Using LSTMs to Predict Five-Day Stock Returns from Price and Volume Data

Summary

The article explains how a long short-term memory network can use sequential market data for short-horizon stock selection. It outlines LSTM’s forget, input, and output gates, then describes a model with an LSTM layer, dropout, and a dense layer. The model predicts five-day returns for A-share stocks using seven basic price and volume inputs transformed into 98 time-series features, with a five-day input window.

The stated workflow clips and cross-sectionally standardizes labels, standardizes and clips features, and fills missing values. It reports tuning the number of LSTM units on an earlier date range, with 256 performing best among the tested choices. However, the supplied text ends before showing the backtest results or detailing the rolling training and portfolio construction. It therefore gives a research setup, not enough evidence to assess profitability, robustness, transaction costs, or data leakage. The page also warns that its platform instructions refer to an older version.

Key ideas

• LSTMs use gates to manage information across sequential inputs and address training difficulties in long sequences.
• The described model combines an LSTM layer with dropout and a dense layer.
• The prediction target is each stock’s return over the next five trading days.
• The input set expands seven basic market variables into 98 derived features and uses a five-day sequence window.
• The article reports selecting 256 LSTM units from three tested sizes, but the provided text omits the associated backtest results.

Tags

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