Machine Learning, Return Predictability, and Implementable Alpha on the Ghana Stock Exchange

This paper examines the return predictability and portfolio performance of machine learning models relative to conventional asset pricing models on the Ghana Stock Exchange using monthly equity, macroeconomic, and commodity price data from 2007:M12-2023:M4. Results show that Elastic Net delivers the strongest out-of-sample performance, achieving an R² of 28.23%, an annualised long-short portfolio return of 88.4%, and a gross Sharpe ratio of 3.28. This substantially outperforms CAPM (-0.08% return; Sharpe-0.003) and ARIMA (14.7% return; Sharpe 0.59). Although the market consists of a concentrated eight-stock liquid universe typical of frontier markets, the performance differential remains economically and statistically significant. After accounting for GSE-specific transaction costs of 1.5%-2.5% round-trip, net Sharpe ratios remain between 1.86 and 2.96. Subsample results further indicate improving predictive performance over time, implying that regularised nonlinear machine learning models can extract persistent alpha in emerging equity markets.