Methodology

Player style map

Every player-season (2015-16 onward) placed on a 2D map by playing style — shot diet, role, rebounding and defense, self-creation, ball handling, shooting mode, touch location, and play types — via PCA for the underlying vector space, a Gaussian mixture for archetypes, and UMAP purely for 2D layout. Nearest-neighbor comps run as a brute-force cosine search entirely in the browser.

Loading eval data…

Limitation: the 30% hand-checked comp sanity rate reads low in isolation — it's measuring exact agreement with one author's subjective sense of a “good comp,” a much stricter bar than the map being generally sensible (most near-misses, like Curry↔Lillard or Giannis↔LeBron, are defensible archetype neighbors, just not the single comp the author had in mind).

Ask the box score

Plain-English questions are translated to a single validated SQL query by Cloudflare Workers AI (llama-3.3-70b-instruct, with 8B/3B fallbacks if the primary model is ever retired), then executed against real Parquet tables entirely in the browser via DuckDB-WASM. The same SQL guard (allowlisted tables, single SELECT/WITH statement, no file or network access, forced row limit) runs on both the Worker before caching a query and the browser before executing it.

Not yet measured: golden-set execution accuracy (`evals/ask/golden.jsonl`, ~60 questions across easy/medium/hard) and Worker CPU time per request. Both are flagged open in `docs/decisions.md` rather than reported with a made-up number — the model choice above is based on manual spot checks, not the full eval, and may change once that eval exists.

Shot quality court

An expected field-goal probability (xFG) model — what an average player would shoot from a given location, situation, and shot type — trained on every field goal attempt since 2015-16. A LightGBM classifier was compared against a logistic-regression baseline with distance/angle splines; the model isn't shipped directly, only a precomputed lookup grid the browser reads at 1-foot resolution per action family.

Limitation: no defender-distance data exists in public shot logs, so the model can't separate a contested layup from a wide-open one — layupis the one action family with a materially wider calibration gap than the rest.

Game flow

Win probability through every game since 2023-24, from play-by-play events plus a FiveThirtyEight-style Elo rating (K=20, +75 home-court points, 30% regression toward the mean each season) computed from the full 2015-16 onward history. No betting-line data anywhere in the pipeline.

Limitation: Elo is built from final scores only — no injuries, rest, or in-season roster changes. Possession is derived from event sequences, not exposed directly by the API, so it can occasionally misfire on unusual sequences (technical fouls, replay reviews).