Portfolio return distributions: Sample statistics with non-stationary correlations

TL;DR

This paper models non-stationary correlations in multivariate normal distributions by averaging over an ensemble of correlation matrices, resulting in a modified distribution that aligns well with empirical asset return data.

Contribution

It introduces a novel ensemble averaging approach to account for non-stationary correlations in portfolio return distributions, extending traditional models.

Findings

The modified distribution differs from the normal distribution in the center and tails.

The model fits empirical Nasdaq return data well, especially centrally.

Ensemble averaging effectively captures non-stationarity in correlations.

Abstract

We consider random vectors drawn from a multivariate normal distribution and compute the sample statistics in the presence of non-stationary correlations. For this purpose, we construct an ensemble of random correlation matrices and average the normal distribution over this ensemble. The resulting distribution contains a modified Bessel function of the second kind whose behavior differs significantly from the multivariate normal distribution, in the central part as well as in the tails. This result is then applied to asset returns. We compare with empirical return distributions using daily data from the Nasdaq Composite Index in the period from 1992 to 2012. The comparison reveals good agreement, the average portfolio return distribution describes the data well especially in the central part of the distribution. This in turn confirms our ansatz to model the non-stationarity by an ensemble average.