A Hybrid LSTM-XGBoost Framework for Multi-Horizon Stock Return Prediction Across Diversified Equity Portfolios
The paper proposes a hybrid two-stage architecture combining LSTM and XGBoost for multi-horizon stock return prediction across 14 U.S. equities in six sectors. LSTM uses two stacked layers with 64 hidden units on 60-day windows of five features to produce 64-dimensional embeddings. These are concatenated with 14 technical indicators to form a 78-dimensional vector for XGBoost, tuned via 3-fold cross-validation. Training uses chronological splits and per-stock MinMaxScaling to prevent look-ahead bias. Evaluation covers 30, 90, 252, and 365 trading day horizons.
A research paper introduces a hybrid LSTM-XGBoost model for predicting stock returns over multiple horizons. It processes sequential market data with LSTM and combines the learned embeddings with technical indicators before feeding them into XGBoost. The approach is tested on a diversified set of U.S. stocks and evaluated across four prediction horizons.
The architecture leverages LSTM's ability to capture temporal dependencies and XGBoost's strength in handling tabular features, potentially improving robustness over single-model approaches. The use of chronological splits and per-stock scaling addresses common pitfalls like look-ahead bias. The multi-horizon evaluation (30 to 365 days) suggests an attempt to assess both short-term and long-term predictive power.
Hybrid deep learning and gradient boosting models are increasingly explored in quantitative finance for their complementary strengths. This work may influence practitioners seeking more accurate return forecasts for portfolio management. The focus on diversified equities indicates applicability across sectors.
Improved stock return prediction can enhance trading strategies, risk management, and asset allocation. Financial institutions may benefit from integrating such hybrid models into their quantitative pipelines, potentially leading to better investment decisions and competitive advantage.
Future work may extend the framework to larger asset universes, incorporate alternative data sources, or explore attention mechanisms. The reported performance could encourage adoption in algorithmic trading systems, pending further validation on out-of-sample data and robustness checks.