Entropy-Weighted Portfolios Overview

• Entropy-weighted portfolios are a family of methods that apply information-theoretic measures, such as Shannon and Rényi entropy, to evaluate portfolio risk and diversification.
• They utilize various approaches by applying entropy to asset returns, allocation vectors, and terminal wealth distributions, leading to multiple allocation and ranking strategies.
• Practical implementations highlight challenges in entropy estimation, sensitivity to parameter choices, and the need to balance theoretical rigor with real-world trading constraints.

Searching arXiv for recent and foundational papers on entropy-weighted portfolios and closely related entropy-based portfolio construction.

Entropy-weighted portfolios are portfolio constructions, benchmark-relative reallocations, or diversification diagnostics in which an entropy functional enters the portfolio problem. The term does not denote a single canonical method. In the literature, entropy may be the Shannon entropy of portfolio returns, the negative information entropy of terminal wealth, the Rényi or Tsallis entropy of a return distribution, the Kullback–Leibler divergence from a benchmark, or the generating function of a functionally generated portfolio in stochastic portfolio theory. Accordingly, entropy can determine weights directly, modify a covariance-based optimizer, rank assets for inclusion in equal-weighted portfolios, or merely evaluate diversification ex post rather than produce an allocation rule ex ante (Urbanowicz, 2014, Lassance et al., 2017, Ruf et al., 2019, Qiu, 28 Oct 2025).

1. Conceptual scope and principal definitions

At the most general level, the entropy-based portfolio literature replaces or supplements second-moment risk with a distributional uncertainty functional. For a continuous return variable $X$ with density $p_X(x)$, the Shannon differential entropy used repeatedly in this literature is

$H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$

while the continuous Rényi family is

$$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$

A benchmark-relative formulation replaces entropy of returns by relative entropy between weight vectors, most cleanly as

$$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$

with $w$ the portfolio and $b$ a benchmark. A distinct strand, originating in stochastic portfolio theory, defines an entropy-weighted portfolio directly through market weights $\mu$, for example

$$\pi_j(\cdot)=\frac{\mu_j(\cdot)\log\mu_j(\cdot)}{\sum_{i=1}^{d}\mu_i(\cdot)\log\mu_i(\cdot)}.$$

A further strand defines diversification as the negative information entropy of the final portfolio value distribution,

$$D(A|I)=\int_{-\infty}^{\infty} p(A_a|I)\log\big(p(A_a|I)\big)\,da,$$

which is explicitly “just the negative of the information entropy” rather than an allocation rule (Rodriguez-Rodriguez et al., 2022, Lassance et al., 2017, Qiu, 28 Oct 2025, Ruf et al., 2019, Kirchner et al., 2011).

This variety of definitions explains a persistent ambiguity in the phrase “entropy-weighted portfolio.” In some papers, entropy is applied to the allocation vector; in others, to the return distribution of each asset; in others, to the return distribution of the portfolio as a whole; and in yet others, to the deviation from a benchmark. A plausible implication is that the topic is better understood as a family of entropy-governed portfolio methodologies than as a single optimizer.

2. Entropy as portfolio-level uncertainty and diversification

A major part of the literature uses entropy primarily as a portfolio-level uncertainty measure rather than as a direct source of weights. In cryptocurrency data, daily simple returns for 18 cryptocurrencies from 10/16/2018 to 12/31/2021, with 1172 daily observations per cryptocurrency, were found to be non-Gaussian and heavy-tailed. The Shapiro–Wilk normality test rejected normality for all cryptocurrencies, with $p_X(x)$0-values “at most of the order of $p_X(x)$1,” and Student-$p_X(x)$2 fits had higher log-likelihoods than normal fits for every asset. In that setting, Shannon entropy was used to evaluate equally weighted portfolios formed under $p_X(x)$3, with assets randomly selected and added sequentially. The main result was that both normal entropy and empirical entropy decreased as the number of assets increased, while normal entropy was always larger than empirical entropy. The paper was explicit that this was not an entropy-based optimizer: entropy evaluated diversification in equally weighted portfolios rather than determined nontrivial weights (Rodriguez-Rodriguez et al., 2022).

A broader equity study treated entropy as a nonparametric total-risk measure using daily excess returns of 150 randomly selected S&P 500 constituents over 1985–2011. That paper estimated Shannon and Rényi entropy from histogram, kernel, and sample-spacing methods and found that, for 10 million random equally weighted portfolios, entropy decreased with portfolio breadth in a way similar to standard deviation. For a portfolio of 10 randomly selected securities, it reported roughly 40% risk reduction relative to a single-security portfolio under standard deviation, Shannon entropy, and Rényi entropy. The same study argued that portfolios in expected return–entropy coordinates are empirically “situated on a hyperbola as in the portfolio theory of Markowitz,” but it did not formulate an explicit entropy optimizer over $p_X(x)$4 (Ormos et al., 2015).

