research on liquidity regimes, complicates reinforcement learning for trading, and limits practical market design interventions.

Figure 1: Schematic illustration highlighting the observational ambiguity in LOB quote movements, the agent-based simulation environment generating ground-truth mechanical crumbling, and the mechanics-constrained detection and probabilistic labeling pipeline.

Simulation-based Ground Truth and Methodological Framework

To systematically resolve the ambiguity inherent in LOB streams, the authors construct a ground-truth labeling regime using the ABIDES agent-based simulation platform. The key innovation is a regime-switching market maker whose side-specific quoting behavior is parameterized via a stochastic skewness process, producing explicit, temporally-resolved intervals of deliberate liquidity withdrawal. This design establishes a set of labeled crumbling regimes (via observable internal agent states) in an otherwise realistic, latency-stratified market environment populated by heterogeneous zero-intelligence and strategic agents.

The detection pipeline operates solely on observable LOB data, excluding any privileged simulation knowledge. The event construction proceeds in two stages:

1. Candidate Generation: Best-quote deterioration steps are identified based on tick-wise movements accompanied by visible queue depletion, subject to minimal replenishment.
2. Mechanics-Consistent Filtering: Candidate events are filtered by mechanically interpretable criteria: book consistency (accounting for visible removals and additions), efficient price stability, one-sidedness (excluding two-sided information moves), and price transience (requiring mean reversion after the event). These hard constraints yield a set of events that are consistent with purely mechanical liquidity erosion.

Event-level feature vectors are derived, summarizing severity (walk depth, depletion speed), refill and spread dynamics, efficient-price displacement, and impact decay. These features support both interpretable rule-based labeling and subsequent continuous labeling with neural models.

Neural Continuous Labeling and Temporal Contextualization

Recognizing the inadequacy of binary rules for calibrating uncertainty and interacting features, the authors introduce a mechanics-gated neural probabilistic labeling function. The model employs a feedforward MLP encoder augmented with temporal context features designed to capture regime persistence:

• Recovery interval ($\Delta t_{\mathrm{rec}}$): Time since the previous event, a shorter interval indicates persistent liquidity stress.
• Recent event count ($n_{\mathrm{rec}}$): Number of events within a lookback window, encoding clustering.
• Cumulative depletion ($\Sigma_{\mathrm{DS}}$): Aggregate order flow imbalance.

The output is multiplicatively gated by the hard mechanics filter, ensuring that soft probabilistic scores are only assigned to mechanically eligible events. The probabilistic targets are KL-calibrated to regime-labeled ground truth intervals via cross-entropy minimization.

Experimental Results and Empirical Validation

The experiments, performed on five simulated equity trading days, compare the discrimination ability of three methods: a rule-based binary detector, a logistic regression classifier, and the mechanics-gated neural MLP. Evaluation is conducted across baseline, bull, bear, and high-volatility regimes, using ROC-AUC as the metric.

• In the baseline regime, the rule achieves an AUC of 0.67, logistic regression 0.84, while the neural model attains 0.91.
• This strong numerical improvement persists across bull, bear, and high-volatility settings, with the neural model demonstrating robust performance under regime drift and order flow shocks.

Figure 2: ROC curves for crumbling detection across multiple market regimes, illustrating the performance gap between the rule-based, logistic regression, and MLP models.

The generalization capacity of the neural model is tested by inducing crumbling regime clustering via a Hawkes process in the market maker’s intervention schedule. The temporal context features enable the model to maintain high discrimination, accurately tracking both isolated and clustered crumbling events.

Figure 3: Mid-price series alongside event-level crumbling detections and continuous neural probabilities, displaying precise temporal correspondence with ground-truth regime switches.

Ablation studies show that the inclusion of temporal context features—specifically recovery interval, event count, and cumulative depletion—substantially enhances classification, especially in the presence of persistent liquidity withdrawal regimes.

Figure 4: ROC curves with and without temporal context features, demonstrating the incremental value of temporal augmentation in the MLP architecture.

Theoretical and Practical Implications

The formalization and detection of mechanically driven liquidity erosion have significant implications for both AI-driven trading systems and microstructure theory:

• Execution Algorithms: The framework provides execution algorithms and RL agents with calibrated, probabilistic signals for adverse microstructural conditions—enabling adaptive reaction to transient liquidity crises not detectable via conventional metrics.
• Market Microstructure Research: The methodology enables systematic measurement and differentiation of mechanical liquidity withdrawal versus information-driven price discovery at high frequency, supporting empirical studies on market impact and latency-centric phenomena.
• Market Design and Regulation: The mechanics-consistent, auditable labeling regime can inform policy evaluation for interventions (e.g., speed bumps, protective order types) designed to mitigate transient liquidity disruptions and adverse selection.
• Generalization: The robustness of the framework to both memoryless and self-exciting (Hawkes) crumbling regimes suggests its applicability to diverse market environments, including those exhibiting clustering dynamics relevant in stress events.

Conclusion

This work establishes a new empirical and methodological standard for the detection and quantification of transient mechanical liquidity withdrawal in limit order books. By providing a regime-agnostic, interpretable, and data-efficient detection pipeline validated against auditable ground truth in agent-based simulation, it advances both the practical monitoring of execution risk and the broader microstructure modeling literature. The robust gains in discrimination, especially in volatile and clustered regimes, underscore the importance of temporal aggregation and mechanics-first feature design. The research paves the way for future developments in adaptive execution, microstructural monitoring, and synthetic market experimentation supporting both academic inquiry and industrial deployment.