Picking a company’s peers is a routine step in valuation, pairs trading, index construction and risk modeling, and the default answer is GICS: one sector per firm, assigned mainly by its largest source of revenue. I tested a text-based alternative. The universe is the S&P 500 as of June 2026: 502 listings, or 499 companies after merging dual share classes. Each firm’s yfinance business summary is embedded with all-MiniLM-L6-v2, a 384-dimension model that runs locally with no API. 58.5% of the summaries are longer than the model’s 256-token window, so I split them into chunks, embedded each one and averaged the vectors. A firm’s peers are its ten nearest neighbors by exact cosine similarity in NumPy; as a check, I loaded the vectors into Qdrant and confirmed it returns identical neighbor lists.

The main result is about returns. Over 2025-07-01 to 2026-06-30, a firm’s mean daily-return correlation with its ten embedding peers is 0.335, against 0.275 with its GICS sector peers (difference +0.060, 95% CI 0.050 to 0.069, tested with a sign-flip permutation test and a bootstrap over firms). The embedding peers are the more correlated set for 72% of the 493 firms with enough price history, and net of the market the gap is +0.070 (95% CI 0.060 to 0.080). Two thirds of embedding peers are in the firm’s own sector, so this mostly reflects the embeddings picking the closest firms within a sector. I didn’t compare against GICS sub-industry peers, which would be the stricter test.

The embeddings also recover the sectors without being told what they are: 66.0% of nearest neighbors share the firm’s GICS sector, against 11.0% expected by chance (6.0×), tested against 10,000 firm-label permutations with Benjamini–Hochberg correction. Once sector size is accounted for, Utilities (14.9×) and Energy (14.8×) are the tightest sectors, and Industrials (3.5×) and Consumer Discretionary (3.6×) the loosest. Six cross-sector links are above chance, the strongest Energy → Utilities (3.7×).

Firms whose neighbors mostly sit in another sector (83 of them at a 0.7 threshold) show no co-movement advantage: −0.014 (95% CI −0.036 to 0.010). The embeddings give a finer peer set inside the GICS frame; they don’t show that GICS puts particular firms in the wrong sector.