Artificial Intelligence for Earnings-Based Market Risk Analysis: A Machine Learning Framework for Post-Earnings Announcement Drift

Financial markets react continuously to corporate disclosures, yet stock prices may not adjust immediately to new information. One of the most widely studied anomalies associated with post-announcement return behavior is post-earnings announcement drift (PEAD), in which stock prices continue to move in the direction of earnings surprises after the announcement date. Although PEAD has been extensively examined in developed markets, limited research has investigated post-earnings announcement return behavior in emerging markets, particularly in the Saudi stock market. Existing studies largely rely on traditional statistical approaches or focus on developed-market datasets, leaving a gap in understanding how artificial intelligence (AI)-driven machine learning techniques can support financial risk analysis and data-driven financial econometrics by classifying post-earnings announcement return behavior across different firm sizes and post-announcement horizons. To address this gap, this study develops an AI-driven financial risk analytics framework for classifying operational PEAD-related post-announcement return observations using financial and technical indicators derived from companies listed on the Saudi Stock Exchange (Tadawul). The analysis utilizes five years of historical data (2020–2025) from 60 firms equally distributed across small-cap, mid-cap, and large-cap categories. Cumulative return is employed as the response variable and evaluated over three forward post-announcement horizons of 5, 21, and 30 trading days. Five supervised learning models, namely Logistic Regression, Random Forest, Gradient Boosting, XGBoost, and CatBoost, are trained and evaluated using standard classification metrics. The results show that operational PEAD-related return classifications are observed across all market-capitalization categories, with the proportion of observations assigned to the high-return class ranging from 20.55% to 23.30%. These proportions arise from our study’s predefined upper-quartile return-based classification criterion and therefore should not be interpreted as estimates of the prevalence or persistence of classical PEAD. Classification performance varies across market segments and post-announcement horizons. Because the upper-quartile threshold produces an inherently imbalanced classification problem, accuracy results are interpreted alongside class-specific performance and F1 scores rather than in isolation. CatBoost achieved the highest reported accuracy of 77.8% at the 21-day horizon for large-cap firms, while also delivering the highest overall-market accuracy of 76.9% at the 30-day horizon. These results indicate that the selected financial and technical indicators contain information useful for classifying the predefined high-return post-announcement observations, although the reported accuracy levels should not be interpreted independently of the majority-class baseline. The proposed framework contributes to AI-enabled financial econometrics by providing a data-driven approach for classifying post-announcement return behavior and supporting market risk assessment, investor decision-making, earnings-based trading strategies, and market monitoring in emerging financial markets.