A Hybrid LSTM-XGBoost Framework for Multi-Horizon S... | AI Research

A Hybrid LSTM-XGBoost Framework for Multi-Horizon Stock Return Prediction Across Diversified Equity Portfolios
Predicting stock market returns is notoriously difficult because financial data is often noisy, nonlinear, and constantly changing. This paper introduces a two-stage hybrid model designed to improve return predictions by combining the strengths of deep learning and traditional statistical methods. By using a neural network to identify complex patterns in historical price data and a gradient-boosted regressor to analyze technical indicators, the researchers aim to create a more robust system for forecasting stock performance across different timeframes.

How the Hybrid Approach Works

The framework operates in two distinct stages. First, a Long Short-Term Memory (LSTM) network—a type of deep learning model well-suited for sequential data—processes 60-day windows of market information. Instead of just predicting a price, the LSTM is used to compress these 60 days of data into a 64-dimensional "temporal embedding," which acts as a summary of recent market dynamics. The same reasoning question is explored in A Unified Physics-Aware Quantum Machine Learning..., which adds a research perspective.
In the second stage, this summary is combined with 14 hand-crafted technical indicators (such as moving averages, RSI, and volume ratios). This creates a 78-dimensional feature vector that is fed into an XGBoost regressor. By using this hybrid approach, the model benefits from both the LSTM’s ability to learn from sequential history and the XGBoost model’s efficiency in finding patterns within structured, domain-specific indicators.

Performance and Results

The researchers tested their model on 14 U.S. stocks across six different industries, looking at prediction horizons of 30, 90, 252, and 365 trading days. The results showed that the hybrid model consistently outperformed a standalone LSTM model. When compared to an XGBoost-only model, the hybrid approach provided selective benefits, particularly for high-volatility growth stocks and certain cyclical industrials. The ai search story also surfaces in Google AI Releases TimesFM 3 for..., adding another angle.
While the model’s directional accuracy—its ability to correctly predict whether a stock will go up or down—reached as high as 97.6% for the 365-day horizon, the authors note that this is largely due to the high base rate of positive returns in the market over that period. They emphasize that the model's performance is most informative at shorter time horizons, where it provides a more distinct signal compared to simply guessing that a stock will always rise.

Key Considerations

The study highlights a few important takeaways for financial forecasting:

• Horizon Sensitivity: The model’s error rates increase as the prediction window gets longer, which is expected given the higher variance of long-term market returns.

• Sector Differences: The added value of the LSTM temporal embeddings varies by sector. Defensive, low-volatility stocks like those in the healthcare sector saw less improvement from the hybrid approach compared to more volatile technology or consumer stocks.

• Evaluation Strategy: The researchers used a multi-stock pooled training method with strict chronological splits. This approach helps the model learn patterns across different companies while ensuring that the training process does not "look ahead" into the future, which is a common pitfall in financial modeling. The same ai evaluation question is explored in Xiaomi-TabLDM, which adds a research perspective.
  Ultimately, the paper suggests that this hybrid architecture can serve as a useful tool for portfolio ranking and decision support, providing a more nuanced view of market trends than either deep learning or traditional technical analysis could provide on their own. as detailed in the full paper on Arxiv