Methodologies

Jump: Eligibility · NHLe · FlameWire Calc · Team Blueprint · DirtyXg · Data · Limits

NHL draft eligibility (age window)

The lab shows an Eligibility column based on the standard NHL Entry Draft age rules (practical approximation for scouting):

• Minimum: player is 18 on or before September 15 of the draft year.
• North American: not older than 20 on December 31 of the draft year (age on Dec 31 ≤ 20).
• International: same floor; max age 21 on Dec 31.
• Players past the max age who were never drafted follow the undrafted free agent (UFA) path — they stay “undrafted” in the draft sense.
• Players already matched to an NHL draft class are labeled Drafted.

Region (NA vs international) is inferred from leagues / hometown when possible; this universe defaults to North American. After July 1 we evaluate the next calendar draft year (e.g. Aug 2026 → 2027 draft).

NHLe (NHL equivalency)

NHLe translates points-per-game from other leagues into an NHL-scale rate using public league factors (classic NHLe tradition). Headline number is NHLe/82 — expected NHL points over 82 games if the translated rate held.

• Multi-year blend weights: 0.50 · 0.30 · 0.20 (most recent first, up to 3 seasons).
• Minimum ~15 GP per translated season (with small allowances for NHL cups).
• Youth / showcase leagues and Memorial Cup rows are skipped.
• Role ladder (ultimate / most likely / floor) is offense-led with age and NHL sample clamps.

League factors (v1)

Junior A factors are intentionally conservative. Scoring-leader lists over-sample offense and can inflate rates versus full rosters.

FlameWire Calc

Proprietary upside score on a 0–100 “OVR-style” scale:

FW = 0.40·P + 0.30·D + 0.30·A (+ agreement · defense boost · young-elite · soft tiers; v0.9)

• P (production) — NHLe blend + trend + confidence + league-relative context. Confidence dings use schedule-relative sample quality (NCAA ~36 GP season treated as full; not raw GP vs CHL ~68). Relative = PPG percentile vs peers, with fallback chain: same age → age band (≤17 / 18 / 19 / 20+) → league+pos → league. Age from season row or birthdate. Season series (peak / GP-weighted mean / trend, age-adj + plain) is on the player payload. Score impact stays neutral on draft-day destiny CV (age-adj level did not beat NHLe+pick for hard tier accuracy on eventual NHLers).
• D (draft prior) — smooth exponential pick chart 100·e−0.011·(pick−1) with years-since-draft decay (Schuckers / pick-value chart family). Undrafted players use an eligibility ladder: ~42 (1st-yr) · ~32 (2nd) · ~24 (3rd/final) · ~16 (aged out / UFA) · ~22 (unknown); pre-eligible ~38.
• A (age / path) — prospect window + path tier + NHL GP sample (draft-day destiny cares a lot about development clock).
• P↔D agreement — small soft term when production and draft capital align (or diverge). Encourages coherent profiles over mixed signals.
• Defense destiny boost — +6 for defensemen. Box-score NHLe understates D role outcomes at the same points rate; draft-day CV preferred this over F-only cutpoints.
• Young-elite ceiling — small age-gated boost when age ≤19–20 with top capital and real production, so Star path is reachable.
• Destiny objective: ordinal role path at draft day (see Regressions and Stress Testing), not exact NHL points. Cuts jointly maximize within-1 accuracy then exact tier hit.
• Soft tier probabilities (v0.9) — softmax over distance to band midpoints (temperature 15, Brier-tuned on draft-day set). Headline path stays hard cuts; soft %s express nearby-tier uncertainty (e.g. 48% middle-6 / 31% top-6). Hard one-hot Brier ~1.26 → soft ~0.75 on the destiny sample — better calibration, same ranking score.
• Hurdle two-stage (v0.9.1 side channel) — Liu/Schuckers-style P(regular NHL path) × role intensity, fit on draft-day history (Data/cache/hurdle_destiny_model.json). Shown as hurdle.p_make + hurdle soft mix. Does not replace the PDA hard path: on 5-fold / year-holdout destiny CV, hurdle hard tiers lost ~4–5pp within-1 vs FW. Paper metric (Spearman with first-7-year NHL GP) is essentially tied with FW (~0.47 vs draft order ~0.45).

