Application of Machine Learning Models for Risk Level Optimization

Abstract

This thesis explores the use of machine learning models to optimize portfolio risk through empirical analysis. Originally the thesis was targeted at stock selection within indices, but disappointing outcomes prompted a shift towards a simpler task of discerning between risk and risk-free assets in portfolios. Machine learning models trained on price data calculated returns and volatility sought to forecast risk-taking decisions using signals like annual returns and volatility, or signals based on avoiding markets with negative returns, increasing volatility, and extreme market conditions. Combining three of these signals yielded the optimal risk-adjusted performance, which has been analyzed in-depth. Results indicate that machine learning portfolios outperform relevant benchmarks, particularly the ensemble model shows potential in balancing risk and return and avoiding large drawdowns. Moreover, alpha persists even after adjusting for common risk factors, and trade analysis confirms the practical viability of implementation, as the portfolio consistently delivers abnormal returns across economic cycles and calendar years. While out-of-sample tests across different markets attest to robustness, the study acknowledges weaknesses in machine learning and backtesting, stressing the need for caution when applying the results in real-world scenarios. Nonetheless, the findings underscore the efficacy of machine learning methods in constructing portfolios with superior risk-adjusted returns, offering valuable insights for investment decision-making.