Week 6 — Reading the Frontier II: Text, Modern Empirics, and the AI Co-Pilot

Text-as-data and machine-learning-flavored empirical finance, plus a full module on using LLMs responsibly as a research co-pilot. See ../../TOC.md for the full plan.

Meta-pack

  • Reading Guide Pack 6 + AI Lab Manual — reading measurement papers, a RAG architecture, a prompt-pattern catalog, an LLM-label evaluation harness, and a responsible-use/disclosure checklist.

Reader’s Guides + the AI chapter

  1. Ch 6.1 — Kogan, Papanikolaou, Seru & Stoffman (2017), QJE 132(2):665–712 (market-based patent value)
  2. Ch 6.2 — Hoberg & Phillips (2016), JPE 124(5):1423–1465 (TNIC; 10-K cosine similarity)
  3. Ch 6.3 — Loughran & McDonald (2011), JF 66(1):35–65 (finance sentiment dictionaries)
  4. Ch 6.4 — Bartlett et al. (2022) + Bhutta–Hizmo–Ringo (paired) (fair lending in the algorithmic era; ties to Sun & Gao 2019)
  5. Ch 6.5 — The AI Co-Pilot for Research (LLM-in-the-Loop) — prompt patterns, RAG over 10-Ks, OOS-validated classification, critical limits, the Anthropic + GMU Azure APIs (env-var keys only)

Notebooks (../../../notebooks/week-06/)

  • nb6.1 Patent-value panel (event-study) · nb6.2 10-K text vectorization & cosine similarity (sklearn) · nb6.3 Loughran–McDonald sentiment pipeline · nb6.4 Mortgage-disparity decomposition (Blinder–Oaxaca; over-controlling trap) · nb6.5 AI co-pilot lab (RAG + OOS-validated classification; runs offline with a local fallback, env-gated live Anthropic/Azure cells, no hard-coded secrets)
  • All verified to run headless.

Problem sets (solutions in Appendix E)

  • PS 6.1 Patent-value · PS 6.2 Cosine/TNIC · PS 6.3 LM vs. naive dictionary · PS 6.4 Fair-lending decomposition · PS 6.5 LLM classification + held-out validation + leakage audit

Mentor, assessment