Value Score methodology

What this is not

Not an official CBA salary-cap sheet. Team payroll is season-roster AAV from curated/import contracts only — no Spotrac scrape, no LTIR / buried / retention / buyout math. Estimated placeholders are gaps, never silent filler in the total.

What we answer

Is this player worth what they're paid — as a share of the cap, at today's ceiling and at the next-deal / FA-summer ceiling as the cap moves up? Contract AAVs are inputs. The product is worth (%cap) vs cost, with tier-matched market anchors as context when a clearing / elite path applies.

Same-season window: 2025-26 MoneyPuck production, NHL season rosters (API /roster/TEAM/20252026, not /current), and 2025-26 contract AAVs. Transparent TypeScript heuristics — not a trained Python model.

Plain English — what drives fair $

Fair AAV is "what would a reasonable mid-market deal look like for this production?" We blend a few seasons of rates, convert that into dollars, then compare to the actual paycheck in % of the salary capso older deals aren't punished for inflation.

Forwards (C / LW / RW)

• Scoring pace — adjusted individual xG and primary points (goals + primary assists), paced to 82 games. Secondary assists and high-danger xG add a bit.
• Minutes — TOI/game scales the offense package (top-line ice ≠ depth ice). Centers get a small bump; wings with real faceoff volume can count as centers for FO credit and comps.
• On-ice impact — 5v5 relative xG% and on-ice xGF/60 (are they tilting the ice?).
• Special teams & two-way — PP xG + PP/PK minutes; penalty differential, takeaways vs giveaways; faceoffs for centers.
• xWAR blend (~28%)— same signals packed into a WAR-style proxy so mid-stars don't slam the ceiling.
• Elite-tier — when prim82 ≥ 90, TOI ≥ 21, and xWAR ≥ 6.5, skip the soft-cap trim, soften age-30 × long-term haircut, raise the forward fair ceiling (~$16.5M tune), floor offense on a prim-led index so mid adjxG cannot veto a franchise package, and skip the comps second pass (peer medians are mid-star heavy). Flagged in notes as elite-tier.
• Context — aging curve, RFA discount, long-term risk, soft market anchor on live mid-star deals, optional comps pull.

Defense

• Role base from TOI — ladder from depth (~$1M) through top-four (~$4.2M) / high top-pair (~$5.5M) to true 1D minutes (~$6.4M) before offense/impact add-ons.
• Offense — adj/HD xG + primary points + a smaller secondary-assist term.
• Impact — rel xG% and xGF−xGA rate, scaled by even-strength minutes (usage-weighted).
• Two-way / ST — blocks get extra weight; PP/PK dollars slightly higher than forwards.
• Same xWAR blend, aging, RFA/term, comps path as forwards (D peer pool).

Goalies

• GSAX— MoneyPuck expected goals against minus actual GA, blended up to three seasons so one ugly (or hot) year doesn't own the fair AAV.
• SV% / GAA — still in the mix; fallback when GSAX is missing.
• Workload — games played vs a ~50-start reference; starter floorsladder by GP × AAV so high-GP / high-AAV creases don't collapse to backup money.
• Aging + RFA/term; AAV-tier capstemper hot GSAX on backup and mid deals so a spike can't invent elite fair.

How inputs become fair $ / %cap

1. Load MoneyPuck all / 5on5 / 5on4 / 4on5 (+ playoffs when present) into a season impact bag.
2. Blend up to three seasons at per-game rates(default 50/30/20; young risers overweight the latest year). Thin GP seasons (<10) are skipped.
3. Position heuristic → raw fair dollars (F / D / G paths above), including ~72% component model + ~28% xWAR dollars.
4. Multiply playoff factor (capped), age / RFA / term / clause factors. Soft market anchor toward the signed deal for non-ELC mid-star extensions still live (age ≤30, ≥2 yrs left, AAV ≥ $4M).
5. Cap-index: treat those dollars as a share of the $95.5M tune ceiling, then mint fair AAV at the live ceiling ($95.5M). Scoring residuals stay on the active deal: live AAV ÷ ceiling in the final year, signing %cap when current enough, or legacy-blended for older fossil starts. Future extensions stay in projection context.
6. Optional second pass: market comps (median peer %cap × live ceiling) gently blended into fair AAV — see Comps below.

Data dictionary — outputs

Stored on value_scores (fair/actual/surplus/label/confidence/notes). %cap fields are computed at score time (ValueResult) even when not every column is persisted.

