Enhancing stock return prediction and portfolio performance with adaptive LASSO

This paper explores the predictability of monthly US stock returns using adaptive LASSO on firm-specific characteristics from June 1990 to December 2022. By efficiently selecting relevant features, such as lagged returns, mean log-volumes, market values, dividend yields, and R&D expenses, the study develops threshold-based portfolios incorporating transaction costs and no-trade zones. Empirical results show that portfolios based on expected Sharpe ratios outperform benchmarks such as the S&P 500, with the adaptive LASSO portfolio achieving a 337.20% cumulative return and a 76.74% annualised Sharpe ratio. Compared with machine learning methods such as random forests and XGBoost, adaptive LASSO offers superior interpretability and robustness, highlighting its effectiveness for dynamic, cost-aware portfolio management while mitigating overfitting.