Abstract: This study explores the short-term predictability of price jumps in modern electronic markets. Using high-frequency data from the French CAC40 index, we demonstrate that the "shape" of the limit order book (LOB)—specifically the liquidity imbalance between buyers and sellers—serves as a robust predictor for immediate price movements. Through logistic regression and LASSO variable selection, we achieve a median AUC of 0.70, identifying best-quote liquidity and trade signs as the most critical predictive signals.
Introduction: The Microstructure of Price Jumps
In traditional financial modeling, stock prices are often treated as continuous processes. However, in high-frequency electronic markets, prices move discretely on a tick-by-tick basis. These movements, or "jumps," are triggered by specific microstructural events: limit order arrivals, cancellations, and market orders.
Predicting the direction of the next price change is a primary objective for market makers and high-frequency traders. This research focuses on two types of observable jumps:
- Inter-trade price jumps: Changes in the best bid or ask price following a market order.
- Trade-throughs: Large market orders that exhaust all available liquidity at the best price level, forcing an immediate shift to the next level.
Methodology and Feature Engineering
The prediction task is framed as a supervised learning classification problem. The goal is to estimate the probability of a price jump given the current state of the LOB.
Feature Extraction
We extract 76 features from the LOB, including:
- Log volumes at the first five price levels ($V^b_i, V^a_i$).
- Price gaps between levels and the bid-ask spread.
- Binary "dummy variables" representing the nature of the most recent event (e.g., a limit order on the bid side).
Model Selection
We employ Logistic Regression combined with LASSO (Least Absolute Shrinkage and Selection Operator). LASSO is crucial for variable selection; it penalizes model complexity, effectively zeroing out non-informative features. This allows us to identify the core drivers of price dynamics among the dozens of potential predictors. To account for intraday volatility patterns, morning and afternoon sessions were analyzed independently.
Data and Scope
The study utilizes tick-by-tick data provided by NATIXIS via the Thomson Reuters "Tick Capture Engine." The dataset covers the 40 constituent stocks of the French CAC40 index for the month of April 2011. The granularity includes every transaction and LOB modification at the millisecond level, specifically tracking the five best limits on both the bid and ask sides.
Key Insights
- The Liquidity Balance Signal: A strong imbalance between bid and ask volumes at the best quotes is a primary indicator of the next trade's direction. High volume at the best bid relative to the best ask significantly increases the probability of a subsequent buy order.
- Predictability is Time-Invariant: Despite varying trade volumes throughout the day, the model's performance remained consistent across morning and afternoon sessions. This suggests that the fundamental mechanics of LOB-driven price moves are stable.
- Level 1 Dominance: LASSO selection consistently highlighted that the most informative variables are the liquidity at Level 1 (the best bid/ask), the size of the most recent market order, and the direction of the previous trade. Deeper levels (2-5) offer diminishing marginal utility for short-term prediction.
Results and Performance
The models achieved a median AUC (Area Under the Curve) of approximately 0.70 across the CAC40 stocks. In financial forecasting, an AUC of 0.70 represents a highly significant and potentially exploitable signal, far outperforming random chance (0.50).
The most consistent predictors for bid-side jumps were the log volume at the best bid ($VB1\_0$) and the binary indicator of a market order at the bid ($BMO\_0$). Interestingly, while trade-throughs cause immediate price changes, they were found to be less informative for predicting future jumps than simple volume imbalances at the best quotes.
Conclusion and Future Directions
This research confirms that short-term price jumps are predictable using simple LOB characteristics. While this study focused on linear models and a one-step-ahead horizon, future work could explore non-linear relationships or the impact of hidden liquidity (iceberg orders). Nevertheless, the stability of the Level 1 liquidity signal across diverse stocks provides a foundational insight into market microstructure.
Summary updated