<![CDATA[Accurately forecasting stock prices remains challenging because of the nonlinear and volatile nature of financial markets, particularly during periods of heightened uncertainty, such as the COVID-19 pandemic. This study evaluates the robustness of three models, LSTM, GRU, and Hybrid LSTM-GRU, for ANTM.JK stock price forecasting using a volatility-oriented evaluation framework. Historical stock data from September 2005 to May 2022 were transformed into supervised time-series datasets using a 15-lag sliding window. The model performance was evaluated using baseline prediction accuracy, 5-fold chronological cross-validation consistency, and synthetic stress scenarios consisting of controlled price drops, price rises, and high-volatility noise. Evaluation metrics included RMSE, MSE, MAE, R, and R^2. The GRU model delivered the top baseline prediction results, achieving the smallest RMSE of 52.95 and MAE of 28.14. In cross-validation, the LSTM model recorded the lowest average RMSE of 119.41. Meanwhile, the Hybrid LSTM-GRU exhibited the highest prediction consistency and robustness across various synthetic stress scenarios. In contrast to earlier research that mainly focused on prediction precision, this study presents a comprehensive framework for evaluating robustness. This framework combines baseline accuracy, consistency through cross-validation, and an analysis of synthetic stress scenarios. The generated robustness map offers a systematic interpretation of model strengths across diverse evaluation goals, facilitating a more thorough assessment of stock-forecasting models in different market environments.]]>
