Fair AAV from open data

Thesis

Open production (MoneyPuck) plus curated contract AAVs can produce a transparent fair-AAV estimate whose market-cohort residuals are measurable and improvable — useful for judging mid-contract NHL deals without licensed WAR or full CBA tooling.

This site is the lab demo. The career asset is the method, the calibration numbers, and the public misses — not a cap encyclopedia and not a takes factory.

Problem

Fans and media argue overpays from vibes. Teams price contracts with private models. Between those poles there is little public, auditable fair-value work that separates (a) bad salary data, (b) ELC structure, and (c) real production–price gaps.

We answer one question per player: given same-season impact and the paycheck, is the deal rich, fair, or a bargain — and how wrong is the model on a held-out-style market cohort?

Data & window

• Production: MoneyPuck official CSVs (2025-26), credited.
• Rosters: NHL season roster API for 2025-26 — not /current FA mix.
• Contracts: curated / import AAV only. No Spotrac scrape. No LTIR / burial / retention math.

Stars Lab 2026-27 desk on gold AAV; Worth production stays on 2025-26.

Method (summary)

Transparent TypeScript heuristics (v1.38) — not a trained ML model. Skater fair AAV blends impact-style rates (adj/HD xG, primary points, 5v5 rel xG%, special teams, two-way proxies) with an open xWAR proxy (~⅓ weight) and a light market-comps blend. Goalies prefer MoneyPuck GSAX, tempered. Aging curves differ by F / D / G. Labels: overpay if surplus ≤ −$2M, underpay if ≥ +$2M.

Full dial list: methodology.

Calibration

We do not judge the model on ELC bargains. Market cohort = curated AAV ≥ $2M, age ≥ 24, 40+ GP, not ELC/bridge.

Live residuals: /research/misses. CLI twin: npm run calibrate.

Known failures

• Stale AAV still drives many loud residuals — fix data before dialing.
• xWAR is a proxy, not Evolving-Hockey GAR/WAR. Do not cite it as licensed WAR.
• Same-season only — this is not yet an ex-ante “will this deal look bad in year 3?” model. That is the 90-day research stretch goal.
• Goalie variance and small samples stay loud; confidence flags exist for a reason.

90-day focus

1. Publish and watch market MAE / bias in pp %cap every model change.
2. Clear AAV errors in the top misses list.
3. One forward test: do large |surplus| names mean-revert in production or next contract? Write it up.

Lab charter (repo): RESEARCH.md