Mutual Information — What Statistical Dependence Reveals About Your Models
Correlation tells you about linear relationships. Mutual information tells you about all of them.
Mutual information in trading models measures the statistical dependence between two return distributions — specifically, how much knowing the distribution of returns from one period tells you about the distribution in another. Unlike correlation, which only captures linear relationships, mutual information detects any form of dependency, including complex non-linear patterns that standard tools miss entirely.
Why correlation is not enough for model evaluation
Most traders are familiar with correlation as a way to measure the relationship between two datasets. It is widely used, easy to calculate, and intuitive. But it has a significant blind spot: it only measures linear relationships. If a model's returns in one period bear a non-linear relationship to its returns in another — which is common in real markets — correlation can return a near-zero value even when a strong statistical relationship exists.
Mutual information does not have this limitation. It is drawn from information theory and measures the reduction in uncertainty about one distribution given knowledge of another. In plain terms: how much does knowing one dataset tell you about the other? This captures all forms of statistical dependence, not just linear ones.
In the context of trading model evaluation, this matters considerably. Markets exhibit non-linear dynamics — volatility clusters, regime transitions, and fat-tailed return distributions that linear tools systematically underestimate. A model that produces structurally similar return distributions across different market periods, measured by mutual information, is more likely to be capturing something persistent about the market rather than fitting historical noise.
Mutual information as a distribution stability measure
In darwintIQ, mutual information is one of several distribution similarity metrics used as part of ongoing model assessment — alongside the KS statistic, Jensen-Shannon Divergence, Wasserstein Distance, and PSI. Each of these approaches the same fundamental question from a different mathematical direction: is this model behaving consistently across time?
A high mutual information value between a model's baseline and recent return distributions suggests the model has captured something persistent about the market. The statistical structure of its returns has remained intact. A low mutual information value suggests the two periods look structurally unrelated — a concern for model reliability regardless of whether headline returns are currently positive.
This distinction is important. A model can be profitable in recent periods and still show low mutual information if its return distribution has changed character. Perhaps it is now making fewer, larger trades. Perhaps it is generating returns in different market conditions from those it was tested on. The profitability is real, but the statistical foundation has shifted. Mutual information flags that discrepancy early.
What mutual information looks like in practice
Mutual information is expressed in nats or bits depending on the logarithm base used, but the unit matters less than how the metric behaves over time. What you are looking for is stability: does the score remain consistent across rolling evaluation windows, or does it drop sharply when market conditions change?
A drop in mutual information often accompanies or precedes a drop in the Robustness Score. Together, they suggest the model is encountering market conditions that differ structurally from its validated history — a form of distributional mismatch that tends to precede underperformance rather than accompany it.
In darwintIQ's continuous evaluation framework, mutual information is recalculated on a rolling 4-hour basis. This means traders can observe in near real-time whether the statistical structure of a model's returns remains intact, or whether it has begun to diverge from its baseline. A model that maintains high mutual information across consecutive evaluation windows is demonstrating something important: its behaviour is consistent, its returns are statistically coherent, and its edge — whatever it captures — has not deteriorated in the current market environment.
Final thoughts
Mutual information in trading models is an information-theoretic measure of statistical dependence between return distributions. It captures non-linear relationships that correlation misses, making it a more sensitive detector of structural change in model behaviour. In darwintIQ, it forms part of a multi-metric distribution analysis that helps identify shifts in how a model operates before those shifts become visible in raw performance figures — allowing traders to act on early signals rather than lagging ones.