Tiers (score floors): Star ≥96 · Top-6 / Top-4 D ≥74 · Middle-6 / 2nd pair ≥62 · Bottom-6 / 3rd pair ≥41 · Bubble ≥26 · else AHL path. Compared side-by-side with NHLe role ceiling.

Accuracy lab results

Draft-day destiny back-test: drafted skaters, age ≤21 at draft, ≥40 NHL GP later, seasons as-of draft only (n≈365). Metrics: exact tier hit and within-1 tier. Live baseline under v0.6–v0.8 PDA blend: ~37% exact · ~87% within-1 (bootstrap 95% CI roughly exact 32–42%, within-1 83–90%).

Literature-inspired formulas tested with 5-fold CV (pure data-only features — NHLe, pick, age/path, CS rank, relative peers, size, F/D):

• HockeyStats / Schuckers family — NHLe + log(CSS) logistic / linear blends; pick-heavy vs NHLe-heavy priors.
• Liu model-tree lite — separate PDA weights by pick band (1–15 / 16–32 / 33–64 / 65–100 / 100+).
• PCS-lite — peer mean destiny in pick × NHLe × position cells.
• Ridge → p82 → role map (and F/D-split ridge), ordinal multi-threshold, one-vs-rest tier classifier, two-stage zero-inflation.
• kNN, ensembles, active relative scales, age-residual relative, size residuals, CS blends into D, F/D-separate weight grids, cut re-optimization.

Result: nothing beat the current PDA blend by more than noise. Near-misses (agree scale 1.5, light CS blend 10–35%) moved within-1 or exact by ~0.3–0.9pp — well inside bootstrap CI, so hard score weights stay put. Aggressive relative / PCS / ridge / tree variants usually hurt within-1 (collinearity of CS≈pick≈0.91; relative≈NHLe on eventual NHLers).

Practical ceiling for 6 hard destiny tiers on draft day with public box data is about 35–40% exact / 85–88% within-1. Targets of 50% exact / 90% within-1 need either fewer tiers, post-draft development seasons, scouting text features, or a different label (e.g. binary NHLer / star as in public logistic models). Soft probabilities are the honest way to express the residual uncertainty without inventing hard precision.

From-scratch formulas without NHLe

We also rebuilt destiny models that never use league translation factors (no NHLe), following public draft research:

• Schuckers / Liu — raw GP/G/A/PPG, age, size, CSS, log(pick), path flags; ridge → NHL P/82 → role map; ordinal multi-threshold; one-vs-rest tier classifier.
• PCS-style cohorts — mean observed destiny in bins of age × raw PPG × height × pick × position.
• Hand blend — in-league relative %ile + age-weighted raw PPG + draft capital + age window (grid-searched weights).
• kNN / ensembles on the same non-NHLe feature set.

On the same draft-day set (n≈365), none beat FlameWire on within-1 (best non-NHLe ~85% vs FW ~87%). One ridge→P/82 path hit slightly higher exact (~38% vs ~37%) while losing ~4pp within-1 — not a joint win. Pure skill (no pick/CS) collapses (~28–37% exact / ~77–80% within-1). So NHLe is not the bottleneck: capital + multi-league context already do most of the ordinal work; dropping NHLe does not unlock a new accuracy regime on this label.

NHL draft status

Draft status is matched against public NHL draft classes from api-web.nhle.com/v1/draft/picks/{year}/all (default years 2005–current). Matching is by normalized full name, with disambiguation on position, age-at-draft, and amateur league when multiple picks share a name.

• Matched players get year / overall / round / team and a real FlameWire Calc D prior.
• Unmatched players stay undrafted (UDFA prior).
• Cache: Data/cache/nhl_draft_index.json. Refresh via UI link or POST /api/sync-draft?force=true.
• Common-name collisions are possible — confidence is marked medium when candidates are close.

