LSTM-Based Factor Stock Selection with Rolling Windows and Portfolio Rules
Summary
The article explains recurrent neural networks and LSTM gates, then outlines an A-share stock-selection example that predicts five-day returns from sequences of seven factors. It describes missing-value handling, outlier filtering, standardization, and rolling five-day input windows. The model uses an LSTM with dense layers, dropout, and regression output; stocks are ranked by predicted return to form a daily portfolio with holding and capital-allocation rules.
The example trains on 2010–2016 data and evaluates selections over 2016–2019. The article claims the backtest outperformed its benchmark but provides no numeric results in the text, so the strength and reliability of that evidence cannot be assessed here. It labels the implementation as an older version for study. Model choices and portfolio parameters are presented as adjustable, and the discussion notes recurrent-network gradient problems and dropout as an overfitting control; it does not establish that the approach will generalize or remain profitable.
Key ideas
- LSTM adds a cell state and input, forget, and output gates to recurrent sequence processing.\nThe example uses seven factors across rolling five-day windows to predict five-day stock returns.\nPreprocessing includes missing-value treatment, extreme-value filtering, and feature standardization.\nPredicted returns are ranked to select stocks under daily rebalancing, minimum holding, and allocation constraints.\nThe text reports benchmark outperformance without numerical results, limiting independent assessment.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.