Are AI Risks Priced in the U.S. Stock Market? Evidence from Financial News Factors

Battle Count: 75

The paper identifies a statistically significant risk premium (0.49%-0.57% monthly alpha) for high-minus-low D3-beta portfolios, which could inform factor-based trading strategies. However, the factor is non-tradable (news-based), the economic magnitude is modest, the paper does not construct factor-mimicking portfolios, and the D3 factor derives from a single LDA topic. The near-orthogonality to conventional factors (max |correlation| = 0.045) suggests incremental information value. The result is robust across specifications but limited to one domain out of four tested. Practical implementation would require real-time news processing and topic-model updating.

The pipeline involves multiple complex stages: (1) keyword-based article filtering from WSJ archive, (2) text preprocessing with NLTK and Gensim, (3) LDA topic modeling with model selection across 16 candidate K values and 5 seeds, (4) manual + AI-assisted article relevance review of 360 articles, (5) sentence embedding and cosine similarity mapping to 24 subdomains with null-distribution standardization and Sparsemax, (6) daily factor construction, (7) AR(1) innovation extraction with rolling estimation, (8) trading-day alignment, (9) firm-level beta estimation across 60+ specifications, (10) quintile portfolio formation and Fama-MacBeth regressions. Requires access to WSJ archive, CRSP, Compustat, French Data Library, and computational resources for LDA and embedding.

The paper provides detailed descriptions of keyword lexicons, LDA hyperparameters (alpha=1/K, eta=0.1, 10 passes, chunk size 2000, 200 iterations), topic selection criteria (K=18, seed 84), taxonomy mapping procedure, AR(1) innovation extraction, and all regression specifications. However, no GitHub repository or code is mentioned. The WSJ archive, CRSP, Compustat, I/B/E/S, and French Data Library are standard but require institutional access. GPT-5.6 Sol usage for article classification introduces a non-reproducible AI component. The MIT AI Risk Repository is publicly available.