New QuantLet Submission: Narrative-Risk-AMLEF

[HoffmannDennis]

Repository Details

• Repository: https://github.com/HoffmannDennis/Narrative-Risk-AMLEF
• Author: Dennis Hoffmann
• Institution: University of Twente — PhD Programme
• Category: Empirical Finance / Machine Learning / Cross-Sectional Asset Pricing

Description

I would like to submit a collection of 10 educational QuantLets demonstrating LLM-extracted
narrative-risk factor methodology on US large-cap 10-K Item 1A risk-disclosure prose (top-30 S&P
500 constituents by market cap, snapshot 2014-12-31, fiscal years 2015-2024).

Each quantlet isolates one methodological axis (signal extraction, sort granularity,
risk-adjustment basis, formation lag, industry neutrality, rolling-window estimation, long-only
construction) so readers can see how each design choice affects the headline alpha. Headline
numbers are explicitly indicative not inferential at N=30 firms.

Compliance Checklist

✅ All 10 modules have properly formatted Metainfo.txt files with:

• Name of Quantlet
• Published in: "AMLEF Applied Machine Learning in Empirical Finance"
• Description (detailed)
• Keywords (comprehensive)
• Author
• Submission date
• Datafile specification
• Input / Output

✅ Repository structure follows QuantLet conventions:

• Each module in separate directory at repo root
• Module names follow NN_NarrativeRisk_<Axis> pattern
• Clear, descriptive naming
• Name of Quantlet matches each notebook filename exactly

✅ Top-level README.md at repository root

✅ All code is self-contained and reproducible:

• Each quantlet ships its frozen LLM-extracted corpus (corpus.ndjson.gz) + provenance.json
• make demo fetches third-party data (Yahoo prices, Ken French factors) on first run, then
  executes the notebook end-to-end
• Headline figure (PDF) and summary CSV bundled per quantlet

✅ Educational focus with executable Jupyter walkthroughs

Modules Included

1. 01_NarrativeRisk_LLM_DisclosureScoring — Tercile-sort spread + FF5+UMD alpha (canonical)
2. 02_NarrativeRisk_LM_DisclosureScoring — Loughran-McDonald dictionary baseline
3. 03_NarrativeRisk_LLM_DecileSort — Decile vs tercile granularity
4. 04_NarrativeRisk_DeltaSeverity — Year-over-year change in composite severity
5. 05_NarrativeRisk_FF3_FF5_Comparison — FF3 vs FF5+UMD risk model
6. 06_NarrativeRisk_LLM_SeverityDistribution — Corpus inspection (no portfolio)
7. 07_NarrativeRisk_LLM_PostFormation — One-month skip-formation lag
8. 08_NarrativeRisk_LLM_IndustryNeutral — Within-sector demean
9. 09_NarrativeRisk_LLM_RollingBeta — 36-month rolling FF5+UMD beta
10. 10_NarrativeRisk_LLM_LongOnly — Long-only T3 portfolio (drops short leg)

Educational Value

• Designed for empirical-finance researchers and graduate students
• Methodological-axis design: each quantlet varies one design choice while holding others constant,
  so readers see the contribution of each axis in isolation
• Quantlet 06 is inspection-only — encourages reading the input distribution before consuming
  derived statistics
• All metrics carry "indicative not inferential" warnings (honest small-sample framing at N=30)
• Fully audited LLM-extraction provenance shipped per quantlet (model SHA, prompt SHA, corpus
  SHA-256)

Request

Please consider:

• Adding this repository to the QuantLet organization
• Or advising on the appropriate category/target repository for these empirical-finance educational
  materials
• Any feedback on structure or formatting improvements

Thank you for maintaining this valuable educational resource!