Folio Lab/Documentation

Entropy (in Portfolio Returns)

Entropy, borrowed from information theory, measures the unpredictability or disorder in a portfolio's return distribution. Higher entropy indicates a more dispersed, less predictable return distribution, while lower entropy indicates returns concentrated around specific values. It provides a model-free measure of distributional complexity that complements traditional volatility measures.

Overview

Shannon entropy, introduced by Claude Shannon in 1948 as the foundation of information theory, quantifies the average information content (or surprise) in a random variable. When applied to financial returns, entropy captures aspects of the return distribution that standard deviation misses: a distribution with multiple modes, fat tails, or complex shapes will have higher entropy than a simple Gaussian with the same variance.

In portfolio analysis, entropy serves multiple purposes. It can measure the diversification of a portfolio (higher entropy of return outcomes suggests greater uncertainty and less concentration of outcomes). It can also be used as an objective function in portfolio optimization, where maximum-entropy portfolios are sought as the least biased allocation given certain constraints. Dionisio et al. (2006) and Zhou et al. (2013) have demonstrated that entropy-based measures provide complementary information to traditional risk metrics and can improve portfolio selection.

Mathematical Formulation

Shannon Entropy

For a discrete random variable with possible outcomes and probability mass function , Shannon entropy is defined as:

Entropy is measured in bits (when using base-2 logarithm) or nats (when using the natural logarithm). By convention, . The entropy is maximized when all outcomes are equally likely ( for all ) and minimized (zero) when one outcome has probability 1.

Discretization: Freedman-Diaconis Binning

Since financial returns are continuous, they must be discretized into bins before computing entropy. The Freedman-Diaconis rule provides a data-adaptive bin width:

where is the interquartile range and is the number of observations. This rule is robust to outliers (unlike Sturges' rule) because it depends on the IQR rather than the range. The number of bins is then .

Probability Estimation

After binning, the probability of each bin is estimated by its relative frequency:

where is the number of observations falling in bin . These estimated probabilities are then substituted into the Shannon entropy formula.

Normalized Entropy

To facilitate comparison across distributions with different numbers of bins, entropy is often normalized by its maximum possible value:

Normalized entropy ranges from 0 to 1. A value near 1 indicates a nearly uniform distribution (maximum unpredictability), while a value near 0 indicates a highly concentrated distribution (high predictability). Note that here refers to the number of bins used in the discretization.

Financial Interpretation

In the context of portfolio returns:

  • Higher entropy: Returns are more dispersed and unpredictable. The distribution occupies more of the return space, suggesting greater uncertainty about future outcomes.
  • Lower entropy: Returns are concentrated in a narrow range. More predictable outcomes but potentially higher concentration risk if the concentration is in the loss region.
  • Comparison with volatility: Two distributions can have the same standard deviation but very different entropies. For example, a bimodal distribution (common in levered strategies) and a Gaussian have different entropies even if their variances match.

Advantages & Limitations

Advantages

  • Model-free: Makes no assumptions about the shape of the return distribution; captures non-Gaussian features.
  • Shape sensitivity: Detects multimodality, fat tails, and other distributional features that variance alone cannot capture.
  • Diversification measure: Can serve as a metric for portfolio diversification quality.
  • Information-theoretic foundation: Grounded in a rigorous mathematical framework with deep connections to coding theory and statistical mechanics.

Limitations

  • Binning sensitivity: Results depend on the choice of binning method and the number of bins. Different binning strategies can yield different entropy values.
  • Sample size requirements: Reliable entropy estimation requires sufficient data to populate the bins; sparse data leads to biased estimates.
  • No directional information: Entropy treats gains and losses symmetrically; it does not distinguish between upside and downside uncertainty.
  • Not widely adopted: Entropy is not a standard metric in the investment management industry, limiting comparability.

References

  1. Shannon, C. E. (1948). "A Mathematical Theory of Communication." Bell System Technical Journal, 27(3), 379-423.
  2. Dionisio, A., Menezes, R., & Mendes, D. A. (2006). "An Econophysics Approach to Analysing Uncertainty in Financial Markets: An Application to the Portuguese Stock Market." European Physical Journal B, 50(1), 161-164.
  3. Zhou, R., Cai, R., & Tong, G. (2013). "Applications of Entropy in Finance: A Review." Entropy, 15(11), 4909-4931.