Strategies · Core
Filing Similarity and Stock Returns
Use how similar this year’s 10-K/10-Q language is to last year’s as a signal — boilerplate vs change as alternative data.
Definition
Compute textual similarity (cosine on embeddings or shingles) between consecutive filings. High similarity can mean stability or laziness; low similarity can mean a real change or a lawyer rewrite. The paper’s signed mapping is the strategy; yours must be tested.
Why it matters
It is a clean NLP factor on mandatory text. Vendor NLP moats are thin once the metric is public.
Case
A similarity drop that was a new risk-factor boilerplate the whole sector added. The signal was regulation, not a firm event.
How to read it
Horizon (annual is slow), sector controls, and a hold-out after transformers got cheap.
No query has been run yet. Which is tragically normal for most knowledge systems, but we are trying to evolve past decorative databases.