Reproduce “Deep Sequence Modeling: Development and Applications in Asset Pricing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lin William Cong, Ke Tang, Jingyuan Wang, Yang Zhang
Deep sequence models are applied to conditional U.S. stock-return prediction using up to 12 months of lagged returns and firm characteristics. Across 1970–2016 data with out-of-sample testing from 1987 onward, gated recurrent models generally outperform a feed-forward neural-network benchmark: the LSTM reaches an out-of-sample R² of 0.45% versus 0.29% for the DNN, while attention-based models offer comparatively lower portfolio turnover and competitive Sharpe ratios.
Key insights
- Gated recurrence is more effective than basic recurrence for this setting: the RNN posts an R² of 0.13%, whereas the LSTM, GRU, and Bi-LSTM achieve 0.45%, 0.42%, and 0.44%, respectively.
- Predictive accuracy and investment performance do not produce identical model rankings; attention-based LSTM models have lower R² than the best LSTM but often deliver lower turnover and strong Sharpe ratios, especially after excluding micro-cap stocks.
- Greater architectural complexity does not guarantee better results: the Transformer performs weakly, with an R² of 0.12%, suggesting that model size and training demands must match the scale and structure of financial data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.