Explainable Machine Learning for One-Year Stock Return Prediction Using Financial Indicators

This study examines whether annual financial indicators can rank U.S. stocks by their subsequent one-year price changes. The dataset contains 22,077 firm-year observations and 221 numeric indicators from 2014–2018. Ridge, Lasso, Random Forest, XGBoost, and LightGBM are fitted using the 2014–2016 observations and evaluated on the 2018 financial-data cohort, whose target covers 2019. The 2017 cohort is excluded from model fitting. LightGBM records the highest test Spearman rank information coefficient, 0.0805, and a top-minus-bottom decile difference of 26.69 percentage points. Lasso has the lowest root mean squared error and mean absolute error, so the preferred model depends on the evaluation criterion. SHAP analysis of LightGBM identifies earnings yield, weighted-average shares outstanding, price-to-book ratio, the Energy sector indicator, and share issuance or buybacks as its five leading inputs. These attributions describe the fitted model rather than causal return determinants. The reported experiment uses full-sample winsorization and lacks filing timestamps; consequently, its chronological split does not eliminate information leakage. The findings are descriptive comparisons within one annual test cohort, and the decile difference should not be interpreted as an implementable investment return.