HYBRID MACHINE LEARNING–DRIVEN FINANCIAL FORECASTING MODELS: INTEGRATING LSTM, PROPHET, AND XGBOOST FOR ENHANCED STOCK PRICE AND RISK PREDICTION

Abstract

This review synthesizes research on hybrid machine learning for equity price and risk forecasting, focusing on combinations of LSTM, Prophet style additive models, and XGBoost. We used a PRISMA guided protocol covering 2015 to 2025 across Scopus, Web of Science, IEEE Xplore, ACM Digital Library, SSRN, and arXiv, with eligibility requiring equity focus, out of sample evaluation that respects time order, and explicit hybridization or risk components. After screening and full text assessment with reasons coded exclusions, the final qualitative synthesis comprised 110 studies. Across this evidence, deliberately engineered hybrids consistently outperform single learners on point accuracy, directional reliability, and risk calibration. Normalized comparisons in one day ahead settings show typical reductions in RMSE near 9 percent and gains in directional accuracy around five to six percentage points, with tighter Value at Risk coverage under quantile aware training. Benefits persist under strict rolling origin validation with nested tuning, and widen during turbulent regimes where dynamic weighting and residual correction add stability. The literature also emphasizes explainability and governance, recommending component plots for structural layers, Shapley value attributions for tree ensembles, and ablations that quantify each module's marginal value. Drawing these threads together, we outline a blueprint that decomposes trend and seasonality with a structural layer, models nonlinear temporal dynamics with an LSTM, learns interaction rich signals with XGBoost, and combines outputs using out of fold stacking and calibrated risk heads. This evidence-based specification offers decision grade forecasts and tail risk estimates for equity markets for deployment.