• fairAav— model worth in today's dollars (fair%cap × live ceiling).
• fairCapPct — fair as share of cap (primary unit).
• actualCapPct — scoring actual: active-deal live AAV ÷ ceiling in the final year; otherwise signing %cap when known, or live AAV ÷ ceiling when no sign share is known. Pre-2023 starts with sign% ≫ live% blend toward live (notes: legacy-%cap). Signed-future AAVs are shown as projection context, not current label burden. Stored cap_pct_at_sign remains for display.
• surplus — fairAav − actualAav (+ = team surplus / underpay).
• surplusCapPct — fair%cap − scoring actual%cap (+ = underpay). Labels use this. Display also shows at sign residual (fair − sign%) and vs live residual (fair − AAV/live ceiling) when both exist.
• label — overpay | fair | underpay.
• confidence — low / medium / high from blended GP + seasons used; dampened for estimated AAV, ELC, sparse MP.
• methodVersion — currently v1.38.
• notes — human-readable keys: position path, adjxG82/prim82/TOI/xWAR or GSAX/SV%, blend label (3yr 50/30/20, riser 80/15/5, …), mkt-anchor, elite-tier, near-elite, live-%cap, legacy-%cap, ELC/bridge, comps …, season window, %cap actual→fair.

Label bands, ELC / bridge, confidence

• Default band — |surplusCapPct| ≥ ~2.1 pp (≈ $2M at the $95.5M tune ceiling) → overpay / underpay; inside → fair.
• Low confidence — wider band (~3.1pp ≈ $3M) so thin samples don't scream.
• ELC / bridge — cheap young deals (roughly ≤$1M age≤27, ≤$1.75M age≤25, or cheap RFA). Notes flag ELC/bridge. Positive surplus usually stays fairunless talent clears a high bar vs a ~$4–5M bridge. Leaderboards hide AAV < $2.5M ("market board").
• Market deal anchor— pulls model fair toward the signed %cap on live mid-contract deals so a just-paid market extension doesn't auto-label underpay. Skipped for ELCs and when skipMarketAnchor (ex-ante research).

Market comps blend

After the model scores, peers are matched on position group, scoring pace, age (±5), and deal size in %cap at signing (curated AAV ≥ $2M, 40+ GP, not ELC/bridge). Forwards prefer C↔C or wing↔wing when enough same-role peers exist.

• Comps-implied fair = median peer %cap × today's ceiling (nominal AAV median if %cap missing).
• That comps fair is mixed 85/15 with the model fair inside the comps helper; then blendWithCompspulls the stored result ~20–36% toward comps when n≥4 (0.20 / 0.28 / 0.36 at n≥4 / n≥6 / n≥8). ELC/bridge and AAV < $2M skip the post-pass.
• Shown on player contract pages and /research/comps.

How we get better

Feeling "off" is usually stale AAV data, ELC talent vs cost (correct but loud), or the dollar scale drifting vs the real market. Fix with a calibration loop — not mystery features.

• Market cohort — curated deals AAV ≥ $2M, age ≥ 24, 40+ GP. Primary MAE in %cap percentage points (npm run calibrate).
• Residual review — biggest |%cap residual|: verify AAV / %cap@sign first, then dial.
• True GAR/WARis licensed. Our xWAR proxy is open and transparent — not pretending it's EH.

Team AAV payroll

Team pages list the NHL 2025-26 season roster only. Payroll is the sum of curated/import AAVs vs the $95.5M ceiling / $70.6M floor for context, plus surplus. Missing and estimated contracts are counted separately and excluded from the payroll total.

Next-year outlook

Player pages and Ask can answer "how should X look next year?" from the same DB: current vs prior MoneyPuck rates, TOI usage, the Value Score age curve, contract context, and an illustrative GP/points band if the role sticks. /research/outlook freezes those bands on past seasons. This is not a trained projection model — small samples stay low-confidence on purpose.

What is not in the model

• Licensed Evolving-Hockey / hockey-graphs GAR/WAR (we use an open xWAR proxy instead).
• Full CBA tooling: LTIR, buried contracts, retention, buyouts, bonuses, escrow, offer sheets.
• Scraped Puckpedia / Spotrac / HockeyStats HTML — curated AAV only.
• Line chemistry, coach systems, zone-entry microstats, tracking, scouting grades, injury history beyond GP sample.
• Draft capital, trade return, intangibles, "culture" narratives.
• A trained ML regressor — every dial is a readable TypeScript heuristic in score.ts.

Still incomplete

Richer microstats, continuous contract feeds, and deeper ex-ante holdouts. Run npm run calibrate after every model change.