Hypothesis Testing and Machine Learning Approaches to Investment Risk Assessment in Romania

This study proposes a hybrid framework for investment risk assessment in the Romanian equity market by integrating classical hypothesis testing with machine learning techniques. Using daily data for the BET, BET-FI, and BET-NG indices, the analysis evaluates return behavior, volatility dynamics, and return direction in an emerging market context. Classical inferential tests indicate no statistically significant structural breaks in mean returns or seasonal patterns, while volatility-related measures, particularly intraday price dispersion, exhibit consistent explanatory relevance across indices. Machine learning models, namely Random Forest and XGBoost, are applied to predict returns, short-horizon volatility, and return direction using price-based indicators, sentiment variables, and a purchasing power index (PPI). The results show that XGBoost systematically outperforms Random Forest, especially for return direction and short-term volatility prediction, highlighting the importance of nonlinear modeling in capturing complex market dynamics. However, overall predictive performance remains moderate, reflecting the inherent limits of predictability in volatile and structurally evolving markets. The inclusion of purchasing power information improves interpretability and, in selected cases, volatility prediction at the sectoral level, but does not materially alter aggregate market predictability. Overall, the findings underscore the complementarity of hypothesis testing and machine learning approaches and provide evidence that investment risk in the Romanian equity market is primarily volatility-driven rather than return-driven.