🤖 machine learning

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

This paper proposes a detection framework using an agent-based simulator and a neural model to distinguish between mechanical liquidity withdrawal ("crumbling quotes") and informational repricing in electronic limit order books, achieving significant performance improvements over rule-based methods.

Original authors: Haohan Xu, Jason Bohne, Pawel Polak, Yurij Baransky, Ajay Alva, Violetta Fedotova, Gary Kazantsev, David Rosenberg

Published 2026-04-27

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Haohan Xu, Jason Bohne, Pawel Polak, Yurij Baransky, Ajay Alva, Violetta Fedotova, Gary Kazantsev, David Rosenberg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a busy auction for rare trading cards. Most of the time, the prices move because people are actually deciding the cards are worth more or less (this is "Informational Repricing").

But occasionally, something weird happens: the person acting as the "bookkeeper" (the Market Maker) suddenly pulls their offers off the table for a few seconds because they are overwhelmed or adjusting their strategy. The price "drops" or "jumps," but it’s not because the cards changed value—it’s just because the person holding the cards momentarily stepped away from the table. This sudden, temporary disappearance of support is what the researchers call "Crumbling Quotes."

The problem? To an outsider, a price drop caused by a "stepping away" looks exactly like a price drop caused by "bad news." If you are a trader and you see the price drop, you don't know if you should panic (because the value is gone) or wait (because the bookkeeper is just coming back).

Here is how the researchers solved this:

1. The "Flight Simulator" (ABIDES)

In the real world, we can't see the "mind" of the bookkeeper. We only see the prices. To solve this, the researchers built a high-tech "flight simulator" for markets called ABIDES.

In this simulator, they play the role of God. They create digital "agents" (traders) and specifically program one agent to "crumble"—to intentionally pull its liquidity away for a set amount of time. Because the researchers are running the simulation, they have the "Ground Truth": they know exactly when the crumbling is happening, even if it looks messy on the surface.

2. The "Sieve" (The Detection Pipeline)

The researchers developed a way to filter out the "fake" price moves. Think of it like a sophisticated sieve used to separate gold from sand:

The "Is it real?" test: They check if the price move was caused by people actually buying/selling (exhausting the supply) or just someone canceling an order.
The "Will it come back?" test: If the price drops but then immediately bounces back to where it was, it’s likely a "crumble" (transient). If the price drops and stays down, it’s likely "real news" (persistent).
The "One-sided" test: They check if the movement is happening on only one side of the market, which is a hallmark of a mechanical error rather than a market-wide shift in value.

3. The "Smart Detective" (The Neural Model)

Even after the sieve, some events are "maybe" crumbles and some are "definitely" crumbles. The researchers trained an AI (a Neural Network) to act like a seasoned detective.

Instead of just saying "Yes" or "No," the AI gives a probability score. It looks at the "vibe" of the market:

“The last three times this happened, it was a crumble, and it’s happening again very quickly...”
“The speed at which the orders disappeared was suspiciously fast...”

The AI doesn't just look at the single moment; it looks at the rhythm of the market (the "temporal context") to see if the market is in a "stressful" mood.

The Result

The researchers found that their AI "detective" was much better at spotting these moments than traditional, rigid rules. It was 36% more accurate at identifying these "crumbling" episodes.

Why does this matter?
If you are an automated trading system, knowing the difference between "the price is crashing because the company is failing" and "the price is crashing because the bookkeeper is taking a coffee break" can save you millions of dollars. This paper provides the "glasses" that allow traders to see the difference.

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

1. Problem Statement

In electronic limit order books (LOBs), a critical execution risk is "crumbling quotes"—episodes where displayed liquidity at the best bid or ask is removed faster than it can be replenished, causing the price to deteriorate by multiple ticks.

The fundamental challenge is observational ambiguity: a sudden move in the best quote can be caused by either:

Mechanical Liquidity Erosion (Crumbling): A transient withdrawal of depth where the underlying "efficient price" remains stable.
Informational Repricing: A persistent change in the asset's value driven by new information.

Distinguishing between these two is vital for execution algorithms and reinforcement learning agents. If an agent mistakes a transient "crumble" for a permanent price move, it may execute suboptimally. However, in real-world historical data, there is no "ground truth" to tell us which mechanism caused a specific quote movement.

2. Methodology

The authors propose a three-stage framework: Simulation $\rightarrow$ Mechanics-Constrained Detection $\rightarrow$ Neural Probabilistic Labeling.

A. Ground-Truth Generation (ABIDES Simulator):
To solve the lack of ground truth, the authors use the ABIDES agent-based simulator. They augment an adaptive market maker with a stochastic regime-switching mechanism. By manipulating a "skewness parameter" ( $\beta_t$ ), they force the market maker to intentionally withdraw liquidity on one side of the book. Because the simulator knows the agent's internal state, it provides perfect, time-resolved ground truth for when "crumbling" is actually occurring.

B. Mechanics-Constrained Detection Pipeline:
The authors develop a "mechanics-first" pipeline to identify candidate events using only observable LOB data. The process involves:

Visible-Depletion Consistency: Ensuring the price move is explained by the exhaustion of the existing queue (via cancellations or executions) rather than sudden additions.
Hard Filters (The "Gating" Mechanism): To exclude informational repricing, they apply four constraints:
- Book Consistency: Most removed volume must be accounted for.
- Efficient-Price Stability: The "microprice" (a proxy for efficient value) must remain stable during and after the event.
- Opposite-Side Stability: The quotes on the other side of the book should not move significantly.
- Transience: The price must exhibit mean reversion (the excursion must be temporary).

C. Neural Continuous Labeling:
Instead of a simple binary "yes/no" label, the authors train a Multi-Layer Perceptron (MLP) to output a calibrated probability of crumbling.

Features: They use event-level features (walk depth, depletion speed, refill ratio, etc.) augmented with temporal context features ( $\Delta t_{rec}$ , recent event counts, and cumulative depletion) to capture market regime persistence.
Training: The model is trained using a KL-divergence (Binary Cross-Entropy) objective against the simulator's ground truth, but it is constrained by the "hard mechanical gate" to ensure the outputs remain physically interpretable.

3. Key Contributions

Endogenous Ground Truth: A novel way to use agent-based modeling to create a "labeled" dataset for microstructure phenomena that are unobservable in real markets.
Mechanics-First Detection: A rigorous mathematical framework to filter out informational moves from mechanical liquidity withdrawals using LOB-only data.
Continuous Probabilistic Labeling: A neural model that moves beyond binary classification to provide a "severity score," allowing traders to rank the risk of liquidity erosion.

4. Results

Superior Performance: The neural model significantly outperformed rule-based baselines. In baseline simulations, the MLP achieved an AUC of 0.91, a +36% improvement over rule-based methods.
Robustness: The framework maintained high performance across various market regimes, including Bull, Bear, and High-Volatility markets.
Generalization: An ablation study using a Hawkes process (to simulate clustered, self-exciting liquidity withdrawals) showed that the model successfully captured the temporal dependencies of complex market regimes.
Feature Importance: Adding temporal context features (like recovery intervals and recent event counts) was shown to be critical for distinguishing isolated noise from sustained liquidity stress.

5. Significance

This work bridges the gap between theoretical market microstructure and practical machine learning for trading. By providing a method to detect and quantify transient liquidity erosion, it enables the development of more robust execution policies and reinforcement learning agents that can "wait out" mechanical volatility rather than reacting to it as a permanent price change. The methodology of using "mechanics-constrained" neural networks provides a blueprint for applying AI to high-frequency financial data where interpretability and physical consistency are paramount.