Risk-adjusted portfolio optimisation using a parallel multi-objective evolutionary algorithm

Abstract

In this article we describe the use of a multiobjective evolutionary algorithm for portfolio optimisation based on historical data for the S&P 500. Portfolio optimisation seeks to identify manageable investments that provide a high expected return with relatively low risk. We developed a set of metrics for qualifying the risk/return characteristics of a portfolio's historical performance and combined this with an island model genetic algorithm to identify optimised portfolios. The algorithm was successful in selecting investment strategies with high returns and relatively low volatility. However, although these solutions performed well on historical data, they were not predictive of future returns, with optimised portfolios failing to perform above chance. The implications of these findings are discussed.

Further evidence in support of a low-volatility anomaly: Optimizing buy-and-hold portfolios by minimizing historical aggregate volatility

“…In-sample analyses, whereby portfolios are developed and tested on the same dataset, are problematic because they invite overfitting. It is often easy to find some manipulation of the data that leads to outperformance by exploiting random variation in the dataset, variation which is unlikely to persist into the future (see Hawkins, 2004;Maguire et al, 2012). Accordingly, it becomes difficult to differentiate between models which have been overfitted to the data versus those whose development has not been influenced by random quirks.…”

Maximizing positive porfolio diversification

“…There will often be groups of stocks which, by chance, happen to appear uncorrelated for the training period, hence dominating the optimization process. The hidden risk is that, out-of-sample, these supposedly independent time series will immediately return to being correlated (see [12] for a demonstration of such overfitting).…”

“…In sum, the weaknesses of DR as a measure of diversification are that it does not discriminate between good and bad diversification, cannot handle short investments, and bases its value on relationships within rather than outside the portfolio, leaving it vulnerable to overfitting (see [12]). The weaknesses of DRP as a measure of diversification are that it does not account for idiosyncratic risk where investment constraints are applied, and imposes an arbitrary cut-off for identifying investable principal portfolios, without regard to their positivity.…”