hockey-stats

hockey-stats is an NHL analytics system that runs on one machine: ingestion, canonical facts, expected-goals model families, venue correction, goalie metrics, and a rink shot map in React and d3. The itch was to build the entire pipeline myself against the two public peers, Evolving Hockey and HockeyViz, whose published methods are cataloged in-repo as a D01–D74 evidence file.

The split is the architecture. A TypeScript/Effect application owns Postgres and every canonical write; a Python batch owns the models; the two meet only through designated tables. The batch is a separate OS process the maintainer launches, never the app. One language for everything, or an RPC seam between services, were the alternatives; the choice is recorded in the decision ledger the repo keeps for every durable choice.

Two more decisions define the statistics. The production posterior engine is validated against NumPyro NUTS reference posteriors, while importing jax or numpyro from production code is a testable defect — the reference dependencies live in their own exactly-pinned group. The alternatives were running MCMC in production, or trusting the approximation unchecked. And CI runs fully offline: every test tier serves the NHL HTTP boundary from a committed fixture corpus, so CI never contacts a live NHL host. Live endpoints in CI, or function-level mocks, were the alternatives.

Numbers, by git log and grep (raw wc -l, blanks and docstrings included): 406 commits in five and a half weeks, 2026-06-23 to 2026-08-01; 57 versioned Postgres migrations; 2,637 Python test functions across 238 files, plus Playwright journeys over the real API and web stack.

What is wrong with it, the project documents itself: four open P0 defects, including origin-recovery validation that leaks the held-out target into the predictors, which blocks all-attempt model selection, and calibration endpoints served as uncertainty bands on the goalie card when they are no such thing — a live statistical defect the status file names. The Direct228 offence-map vertical, the subject of most of the final month's commits, never shipped: at HEAD it is blocked on a CHOLMOD ABI mismatch, and zero Direct model rows exist. Skater valuation and the goalie isolates are written but dormant: no campaign, no selection, no cutover. The status doc states that known gaps prohibit any parity claim against the peers, so I make none.

What I trust most in the repo is the register of what is not done. The models are behind the bookkeeping, and the bookkeeping says exactly how far.