The Dispatch

Why raw data is a commodity and signal memory is not — and why Nike's nineteenth consecutive quarter of operating margin pressure is not in Bloomberg.

We embedded 150 million sentences from SEC filings into 768-dimensional space. HDBSCAN found 551 clusters. One stopped us cold: 3,282 sentences about nothing happening.

151 million sentences. 924,000 filings. 472 classifiers — 159 hand-designed, 313 discovered by machines. Here is what a decade of reading every SEC filing taught us about predicting stock moves: why AUC lies, why trajectories beat events, and why 4 of our 5 most predictive signals were concepts nobody thought to look for.

The SIC system was designed in 1937. GICS in 1999. Both classify companies by what they sell. We classified 3,000 companies by what they say — 469 binary classifiers across 150 million SEC sentences — and found clusters that predict stock behavior better than any existing taxonomy.