Macro-Informed Machine-Learning ROE Forecasts and Residual Income Valuation: Evidence from U.S. Commercial Banks

This study develops a two-stage empirical framework that integrates macroeconomic information, panel econometrics, machine learning, and residual income valuation to improve bank equity valuation. Stage 1 forecasts one-quarter-ahead return on equity (ROE) for 20 U.S. commercial banks using 680 bank-quarter observations from 2017Q1 to 2025Q4. Twelve econometric and machine-learning specifications are evaluated using a chronological out-of-sample design. LightGBM achieves the strongest point-estimate performance, with a test RMSE of 3.308 and R² of 0.405, compared with 4.399 RMSE for Pooled OLS. Diebold–Mariano tests, however, do not establish statistically significant superiority of LightGBM over the main benchmark models. Stage 2 incorporates the ROE forecasts into an enhanced residual income valuation framework and applies the methodology to JPMorgan across nine valuation dates from 2024Q1 to 2026Q1. Compared with a traditional persistence-based residual income valuation approach, the enhanced framework reduces MAPE from 44.07% to 21.91% and RMSE from $121.25 to $70.16. The results provide evidence that macro-informed machine-learning ROE forecasts can improve the empirical accuracy of residual income valuation in this setting, while the statistical tests and limited valuation sample indicate that the findings should not be interpreted as evidence of universal machine-learning superiority.