Measuring Information Leakage with Mutual Information Metrics

Execution algorithms attempt to hide their trading intent from the market. However, the sequence of orders submitted, their sizes, timing, and aggressiveness inadvertently reveals information. Informed traders and market makers analyze these signals to infer the algorithm's true goal, then position against it. Measuring information leakage allows algorithms to minimize adverse selection and improve execution.

What Is Information Leakage?

An execution algorithm breaks a large order into smaller tranches to minimize market impact. But this very act of breaking the order into many pieces reveals that a large order exists. Over-time, by observing order flow patterns, the market can infer the total size of the algorithm's parent order.

This inference is valuable to informed traders. If they know a large buyer is executing, they can position ahead of the execution, buying inventory beforehand and profiting as the buyer's demand pushes prices upward. This is "adverse selection"—the algorithm ends up buying at unfavorably high prices.

Information-Theoretic Framework

Mutual information (MI) is an information-theoretic measure of how much knowing one variable tells us about another. In execution context:

MI(order_flow ; parent_order_size) measures how much the observed sequence of orders reveals about the true parent order size. High MI means the order flow strongly reveals the parent order; low MI means revelation is minimal.

The goal of execution algorithms is to minimize this mutual information. An algorithm that submits orders in a pattern uncorrelated with the true parent order size leaks less information.

Estimating Mutual Information

Computing mutual information requires estimating probability distributions: the joint distribution of observed order flow and parent order sizes, and the marginal distributions. For high-dimensional data, this is challenging.

One approach uses k-nearest neighbors (k-NN) to estimate local probability densities. Another uses kernel density estimation or neural network density models. Practical implementations often use plug-in estimators or binned approaches for computational efficiency.

Multi-Information Framework

Rather than measuring total information leakage, algorithms often break it down by information type:

• Size information leakage: How much does order flow reveal about total order size?
• Direction information leakage: Does order flow reveal whether the algorithm is buying or selling?
• Timing information leakage: Does the timing of orders reveal when execution will complete?
• Urgency information leakage: Does the order flow reveal how time-critical the execution is?

Different algorithms will tolerate different levels of leakage in different dimensions. An algorithm with unlimited time might accept direction leakage but avoid urgency leakage.

Practical Measurement Approaches

In practice, measuring information leakage requires:

• Historical order-flow data with known parent order sizes (often obtained from proprietary trading records)
• A model of informed traders' inference process (how they estimate parent size from observed flow)
• Measurement of how much inference accuracy improved after observing the order flow

One approach: train a machine learning classifier to predict parent order size from observed order-flow sequence. The classifier's accuracy before and after observing a new order quantifies information leakage from that order.

Relationship to Market Impact

Information leakage and adverse selection are closely related. When traders infer the execution algorithm's intent, they position against it, raising prices paid for buy orders (and lowering prices received for sell orders). This manifests as increased market impact.

Algorithms designed to minimize information leakage naturally achieve lower market impact. The two goals align—they are two ways of viewing the same underlying economic reality.

Game-Theoretic Perspectives

The interaction between execution algorithms and informed traders can be modeled as a game. The algorithm wants to minimize information leakage; informed traders want to infer the algorithm's intent. The equilibrium of this game determines actual information leakage in the market.

Theoretical work shows that in equilibrium, some information must leak—complete secrecy is impossible. The question becomes: how much is inevitable, and how much is excess?

Obfuscation Techniques

Algorithms can reduce information leakage by introducing randomness or obfuscation:

• Random order sizing: vary order size unpredictably
• Random timing: vary submission intervals
• Order placement: sometimes place bids, sometimes asks (if not directional)
• Venue diversity: spread orders across multiple venues unpredictably

However, excessive obfuscation can itself signal something unusual is happening, potentially worsening informed trading against the algorithm.

Conclusion

Information leakage is an inherent feature of fragmented markets with informed trading. Mutual information and related metrics allow quantifying how much execution algorithms reveal about their intent. By measuring and managing information leakage, algorithms improve execution prices and reduce adverse selection losses.