The cross-sectional distribution of portfolio returns and applications

TL;DR

This paper introduces advanced mathematical and computational tools to model and analyze the distribution of long-only portfolio returns across thousands of assets, enabling efficient calculation of moments and application in portfolio optimization.

Contribution

It develops novel formulas and algorithms for computing the distribution and moments of portfolio returns, handling large portfolios efficiently and generalizing previous methods.

Findings

Efficient algorithms compute moments up to order 40 in seconds.

First four moments are directly mapped from asset return moments.

Portfolio score-based strategies show less concentration and risk.

Abstract

This paper aims to develop new mathematical and computational tools for modeling the distribution of portfolio returns across portfolios. We establish relevant mathematical formulas and propose efficient algorithms, drawing upon powerful techniques in computational geometry and the literature on splines, to compute the probability density function, the cumulative distribution function, and the k-th moment of the probability function. Our algorithmic tools and implementations efficiently handle portfolios with 10000 assets, and compute moments of order k up to 40 in a few seconds, thus handling real-life scenarios. We focus on the long-only strategy which is the most common type of investment, i.e. on portfolios whose weights are non-negative and sum up to 1; our approach is readily generalizable. Thus, we leverage a geometric representation of the stock market, where the investment set defines a simplex polytope. The cumulative distribution function corresponds to a portfolio score capturing the percentage of portfolios yielding a return not exceeding a given value. We introduce closed-form analytic formulas for the first 4 moments of the cross-sectional returns distribution, as well as a novel algorithm to compute all higher moments. We show that the first 4 moments are a direct mapping of the asset returns' moments. All of our algorithms and solutions are fully general and include the special case of equal asset returns, which was sometimes excluded in previous works. Finally, we apply our portfolio score in the design of new performance measures and asset management. We found our score-based optimal portfolios less concentrated than the mean-variance portfolio and much less risky in terms of ranking.