Deep Learning for Stock Factors: Embeddings and Market-Aware Prediction
Summary
This overview discusses two research approaches to using deep learning for equity factor discovery. The first builds stock embeddings from co-investment patterns in fund portfolios: random walks on a stock-fund graph produce sequences for skip-gram training. A small rescaling network then assigns stock-specific weights to technical indicators, with information coefficient as its objective and periodic updates to reflect changing conditions. The article reports tests on historical stock data and notes that simple normalization did not improve raw indicators, while a more complex neural network showed signs of overfitting.
The second approach uses matrix factorization of fund holdings to infer static stock characteristics, then combines them with dynamic factor representations. Market state is estimated from recent winners, while an LSTM models changing market trends; the resulting signals feed a prediction model trained with regression and ranking losses. Reported tests favor market-aware integration over direct concatenation and baseline sequence models. The authors caution that holdings also reflect trends and risk management, so long histories are needed to better isolate intrinsic stock attributes. Results are historical and do not establish live trading performance.
Key ideas
- Fund co-investment patterns can be converted into stock embeddings using graph walks and skip-gram training.
- A lightweight rescaling model can tailor technical-indicator weights to individual stocks.
- Factor normalization and model complexity can affect results, with complex models vulnerable to overfitting.
- Fund-holdings matrix factorization can provide static stock representations for dynamic prediction models.
- Market state and trend signals can help integrate stock characteristics with time-varying factors.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.