Blog / 28 Sept 2026
Systematic Alternative Data Research with Language Models
A working paper on how language models can find alternative data, reconstruct what was knowable at the time and test whether a company relationship is genuinely useful.
Alternative data is easy to chart and hard to trust. A customs record, public industry series or product activity measure may seem to move with a company result, but entity errors, changing coverage, pricing, reporting lags or hindsight can create a convincing relationship that means very little.
Systematic Alternative Data Research with Language Models describes a research system that searches for potentially relevant datasets, reconstructs what was knowable at the time and subjects each relationship to the same company-specific checks an analyst would use. It separates language-model reasoning from deterministic data work, preserves the definition and timing of every source, and investigates the periods when a relationship breaks. The paper works through three examples: the U.S. housing pipeline and IBP’s single-family installation revenue, RV wholesale shipments and THOR’s North American Towable sales, and dairy commodity prices alongside Saputo USA reported sales.