Risk-Aversion Adjusted Portfolio Optimization with Predictive Modeling

We propose a multi-period portfolio optimization method rooted on the mean-risk framework, in which the risk aversion coefficient is adjusted in response to the market trend movement predicted by machine learning models. We use the Gini's Mean Difference to characterize the risk of the portfolio, and employ a series of technical indicators as the features to feed the predictive machine learning models. A set of comprehensive computational tests are carried out within a rolling-horizon approach to evaluate the performance of the generated portfolios. The empirical results show that the regularized logistic regression model provides the best prediction of market trend movement, while the proposed dynamic risk-aversion adjusted portfolio rebalancing strategy generates portfolios with higher time-series cumulative returns than a static strategy with fixed risk-aversion coefficient.