NHL Central Scouting rankings

Public midterm + final prospect rankings from api-web.nhle.com/v1/draft/rankings/{year}/{category} (NA skaters = 1, international skaters = 2; years ~2008–current). Cached under Data/cache/nhl_cs_rankings.json.

• Matched onto the universe by name (+ DOB / position / draft year) similar to draft picks.
• On the draft-day destiny set (~93% CS coverage), CS final rank Spearman with destiny (~0.54) is close to pick (~0.54) and below full FW (~0.58). CS is nearly collinear with pick. Light CS→D blends (10–35%) were noise- level on CV; drafted D stays pure pick chart.
• Undrafted but CS-ranked players use CS final rank as D when it exceeds the eligibility ladder (data-only capital for ranked UDFAs).
• Height/weight from CS backfill missing bio when available.

Ages & birthdays

Missing ages are filled without HTML scraping, in this order:

1. Spreadsheet — Age column when present.
2. CHL LeagueStat JSON (OHL / WHL / QMJHL) — full Tier-1 skater boards from 2010–11 onward (cached under Data/chl/seasons/). Merged into player bios (DOB, height, weight, shoots, hometown) and multi-year season rows (GP/G/A/P/+/−/PIM) for NHLe. Refresh: POST /api/sync-chl-dump or python -m app.chl_dump --from-year 2010.
3. NHL API — name search + player landing for official birthDate.
4. NCAA D1 (College Hockey News) — primary college source: full team rosters (DOB, height, weight, class, hometown, draft note) plus full-season overall scoring (~1.3k–1.6k skaters/year), seasons 2010–11 → current including 2022–23. Cached under Data/chn/. Refresh: POST /api/sync-chn or python -m app.chn_dump --from-year 2010.
5. NCAA D1 leaders (ncaa.com) — optional gap-fill via ncaa-api.henrygd.me (leaderboards only; no bios). Data/ncaa/; POST /api/sync-ncaa.
6. NCAA class estimate — Fr/So/Jr/Sr/Gr + season year → approximate age (synthetic July 1 birthday). Marked with ≈. Directional only; CHN DOB preferred.

Cache: Data/cache/birthdays.json. Refresh via POST /api/sync-chl-birthdays or POST /api/sync-birthdays.

Team Blueprint

Team Blueprint diagnoses NHL clubs against historical success ladders (regular-season tiers + playoff depth) using official NHL standings/summary rates, then layers NHL EDGE process metrics from season 2021–22 onward:

• DirtyXg / DirtyXgA per GP — chance quality for and against (v4 logistic; 5v5 + all-situations).
• DirtyXg% (5v5) — even-strength chance share (preferred process read; PP/PK less muddying).
• DirtyXg% (all) — all-situations share for full-game context.
• G − DirtyXg — finishing residual (5v5 and all-sits).
• HD SOG / GP and HD SOG share — high-danger volume and mix.
• OZ% (ES) — even-strength offensive-zone time share.
• CF diff (EDGE) — signed shot-attempt differential from EDGE (not a 0–1 Corsi%; higher = more attempts for).

EDGE dumps cache under Data/blueprint/edge/. Pre-2021–22 seasons keep counting stats only (no EDGE fields).

DirtyXg (experimental) vs public xG

Expected goals (xG) in full public models (e.g. MoneyPuck, Natural Stat Trick, Evolving-Hockey) score every unblocked shot attempt with many features — distance, angle, shot type, rebound/rush, score effects, etc. — trained on large multi-season samples.

DirtyXg is Hockey Scout’s independent chance-quality estimate (FlameWire Lab lineage). Current Blueprint default is v4: a pure-Python L2 logistic fit on NHL play-by-play, trained on goals (not MoneyPuck coefficients). It is not a MoneyPuck/NST feed — think of it as a lab process metric for Blueprint ladders and finishing residuals. MoneyPuck remains a side-check only.

