This research investigates the predictability of Bitcoin price directions following investor overreaction, a prominent anomaly widely documented in behavioral finance. Leveraging a multimodal dataset—integrating transactional variables, technical indicators, on-chain metrics, and social signals derived from Google Trends—this study analyzes market dynamics from August 2017 to December 2023 and identifies 303 distinct overreaction episodes. To construct a reliable predictive framework, a rigorous three stage feature selection pipeline was implemented, combining Random Forest screening, correlation matrix analysis, and Variance Inflation Factor (VIF) assessment. This process yielded 11 highly informative and non multicollinear variables that served as the input features for the proposed multivariate Long Short Term Memory (LSTM) model. The LSTM model achieved a directional prediction accuracy of 67%, substantially outperforming the classical Logistic Regression benchmark with 53.4% accuracy. Moreover, the consistently declining and stable training and validation loss curves demonstrate strong generalization performance and confirm the model’s robustness against overfitting. Empirical findings reveal statistically significant short term market inefficiencies and predictable abnormal returns following overreaction events at the 99% confidence level. Overall, the study provides compelling evidence of exploitable behavioral patterns within the Bitcoin market and offers a practical foundation for developing advanced algorithmic trading strategies and dynamic risk management systems in cryptocurrency markets.