Can Digital-Transformation Disclosure Improve Following-Year ROA Estimation? A Small-Sample Machine-Learning Study of Chinese Listed Restaurant Firms

Abstract

This small-sample study tests whether an annual-report disclosure signal improves following-year return-on-assets (ROA) estimation in Chinese listed restaurant firms. The panel contains three firms, 20 transitions, 11 chronological training observations, and nine holdout observations. A frozen 12-term Chinese dictionary measures digital-transformation disclosure density. We compare historical-mean and current-ROA persistence references with controls-only, controls-plus-disclosure, and controls-plus-year versions of fixed Ridge, random-forest, and gradient-boosting regressors. Disclosure lowers holdout ROA mean absolute error (MAE) relative to controls in all three model families: 0.1027 to 0.0615 for Ridge, 0.0754 to 0.0672 for random forest, and 0.0746 to 0.0633 for gradient boosting. Only Ridge is lower than the persistence MAE of 0.0626, and disclosure density is correlated with feature year in the training panel (Spearman 0.787). Results are descriptive, without a hypothesis test. Disclosure density is therefore an auxiliary feature for longer-panel validation, not evidence of stable independent, causal, or generalizable forecasting value.