Relevance
8/10
Highly relevant: directly addresses predicting announcement-day stock returns using both hard (earnings surprise) and soft (press release text) information. Demonstrates market efficiency (no excess returns at market open) but shows predictive advantage if information leaks. The 56% accuracy in identifying top-10 performers with combined signals is actionable for event-driven strategies. The online learning framework supports real-time deployment.
Implementation Complexity
7/10
Moderate-to-high complexity: requires multiple NLP pipelines (LDA, BERT embeddings, FinBERT, MPNET), rolling window training, Lasso regression, SHAP computation, GPT-4o API calls for topic labeling, TAQ data processing for bid-ask filtering, and Fama-French factor models for alpha testing. The combination of five feature extraction methods and the online learning framework adds engineering overhead.
Reproducibility
3/5
The paper uses publicly available SEC Edgar press releases, but CRSP daily stock returns, LSEG IBES analyst forecasts, and NYSE TAQ intraday data require institutional subscriptions. The rolling window methodology and hyperparameters (lambda=10^-5, oLDA parameters) are specified. GPT-4o is used for topic labeling, which introduces non-determinism. No code repository is mentioned.