An even more general formulation measures diversification from the full distribution of final portfolio value rather than from return moments or weight concentration. For a normally distributed final portfolio value with variance $p_X(x)$5,

$p_X(x)$6

so the measure is a logarithmic function of variance in the Gaussian case, but it remains defined for non-Gaussian and even infinite-variance cases such as the Cauchy distribution. This line of work shifts the portfolio question from “how spread are the weights?” to “how concentrated is the induced terminal wealth distribution?” and thereby places entropy at the portfolio-outcome level rather than the allocation-vector level (Kirchner et al., 2011).

3. Entropy inside optimization and weight determination

Direct entropy-weighted portfolio construction appears in several distinct forms. One modifies the classical Markowitz program by adding an asset-specific Shannon entropy penalty computed from each asset’s return distribution,

$p_X(x)$7

The corresponding free-utility functional is

$p_X(x)$8

so the covariance matrix is replaced by an entropy-adjusted quadratic form $p_X(x)$9. In this construction, entropy is not imposed on the weight vector itself; it is an asset-specific diagonal correction to covariance risk, derived from the return time series of each asset (Urbanowicz, 2014).

A second, more direct approach minimizes the entropy of the portfolio return distribution itself. The minimum Rényi entropy portfolio is defined by

$H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$0

where

$H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$1

In this formulation, entropy is the portfolio’s core risk criterion, not a regularizer on weight concentration. The paper further imposed the simplex constraint and, in its empirical study, a global variance-based constraint. A Gram–Charlier expansion showed that the criterion depends on variance, skewness, and kurtosis, with the parameter $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$2 governing the relative importance of central versus tail behavior. Empirically, low $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$3 values such as $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$4 and $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$5 yielded better risk-return-turnover trade-offs than minimum-variance benchmarks on six Kenneth French datasets (Lassance et al., 2017).

A third construction derives weights from entropy indices extracted directly from time-series structure rather than from return moments. Using detrending moving average cluster entropy, the method computes for each asset an entropy index $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$6 from the distribution of cluster durations in the volatility series and then sets

$H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$7

This is a literal entropy-weighting rule: weights are normalized entropy-derived indices rather than solutions of a mean-variance, mean-CVaR, or benchmark-tracking program. The empirical illustration on five high-frequency equity indices emphasized diversity, robustness, and stability of the resulting weights rather than a conventional out-of-sample performance table (Murialdo et al., 2021).

A fourth modern construction chooses the portfolio closest to a benchmark in KL divergence subject to factor-exposure constraints. With benchmark $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$8, exposure matrix $H_1(X)=-\int p_X(x)\ln p_X(x)\,dx,$9, and target exposure vector $$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$0, the baseline problem is

$$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$1

and the unique strictly positive solution, under strictly positive benchmark and feasible targets in the convex hull of exposures, has exponential-tilt form

$$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$2

The dual objective is

$$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$3

with gradient $$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$4 and Hessian $$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$5. This benchmark-relative entropy minimization is a particularly clean example of an entropy-weighted portfolio in the sense of multiplicative reweighting from a prior allocation (Qiu, 28 Oct 2025).

4. Functionally generated entropy-weighted portfolios

Within stochastic portfolio theory, entropy-weighted portfolios form a canonical class of functionally generated portfolios. In one empirical implementation, the target weights are

$$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$6

where $$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$7 are market capitalization weights. Because $$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$8 implies $$H_\alpha(X)=\frac{1}{1-\alpha}\ln\int f(x)^\alpha dx.$$9, the weights are positive and sum to one. Relative to cap-weighting, this construction tilts toward smaller stocks, but less aggressively than equal weighting. In CRSP U.S. stock data from January 2, 1962 to December 30, 2016, the entropy-weighted portfolio generally beat the index-tracking portfolio in frictionless settings and proved markedly more robust to proportional transaction costs than equal weighting; at $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$0, weekly or monthly trading often preserved slight outperformance over the index-tracking benchmark, whereas daily trading could erase it (Ruf et al., 2019).

A more recent stochastic-transaction-cost extension treated entropy-weighted portfolios as functionally generated portfolios under random proportional costs. In its empirical section, the entropy generator was

$$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$1

with monthly rebalancing on the 1,000 smallest U.S. common stocks over 2 January 1994 to 31 December 2024. Using intraday bid-ask spreads as the cost proxy, the entropy-weighted portfolio achieved gross CAGR $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$2, net CAGR $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$3, average turnover $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$4 per month, annual cost burden $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$5 bps/year, and outperformed the value-weighted benchmark by $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$6 percentage points annually after cost deduction. Relative to diversity-weighted portfolios, it traded less, paid lower costs, and suffered slightly smaller maximum drawdown (Karimi et al., 12 Jul 2025).

The same tradition also yields a negative benchmark result. When the benchmark itself is the entropy-weighted portfolio generated by

$$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$7

no distinct $$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$8 functionally generated portfolio can dominate it on all compact subsets of the simplex under the same broad diversity-and-volatility conditions that allow relative arbitrage against the market portfolio. In that sense, the entropy-weighted portfolio is a maximal benchmark within a large smooth functionally generated class (Wong, 2014).

