Predicting Stock Price Direction on Earnings Announcement Days using Multi-modal Deep Learning

Battle Count: 190

The paper directly addresses event-driven trading around earnings announcements, a critical window for quantitative strategies. The multi-modal feature engineering (fundamentals + technicals + NLP sentiment) and the asymmetric cost framework are practically relevant for trading system design. However, the modest predictive performance (Macro-F1 ~0.39, accuracy ~44%) and limited data window reduce immediate deployability. The LSTM's conservative strategy and Transformer's sensitivity trade-off provide useful guidance for risk-managed signal generation. The work is more of a proof-of-concept than a production-ready trading system.

The architecture is moderately complex: a 2-layer LSTM (hidden dim 64) and a 2-layer Transformer (4 heads, FF dim 256) with standard PyTorch implementations. The multi-modal feature engineering pipeline (FactSet data extraction, FinBERT inference, temporal alignment, imputation, z-score normalization) adds significant data engineering complexity. The weighted cross-entropy loss and custom cost metric are straightforward. Overall, the modeling is accessible but the data pipeline requires proprietary data access and careful temporal alignment.

The complete implementation including preprocessing scripts, model architectures, and training logs is available on GitHub. Models are implemented in PyTorch with specified hyperparameters (Adam optimizer, lr=5e-5, batch size=8, 15 epochs, dropout=0.5). However, the underlying data sources (FactSet fundamentals, Markets API news feed) are proprietary and not publicly available, limiting full end-to-end reproducibility. The ETH Student Cluster environment is also not publicly accessible.