Formula (v4 preferred)

v4 (current): every unblocked PBP attempt (goal + SOG + miss) gets P(goal) from a logistic model with continuous distance/angle (plus squares and interaction), rebound (≤3s since prior shot event), rush (≤4s after NZ/DZ prior event), 5v5 flag, log time-since-last-event, and shot-type one-hots (wrist baseline). Empty-net and miss flags are not features (EN exploded the intercept; miss is label leakage — misses never score). Coefficients live in Data/blueprint/pbp/dirty_xg_v4_model.json. Train calibration is fit so Σ predicted ≈ Σ goals on the training sample (no constant post-hoc scale).

DirtyXg_for = Σ logistic(xi) · 1.0 · same sum for against · method tag pbp_v4

Lanes: every unblocked attempt is scored once, then rolled up as all-situations and 5v5 only (situationCode = 1551). Blueprint ladders and diagnosis prefer DirtyXg% (5v5) when present — cleaner even-strength process — and still show all-sits for full-game context.

v3 (prior PBP): unblocked attempts mapped to EDGE bands, scored with EDGE league conversion rates + shot-type multipliers + miss discount, then ×0.684 so league Σ DirtyXg ≈ Σ goals.

DirtyXg_for = Σ xG_band(unblocked) · 0.684

v2.1 (EDGE-only fallback): season SOG counts by EDGE band high / mid / long (+ residual), then ×1.048:

DirtyXg_raw = N_high·r_high + N_mid·r_mid + N_long·r_long + N_other·r_other

DirtyXg = DirtyXg_raw × 1.048

Typical EDGE rates: high ≈ 0.19 · mid ≈ 0.11 · long ≈ 0.03 · other ≈ 0.06. Blueprint prefers PBP v4 when the model + season rollups are present; EDGE rollup remains as dirty_xg_edge for comparison. Pre-EDGE / no-PBP seasons keep counting stats only.

Spot-check vs MoneyPuck (v4)

We compared DirtyXg v4 to MoneyPuck team xGoalsFor / xGoalsAgainst (situation = all) using the official MoneyPuck data downloads (2023–24 and 2024–25 regular season CSVs) — not HTML scraping. Credit: MoneyPuck.com. Non-commercial research use. Goal is accuracy across the board, not cloning MP; MP is a reference for rank/process fidelity.

Ladder sanity (DirtyXg%, EDGE-era RS): Presidents Trophy ~53% · playoff teams ~52% · missed playoffs ~49% · Bottom 5 ~45%. Example: CGY 2024–25 ≈ 225 DirtyXg vs 220 goals (slight underperformance residual).

Read: v4 tracks public chance creation closely enough for Blueprint process ladders, share metrics (DirtyXg%), and finishing residuals. Remaining gaps vs full commercial models are mostly score/venue adjustments, flurry effects, and richer sequence context — not basic location.

Lineage — what each version adds

Inspired by the public expected-goals tradition (MoneyPuck and peers). DirtyXg is an independent lab estimate — not affiliated with or endorsed by those sites. Always credit MoneyPuck.com when citing their downloadable data.

Data source

Universe is built from the full CHL LeagueStat dump (2010+), College Hockey News NCAA D1 (2010+), and Data/Hockey_Scoring_Leaders_Master.xlsx (other leagues / sheet fills). Players are merged by name (and CHN/CHL ids when present) for multi-year NHLe. Team Blueprint uses official NHL APIs (standings, team summary, playoff carousel) plus NHL EDGE team endpoints. Hit POST /api/reload (or restart) after refreshing dumps; Blueprint/EDGE via POST /api/sync-blueprint or Refresh on /blueprint.

Limits

• CHL / sheet scoring boards can still omit non-scorers; CHN rosters are full-team.
• CHN is a personal research scrape of public pages — be polite; do not republish wholesale.
• Goalie NHLe is not applied (same as FlameWire).
• A minority of roster rows lack DOB (esp. commits / early listings).
• DirtyXg is an experimental process metric — not a substitute for full public xG models (no score/venue/flurry adjustments).
• DirtyXg v4 needs cached PBP (EDGE era ~2021–22+); older seasons fall back or omit process xG.
• NHL EDGE history for Blueprint starts ~2021–22; endpoints are undocumented and can change.