New Hybrid Model Combines ResNet and LSTM to Boost Stock Price Prediction Accuracy

Machine learning researchers have developed a new forecasting model designed to tackle one of finance's persistent challenges: predicting stock price movements. The approach, called ResNLS, combines two complementary neural network architectures to better account for how stock prices move in relation to one another over time.

How ResNLS Works#

The model pairs ResNet with LSTM (Long Short-Term Memory) networks in a hybrid design. ResNet functions as a feature extractor, identifying patterns and dependencies that exist between consecutive stock prices. The LSTM component then processes the original time series data while incorporating these detected dependencies, treating them as residual signals that help refine predictions.

The architecture recognizes that stock prices don't move in isolation—each day's price is influenced by patterns in recent trading activity. By explicitly modeling these interconnections, the system can capture relationships that traditional models often miss.

Peak Performance at Five-Day Window#

The research team tested the model using different input windows—ranging from recent trading data to longer historical periods. ResNLS-5, the configuration using closing prices from the five most recent trading days, emerged as the strongest performer. This variant achieved at least a 20% improvement in accuracy compared to current state-of-the-art baseline models, suggesting that immediate price history carries the strongest predictive signal.

Real-World Trading Validation#

To assess whether these improved forecasts translate to actionable profit potential, the researchers developed a quantitative trading framework and conducted backtests using the ResNLS-5 predictions. The results indicate that strategies built on the model's forecasts can limit losses during market downturns and capture gains during rallies, bridging the gap between academic prediction and practical trading application.