["Business Strategy and the Environment, EarlyView. ", "\nABSTRACT\nESG ratings prediction provides critical reference for investment decisions, as they reflect firm performance and risk management while measuring corporate social responsibility performance. However, existing studies exhibit limitations in predictor completeness and temporal feature utilization: most predictions rely on annual financial data without considering nonfinancial factors such as social or environmental governance. Besides, high‐frequency financial information and its time‐varying features remain underexplored, limiting dynamic assessments of ESG performance. Hence, this paper proposes a novel ESG ratings prediction framework—the RMIDAS (restricted mixed data sampling)—ML (machine learning) framework. The framework integrates heterogeneous information across social, environmental, and financial dimensions when constructing the ESG predictor system. We utilize high‐frequency financial data and introduce internal pay gaps and green innovation achievement as social and environmental indicators to enhance predictive accuracy. Meanwhile, RMIDAS models promote the utilization of mixed‐frequency information and explore the time‐varying patterns of predictors through weight adjusting. We evaluate ML models with single‐frequency predictors and the RMIDAS‐ML framework with mixed‐frequency data. The results demonstrate that RMIDAS‐ML outperforms other models, with the RMIDAS‐RF (random forest) model performing best. Then, by analyzing feature importance and SHAP values, it is observed that firm size, profitability, and internal pay gaps play a significant role. The findings further reveal a nonlinear relationship between ESG ratings and internal pay gaps beyond a U‐shaped correlation. Therefore, our proposed framework not only effectively utilizes time‐series features of mixed‐frequency data for ESG forecasting but also uncovers additional details regarding the connection between internal pay gaps and ESG ratings.\n"]