Direct Answer
Mutual information measures how much knowing one variable reduces uncertainty about another. Unlike correlation it captures any dependence, linear or not, and equals zero only when the variables are independent. Zero indicates independence, and larger values indicate stronger dependence of any shape.
Mutual Information: Formula, Meaning, Signals, Examples and How to Use It
How Is Mutual Information Calculated?
Mutual information measures how much knowing one variable reduces uncertainty about another. Unlike correlation it captures any dependence, linear or not, and equals zero only when the variables are independent.
Formula: I(X; Y) = sum over x, y of p(x, y) * ln( p(x, y) / (p(x) * p(y)) ), measured in nats with the natural log
Inputs:
- Two aligned series
- Binning or estimator choice
- Window length
Worked example. Two perfectly dependent fair binary variables, each outcome pair having probability 0.5, give a mutual information of 0.6931 nats. The inputs are illustrative numbers, not market data.
Reading it. Zero indicates independence, and larger values indicate stronger dependence of any shape. The maximum depends on the entropies of the variables, so values are often normalized to compare pairs.
Limits. Estimates from continuous data depend on binning or kernel settings and are biased upward in small samples. It does not show direction or sign of the relationship.
What is Mutual Information?
Mutual information measures how much knowing one variable reduces uncertainty about another. Unlike correlation it captures any dependence, linear or not, and equals zero only when the variables are independent.
How to use Mutual Information
Zero indicates independence, and larger values indicate stronger dependence of any shape. The maximum depends on the entropies of the variables, so values are often normalized to compare pairs. A reading alone is never a buy or sell signal; it needs context from price structure, volume, volatility and the broader market regime.
Strengths
- It registers any dependence between variables, including nonlinear patterns that leave Pearson correlation near zero.
- The population value is zero only when the variables are independent, which gives it a precise null case.
- Reproducible when the formula and data convention are fixed.
Weaknesses and limitations
- The unit is nats and the value has no upper bound of 1, so it is not directly comparable with a correlation coefficient.
- Sparse regions of the joint distribution carry heavy weight in the log term, so a few rare combinations can drive the estimate.
- Correlated indicators can create false confidence by repeating the same underlying information.
Editorial and risk note
This page is educational content, not individualized investment advice. It should not imply guaranteed prediction, accuracy or outperformance.
Frequently Asked Questions
Is Mutual Information a buy or sell signal?
No. Zero indicates independence, and larger values indicate stronger dependence of any shape. Estimates from continuous data depend on binning or kernel settings and are biased upward in small samples. A trade built on it still needs a trigger, an invalidation rule, a position size and realistic execution assumptions.
What is the best setting for Mutual Information?
Mutual Information takes these inputs: two aligned series; binning or estimator choice; window length. There is no universal best setting for the tunable ones, so test them across instruments, regimes and out-of-sample periods and favor stable parameter regions over one historical winner.
Can Mutual Information be used by itself?
Estimates from continuous data depend on binning or kernel settings and are biased upward in small samples. It does not show direction or sign of the relationship. For that reason it is read next to measures of direction, participation, valuation, liquidity and risk, not alone.
Does Mutual Information work on every timeframe?
The calculation (I(X; Y) = sum over x, y of p(x, y) * ln( p(x, y) / (p(x) * p(y)) ), measured in nats with the natural log) can be run on any bar size, but behavior is not portable: a reading on a five-minute chart describes a very different market window from the same reading on a daily chart. Validate the exact timeframe before relying on it.
Why does Mutual Information give false signals?
The unit is nats and the value has no upper bound of 1, so it is not directly comparable with a correlation coefficient. Sparse regions of the joint distribution carry heavy weight in the log term, so a few rare combinations can drive the estimate. Lag, noise and regime changes add to this, because market participants react to new information after the reading is calculated.
How should Mutual Information be backtested?
Rebuild Mutual Information from two aligned series using point-in-time data, write the rules before testing, include delisted securities where relevant, model realistic fills and costs, reserve a validation sample and compare with a simpler baseline.