INVESTMENT FORECASTING USING A HYBRID LSTM-CATBOOST-RANDOM FOREST ENSEMBLE WITH SHAP EXPLAINABILITY

Abstract

In the modern context of high financial market volatility, investment forecasting has become a crucial challenge. Forecast accuracy directly affects asset management strategies and overall economic stability. Traditional econometric models (ARIMA, VAR, GARCH, etc.) have limitations in capturing nonlinear and high-dimensional dependencies inherent to financial time series. Contemporary machine learning methods such as LSTM networks, gradient boosting, and ensemble algorithms have demonstrated significant improvements in forecast accuracy. However, the “black box” problem remains, making it difficult to interpret their outputs. This paper proposes a hybrid investment forecasting method that combines LSTM, CatBoost, and Random Forest within an optimized weighted ensemble and integrates SHAP (SHapley Additive exPlanations) to provide model interpretability. The methodology leverages models capable of capturing both temporal patterns and structured macroeconomic and market indicators, with results integrated via a weighted scheme. SHAP analysis enables quantitative assessment of each feature’s contribution to the final prediction, ensuring that the model is not only accurate but also transparent and applicable in financial analytics, where decision explainability is critical. Experimental validation on macroeconomic and stock market data demonstrates that the ensemble provides higher forecasting accuracy compared with the individual models. These results confirm the effectiveness of hybrid approaches in investment forecasting and highlight the potential of Explainable AI applications in financial systems.

Keywords: investment forecasting; machine learning; LSTM; CatBoost; Random Forest; model ensemble; SHAP; explainable artificial intelligence; financial analytics,  information technology, information system..