When Quotes Crumble: Detecting Transient Mechanical Liquidity Erosion in Limit Order Books
New neural model spots mechanical liquidity erosion before it hits your portfolio.
A team of researchers—including Haohan Xu, Jason Bohne, Pawel Polak, Yurij Baransky, Ajay Alva, Violetta Fedotova, Gary Kazantsev, and David Rosenberg—has published a paper accepted at ICLR 2026's Workshop on Advances in Financial AI. They tackle the critical market microstructure problem of detecting "crumbling quotes," where observable quote deterioration may stem from either mechanical liquidity withdrawal (e.g., a market maker rebalancing) or informational repricing (e.g., new fundamental news). The distinction is vital for traders and exchanges, as mechanical erosion can trigger cascading liquidity crises if misinterpreted.
The team built a multi-agent environment using the ABIDES agent-based simulator, where crumbling emerges from stochastic regime switches in a market maker. This provides time-resolved ground truth—a luxury unavailable in real market data. Their detection pipeline extracts order book features and trains a neural model to output calibrated crumbling probabilities. In experiments, the neural model achieved a +36% AUC improvement over rule-based baselines and maintained robust performance across normal, high-volatility, bull, and bear market conditions. Ablation studies confirmed generalization across both independent and autocorrelated liquidity withdrawal dynamics, making this a practical tool for real-time market monitoring.
- Neural model achieves +36% AUC improvement over rule-based baselines for detecting crumbling quotes.
- Uses ABIDES agent-based simulator to generate ground truth with stochastic regime switches in market makers.
- Robust performance across normal, high-volatility, bull, and bear market conditions.
Why It Matters
This AI gives traders and exchanges a reliable way to spot mechanical liquidity erosion, preventing panic and systemic risk.