Methods · the metrics canon
What every number on this site is, where it comes from, and what we deliberately do not render. One definition per metric, enforced at the bake — this page is generated from the canon doc, so it cannot drift. See also the reads index for how model claims are judged.
METRICS — the canonical dictionary
What renders *everywhere*, in what form, with which splits — one dictionary so two tables never show the same stat under two names. Sources: ESPN box spine (§0-validated), balldontlie categories + v2-advanced (§0b-validated), and the KB glossaries (NBA.com stats definitions; closingfive; Cleaning the Glass conventions). Judgment-color rule everywhere: pos/neg against the season's own league bands (baked p10/25/50/75/90), flag for partial or small samples — never both at once.
1. The everywhere-set (basic; every player/team surface)
Shown per-game AND per-36 (per-game primary, per-36 as the context column; "small sample" flagged under 15 GP / 250 min):
| metric | form | source |
|---|---|---|
| GP, GS | counts | ESPN box (GS null on BDL seasons) |
| MIN | per-game | both |
| PTS, REB, AST, STL, BLK, TOV, PF, +/- | per-game, 1dp | both |
| OREB, DREB | per-game | both (board rows + shoot block) |
| FG / 3P / FT | m–a and % | both (player docs shoot, board fgp/tpp/ftp) |
Team tables carry the same set minus GS/PF, plus PTS/opp PTS and point differential (colored pos/neg at ±0 — wins do that job).
2. The derived canon (advanced; one definition each)
| metric | definition | rendered |
|---|---|---|
| eFG% | (FGM + 0.5·3PM)/FGA | player page, /stats (advanced block) |
| TS% | PTS / (2·(FGA + 0.44·FTA)) — BDL's calc | advanced tables, chips |
| USG% | BDL usage_percentage (est. version) | advanced tables (neutral — load, not grade) |
| AST% / ORB% / DRB% / REB% | BDL percentage fields | advanced table |
| ORtg / DRtg / NET | BDL ratings (minutes-weighted roll-up) + est. variants | advanced table, /stats |
| PIE | BDL pie | advanced table, /stats |
| Pace / est. pace | BDL | doc only (context, not rendered as grade) |
| Tracking per-36 | deflections, contested, charges, box outs, screen AST (+pts), loose balls, passes, touches, 2nd AST, FT AST, reb chances, blocks against | Defense & activity panel |
| Matchups | minutes matched, opp FG% / 3P%, rim FG% allowed (lower better) | Defense & activity panel |
| Conversion-of-chances | reb36 ÷ reb chances per 36 | defense panel chip — context, not skill (explorations §4) |
3. The advantage framework (where a player has an edge)
| split / reference | status | where |
|---|---|---|
| vs league bands (percentile) | live | percentile chips, advanced tables, /stats |
| by shot zone vs league FG% (±3 / ±10 pts coloring) | live | shot zones table |
| shot-making vs expected (xeFG) — RESULTS §9 | live | zones panel chips |
| vs play-type league percentile | live | play type bars (BDL percentile baked) |
| defensive matchup quality | live | defense panel |
| vs position pool (G/F/C percentiles) | next — position comes from board pos; pools need a floor of ~40 players | — |
| rest days (0/1/2+), B2B | computable now from game-log dates; next bake | — |
| home/away, W/L, vs .500+ teams | computable from game logs + standings; next bake | — |
| by month / pre-post deadline (traded players) | computable from game logs (BDL rows carry team changes implicitly via team) | — |
| clutch | live (deltas, gp-flagged) | clutch panel |
| schedule strength / opponent-adjusted scoring | needs opponent ratings (Track A/B) | MODELS-ROADMAP |
4. Splits canon (what "everywhere" means)
Every player-season table eventually carries these cut dimensions, built from the game-log grain (the only honest place to cut):
1. Season (the unit everywhere). 2. Venue: home/away. 3. Result: wins/losses (descriptive context, never a grade). 4. Rest: 0 days (B2B), 1, 2+. 5. Opponent tier: vs winning / losing records (standings join). 6. Role: starter/bench; team tenure (games 1–15 post-trade vs rest). 7. Context windows: last 10, last 30 (rolling form for live seasons). 8. Possession contexts: overall vs clutch (BDL categories); by play type; by zone.
Rules: splits render only from the game-log grain (never aggregated-then-cut); any split cell under 10 games gets the flag color; a split never colors a grade until its persistence is tested (the explorations lesson).
5. What we deliberately do NOT render
- Zone-rebound "skill" (conversion is context — explorations §4).
- Cohort aging curves as trajectories (survivorship — explorations §5).
- Clutch deltas without the gp flag.
- Any rating without a RESULTS-HOOPS registration behind it.
rendered from docs/METRICS.md at bake — the metrics canon