Intelligent Long-Horizon Stock Profit Prediction using Technical Feature Engineering and Stacked Recurrent–Boosted Tree Models

Abstract

Stock markets exhibit complex, nonlinear, and uncertain behavior, making long-horizon profit forecasting a challenging task for investors and analysts. Traditional price prediction approaches often focus on short-term fluctuations, offering limited actionable insight for long-term investment planning. In response to this limitation, the study presents an intelligent profit-oriented stock prediction framework that identifies whether a stock can achieve a predefined long-term return target. Technical feature engineering is employed to extract trend, momentum, and volatility characteristics using indicators such as Exponential Moving Averages (EMA), Relative Strength Index (RSI), Moving Average Convergence Divergence (MACD), and Bollinger Band midlines. These engineered features serve as inputs to two complementary predictive models: a boosted decision tree classifier (XGBoost) and a stacked recurrent neural network (Stacked-LSTM). The models are evaluated on multiple Indian stocks over extended historical horizons to assess their ability to classify profitable long-term return opportunities. Experimental results demonstrate stable generalization performance across diverse stocks, with Stacked-LSTM consistently achieving slightly higher predictive accuracy than XGBoost. The findings confirm that integrating technical feature engineering with Machine Learning(ML) and Deep Learning(DL) architectures provides a reliable decision-support tool for long-term profit-oriented stock market forecasting.