Industry-Specific Stock Forecasting with RNNs, LSTMs, and Multiple Factors
Summary
This study summary describes using recurrent neural networks, including long short-term memory cells, to forecast stock movements separately by industry. The rationale is that firms in one sector may share common influences, while industries can differ in their lifecycle and price behavior. Separate models are intended to capture those sector-specific patterns without blending influences across industries.
The modeling approach feeds all 69 available factors into the network rather than preselecting them manually, allowing the model to learn which inputs matter. Because the training data are monthly and relatively sparse, a separate validation set is used to monitor overfitting and limit unproductive training. The reported validation accuracy is around 40% for most industries, broadly similar to results from a pooled model. Performance varied by industry and training duration, and larger samples offered more room for validation accuracy to improve before overfitting. The summary gives no full experimental details, benchmark definitions, or evidence of live trading profitability.
Key ideas
- Separate industry models are intended to capture sector-specific effects and avoid mixing factors across sectors.
- The recurrent network uses temporal connections so prior information can influence later predictions.
- All 69 factors are provided to the model instead of being manually screened in advance.
- A held-out validation set is used to monitor overfitting with sparse monthly observations.
- Most industries achieved about 40% validation accuracy, similar to the pooled approach, with results varying by industry and sample size.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.