Testing LSTM Networks for Daily Equity Index Return Signals

Summary

This study examines whether LSTM recurrent neural networks can classify the direction of near-term SPY returns from historical daily returns. Its base model uses sequential return inputs, stacked LSTM layers, dropout and batch normalization, and predicts whether the average return over the next few days is positive or negative. The data are split chronologically into training and validation periods, with scaling parameters learned from training data. The article explains how gated memory lets an LSTM retain information over time and contrasts this approach with models that depend on manually selected technical indicators.

Key ideas

• The base model classifies the sign of a multi-day average SPY return using recent daily returns.
• The network combines stacked LSTM layers with regularization and normalization layers.
• The authors repeated model fitting to assess variability caused by random initialization and stochastic optimization.
• Validation losses rose as training continued, which the article interprets as evidence of overfitting.
• The article notes that LSTMs are computationally expensive and that its reported results do not guarantee profitable trading.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.