Equity2Vec: End-to-end Deep Learning Framework for Cross-sectional Asset Pricing
Equity2Vec: End-to-end Deep Learning Framework for Cross-sectional Asset Pricing
The paper introduces Equity2Vec, an end-to-end deep learning framework for cross-sectional asset pricing that addresses two key limitations of prior work: overlooking cross-sectional effects and underutilizing heterogeneous data. Equity2Vec learns stock embeddings from news co-mentions using matrix factorization on a global co-occurrence matrix to capture long-term static relations, and a temporal graph with attention mechanism to capture evolving dynamic relations. These embeddings are combined with technical factors and news signals (via Word2Vec) as inputs to an LSTM with temporal attention for predicting future returns.
Experiments on a large Chinese stock market dataset (3,600 stocks, 2009-2018) show that Equity2Vec outperforms state-of-the-art methods (SFM, HAN, AlphaStock, ARRR, VAR) in terms of correlation, t-statistic, and trading profitability. Ablation studies validate the contribution of each component, and interpretability analysis reveals that learned embeddings align with industry sectors and temporal attention weights prioritize recent days. The framework effectively monetizes cross-sectional signals, demonstrating significant practical value for quantitative trading.
Highlights
- 1Proposes Equity2Vec, a graph-based component that learns stock embeddings from news co-mentions to capture both long-term static and evolving dynamic cross-sectional relations.
- 2Develops an end-to-end deep learning framework that integrates heterogeneous data sources: stock embeddings, technical factors, and financial news signals.
- 3Outperforms state-of-the-art baselines (SFM, HAN, AlphaStock, ARRR, VAR) on real-world Chinese stock market data (3,600 stocks, 2009-2018) in terms of correlation, t-statistic, and trading PnL.
- 4Demonstrates interpretability of learned embeddings aligning with Barra industry sectors and temporal attention weights favoring recent days.
Methods
- MMatrix factorization on a global stock co-occurrence matrix to extract static stock embeddings.
- MTemporal stock graph construction with exponential moving average and k-nearest neighbor attention mechanism to capture evolving relations.
- MLSTM with temporal attention for sequential modeling of combined stock embeddings and dynamic features (technical factors and news vectors).
- MAblation study and comparison with variants (NoEmd, PoorMan, SectEmd, OneEmd, RadiusEmd) to validate components.
Results
- REquity2Vec consistently achieves higher correlation and t-statistic than all baselines across multiple out-of-sample years (2016-2018).
- RMarket trading simulations show Equity2Vec generates higher cumulative PnL in both long-short and long-index portfolios.
- RThe optimal number of neighbors k in the attention mechanism balances performance; too few or too many neighbors reduce correlation.
- RAblation study confirms each component (global embedding, temporal graph, attention) contributes substantially to overall performance.
- RLearned stock embeddings visualized via t-SNE align well with Barra industry sectors, indicating interpretability.
Analyze Paper
Generate insights from "Equity2Vec: End-to-end Deep Learning Framework for Cross-sec...".