A complementary pathwise account decomposes relative log performance against the market as

$$D_{\mathrm{KL}}(w\|b)=\sum_{i=1}^N w_i\log\frac{w_i}{b_i},$$9

where energy is the discrete excess growth rate, relative entropy is $w$0, and control is an investor-chosen rebalancing term. This yields the $w$1-strategy, which spends a fixed fraction of harvested volatility to move the portfolio toward the market and retain the remainder as drift. The framework is not a direct entropy formula for weights, but it makes entropy the core state variable governing benchmark-relative performance (Pal et al., 2013).

5. Entropy-based ranking, sorting, and stock selection

Another important strand uses entropy not to compute weights directly but to rank securities, form risk bins, and hold equal-weighted portfolios within each bin. In one such framework, Tsallis relative entropy (TRE) is computed for each stock relative to the market, stocks are ranked by that entropy-based risk score, divided into bins, and then equally weighted within bins. On S&P 500 data from 4 January 2000 to 30 May 2018, using five-year estimation windows and six-month holding periods, risk–excess-return profiles based on TRE were more consistent than those based on CAPM beta, Kullback–Leibler relative entropy, or relative standard deviation (Devi, 2019).

An asymmetric extension replaced symmetric q-Gaussians by separate half q-Gaussians for negative and positive returns and defined asymmetric Tsallis relative entropy as $w$2. In the corresponding portfolio procedure, stocks were again sorted by entropy risk and equal-weighted within bins. Over long and short samples covering the dot-com bubble, the 2008 crash, and the COVID period, ATRE and TRE produced broadly similar positive linear risk–excess-return relations, but the downside-only measure $w$3 showed much steeper slopes during chaotic periods and, in 90th-percentile risk portfolios, mean earnings of $w$4 against $w$5 for SPY (Devi et al., 2022).

Cross-sectional intrinsic entropy (CSIE) contributes another indirect portfolio use. Here the entropy-like term is built from daily traded-value shares,

$w$6

and combined with OHLC return terms to form daily cross-sectional volatility estimates. The resulting CSIE-based betas compare the volatility of a stock or selected stock set with the volatility of the whole market. The paper explicitly stated that portfolio allocation optimization was not its aim and that constituent weights were considered equal; entropy served as a volatility benchmark and screening device rather than an allocation rule (Vinte et al., 2023).

A further decision-theoretic line introduces normalized fractional order entropy,

$w$7

with $w$8 interpreted as a risk-attitude parameter. This entropy is then combined with expected utility, and optionally variance, in the NEU-FE and NEU-FEV scores for ranking NIFTY50 stocks. The framework selects the lowest-risk actions but does not derive a budget-constrained portfolio vector $w$9. It therefore belongs to the stock-selection rather than the direct portfolio-weighting branch of the literature (Paul et al., 27 Jan 2026).

6. Misconceptions, limitations, and adjacent formulations

A central misconception is that any paper using entropy in a portfolio context is automatically a paper on entropy-weighted portfolios in the narrow sense of a weight formula. Several influential papers explicitly reject that interpretation. The cryptocurrency diversification study, the portfolio-diversification measure based on final wealth entropy, and the entropy-as-risk asset-pricing study all used equal-weighted or randomly assembled portfolios and treated entropy as a complementary risk or diversification measure rather than as an optimizer over $b$0 (Rodriguez-Rodriguez et al., 2022, Kirchner et al., 2011, Ormos et al., 2015).

A second misconception is that entropy methods all act on the same object. Some act on the allocation vector, some on the return distribution of each asset, some on the portfolio return distribution, some on the terminal wealth distribution, and some on scenario probabilities. Maximum-entropy approaches under tail constraints, for example, derive the least-informative distribution of total portfolio return subject to VaR and expected-shortfall constraints and naturally produce barbell-like payoff shapes, but they do not solve for asset weights. Relative-entropy and polynomial-divergence measure-updating frameworks similarly reweight probability measures under moment or marginal-distribution constraints; they are portfolio-model updating devices, not direct allocation formulas (Geman et al., 2014, Dey et al., 2012).

A third distinction concerns “weighted entropy” in the Kelly literature. There the objective is a weighted expected log-growth,

$b$1

and the companion entropy object is a weighted Kullback–Leibler divergence. The resulting optimal strategy is proportional betting under no-ruin constraints. This is an entropy-weighted portfolio only in the sense of outcome-weighted capital growth, not in the cross-sectional sense of benchmark-relative or diversification-based portfolio allocation (Kelbert et al., 2017).

Across the literature, practical limitations recur. Entropy estimation usually requires histogram, kernel, spacing, or parametric density estimation; hence estimates are sensitive to bins, bandwidths, parametric family, and sample size. Some constructions stop at selection or ranking rather than full allocation; others ignore turnover, taxes, or transaction costs; and some recent methods rely on strict positivity, convex-hull feasibility, or well-conditioned exposure covariance for existence and uniqueness. Taken together, these points suggest that “entropy-weighted portfolios” is best read as an umbrella label for several technically distinct ways of inserting information-theoretic structure into portfolio choice, risk measurement, benchmark-relative reweighting, and diversification analysis.