Surrogate Modeling of Exchange Matching Engines

Exchange matching engines—the computers that match buy and sell orders and execute trades—are opaque. Traders submit orders, and the exchange determines executions according to its rules, but the exact mechanics are proprietary. Understanding how the matching engine will behave in response to your order submission is valuable. Surrogate modeling attempts to reverse-engineer this behavior using machine learning.

Why Model the Matching Engine?

Understanding matching-engine behavior helps with several decisions:

• Order placement: Where should I place my order (which price level, which size) to maximize execution probability?
• Queue position: How will other orders' arrivals and departures affect my queue position?
• Execution timing: If I cancel and resubmit, will I get a better queue position?
• Partial execution: How much of my order will execute versus remain on the book?

A surrogate model (learned from observation) can answer these questions without access to exchange internals.

Data and Feature Engineering

Building a surrogate model requires historical order-by-order data. The training set includes:

• Submitted orders: price, size, type (limit vs market, buy vs sell)
• Current order-book state: volumes at each price level before the order
• Outcomes: whether the order executes, partially or fully, and at what price

Feature engineering extracts predictive signals from this data. Examples include:

• Order's position relative to current best bid/ask
• Relative order size compared to visible book
• Recent order-flow intensity and direction
• Time since last trade at this price level
• Queue depth ahead of this order

Model Architecture

Classification models predict execution (yes/no). Regression models predict execution size or price. More sophisticated approaches use:

• Gradient boosting (XGBoost, LightGBM) for feature importance and robustness
• Deep learning (RNNs, Transformers) for temporal dynamics in order book
• Probabilistic models (Bayesian approaches) to quantify uncertainty

A multi-task learning approach can simultaneously predict multiple outcomes (execution, size, price) while sharing learned representations.

Behavioral Patterns the Model Captures

Learned surrogate models naturally capture several real matching-engine behaviors:

• Price-time priority: Orders at better prices execute first; among orders at same price, earlier arrivals execute first
• Pro-rata allocation: Some exchanges pro-rata allocate when multiple orders compete for the same execution quantity
• Order-type effects: Market orders (execute immediately) behave differently than limit orders
• Partial execution: Incoming orders that cross multiple prices experience step-like execution
• Queue dynamics: Orders' queue position changes as ahead orders execute or cancel

Exchange-Specific Models

Each exchange has different matching rules. A model trained on NASDAQ data will not transfer directly to NYSE. Practitioners typically maintain separate surrogate models for each exchange and venue. Transfer learning can partially address this—training on one exchange then fine-tuning on another reduces data requirements.

Validation and Performance Metrics

Surrogate models must be evaluated carefully. Standard classification metrics (accuracy, precision, recall) may not capture trading-relevant performance. For example, if 95% of orders execute, a naive model that always predicts "execute" achieves 95% accuracy but is useless.

Better metrics are those aligned with trading: prediction accuracy on execution boundary (orders near the margin of execution), calibration of predicted probabilities, impact on trading performance if using the model's predictions.

Causal vs Correlational Prediction

A surrogate model captures correlations in historical data, not causal relationships. For example, if large orders at market open are unlikely to execute at aggressive prices, the model learns this correlation. But if the model is applied to your own order submission, the causality might differ: your order's large size might be obvious to the exchange (if it's known to be from you) or might be hidden (if submitted anonymously).

This causality gap means surrogate models require careful validation on new data before deployment.

Practical Applications

Deployed surrogate models enable:

• Smart order placement algorithms that choose prices maximizing execution probability
• Real-time order-book prediction and liquidity estimation
• Backtesting simulators that more accurately estimate execution outcomes

Conclusion

Surrogate modeling reverse-engineers opaque exchange matching engines using machine learning. While not perfect—models capture correlations rather than causal mechanisms—they provide valuable insights into execution behavior and enable more intelligent order submission strategies.