Scientific Portfolio Optimization: A Risk-Adjusted Approach to Asset Allocation

Abstract

This dissertation evaluates the robustness of traditional, risk-based, and machine learning (ML) portfolio optimization methods under realistic market conditions. Classical Mean–Variance Optimization (MVO) is elegant in theory but fragile in practice due to estimation error and instability in crises. Risk-based approaches such as Risk Parity (RP) and Hierarchical Risk Parity (HRP) provide more resilient alternatives by allocating on volatility and correlation structures instead of unstable return forecasts. ML-enhanced MVO (ML-MVO), which substitutes predicted for historical returns, remains of uncertain value. A modular Python artefact was developed to compare these strategies using rolling five-year windows, monthly rebalancing, and strict walk-forward validation, complemented by an interactive dashboard interface. Performance was assessed through risk-adjusted metrics (Sharpe, Sortino, maximum drawdown, volatility) across both normal and crisis regimes, including the Global Financial Crisis (2008–2009) and the COVID-19 shock (2020). Sensitivity analysis with realistic weight constraints was also conducted to test robustness under practical implementation settings. Results show HRP consistently achieved the most robust risk-adjusted outcomes, outperforming MVO and ML-MVO in both full-sample and stressed settings. RP and equal weighting remained competitive baselines, while ML-MVO underperformed despite moderate predictive accuracy. Overall, the findings suggest ML contributes more effectively to restructuring optimization processes, as in HRP, than to direct return forecasting. The study also highlights inherent limitations of short-horizon ML forecasting and points to future research extending horizons, incorporating richer features, and exploring ML-enhanced risk estimation.