Rating
1861
Battle Count: 82
Relevance
9/10
Directly addresses when and how news is priced in, provides a per-tag drift table usable as a prior in forecasting models, demonstrates that news direction is largely spent by publication while width/volatility effects persist, shows that naive news-fading strategies are indistinguishable from shorting all covered small caps, and provides concrete economic significance estimates (Sharpe ratios, annualized returns) for calendar-time portfolios. Highly actionable for news-driven trading systems.
Implementation Complexity
7/10
The pipeline involves: (1) LLM teacher labeling with structured JSON prompts, (2) distilroberta-base fine-tuning with multi-head architecture (7 heads), (3) active learning with uncertainty sampling and rare-tag quotas, (4) embedding-based clustering with per-tag lifetime calibration, (5) rolling OLS beta estimation, (6) cluster bootstrap with 5,000 draws, (7) placebo-adjusted abnormal return computation. Requires access to large-scale financial news data, GPU for model training, and careful event-study methodology. The inference runs at 125-160 articles/sec on a MacBook Pro M4 Pro.
Reproducibility
3/5
The methodology is described in extensive detail (taxonomy, distillation pipeline, clustering thresholds, bootstrap parameters, beta estimation). However, the NewsWitch corpus is proprietary (commercial product from Zanista AI), and the GPT-5-mini teacher is Azure-hosted. The distilroberta-base student architecture and training procedure are specified. Contact info provided for data access. The 2023 backfill and live production setup limit full replication.
About this paper
Methodology: Distilled LLM Tagging with Placebo-Controlled Event Study. Problem types: Classification, Active Learning, Clustering, Natural Language Processing, Transfer Learning, Dimensionality Reduction Techniques (PCA, ICA, t-SNE, UMAP).
The interactive Everscope explorer (charts, battles, favorites) loads below.