Interpretable decision support for international stock index investments : using heterogenous machine learning stacking for more accurate and resilient mid-horizon predictions
MetadataShow full item record
- Master Thesis 
It remains challenging to beat the wisdom of the crowd by predicting future stock index returns. With the rise in financial machine learning, many new approaches are introduced but predictive accuracy, financial performance, and interpretability are rarely connected. We show that it remains possible to predict future MSCI index returns five years ahead and across seven country indices. We train six base-models that differ in methods and data pools to reduce generalization errors. The stacked combinations of these models reduce MAPE by up to 60.5% if compared out-of-sample to the historical mean forecast bench-mark, while passing the Diebold-Mariano significance tests. In financial terms, our stacked models outperform the equal-weight portfolio by 1.4% to 2.1% in yearly CAGR and by 34% to 72% in Sharpe Ratios. Our research fills a gap for portfolio optimization, by provid-ing more reliable inputs for future returns, and potential for fundamental research, by in-troducing a method that combines predictive accuracy with interpretability.