Cross-stock information and stock price predictability: evidence from ensemble learning

The aim of this study is to overcome the limitations of information and data leakage, instability, and reuse of the model to develop a closing price forecasting model with novel forecasting features for a specific security. Through the Stacking Ensemble approach, the study integrated Elastic Net and Random Forest Gradient Boosting to measure the influence of exchange rates and cross-stock data strength. In addition, the use of the Walk-Forward Validation method overcomes data leakage and increases the integrity of the experiment out-of-sample. Data-driven Monte Carlo simulation with a 243-day trading period helps to forecast long-term price scenarios and enhance the model’s risk assessment capabilities. Experimental results out-of-sample demonstrate the effectiveness of the Stacking model with low forecasting errors, specifically RMSE of 353.2 Viet Nam Dong (VND) and MAE of 274.2 VND. The study also underlines the criticality of sound validation techniques in financial forecasting to circumvent overly optimistic results. The study contributes a standardized process from multidimensional data extraction, characterization techniques, hyperparameter optimization, and performance evaluation, and expands the application of Monte Carlo simulation for long-term scenario forecasting in the stock market. Other macro and micro environmental factors can be integrated into this model to explore deeper learning insights and generalize the model further to other emerging financial markets.