Methodology

What the model is, how its probabilities are calibrated, and how the whole pipeline was validated before a single number went public.

The model

Each team gets an attack rating and a defence rating, estimated from every match in a rolling six-season window. A match's expected goals for each side come from those two ratings plus a single global home-advantage term. Goals are modelled as Poisson counts, with a small, well-established correction (Dixon & Coles, 1997) for how often low-scoring results (0–0, 1–0, 0–1, 1–1) actually occur. Older matches count for less: each match's weight decays with its age, roughly halving over a season. Ratings are fit by maximising a penalised likelihood (a ridge penalty keeps unstable, sparsely-observed teams close to a league-average prior) rather than a hand-tuned heuristic.

Newly promoted clubs get no special-case treatment beyond their prior: a team with zero matches in the window starts at the historical average rating for previously promoted teams, then moves as its own matches arrive. Everything above — the decay half-life, the ridge strength, the promoted-team prior, the training-window length — was chosen once, on data through 2011/12, and has not been touched since (see Validation protocol below).

Calibration layer

Raw model probabilities are passed through a light online recalibration step before publication: a temperature (how confident the raw probabilities are, overall) plus separate home/draw offsets, refit monthly on every match seen so far. This is a hygiene device, not a performance device — on its own it moves accuracy by a fraction of what the base model contributes. It exists because J1's home-win and draw rates swing more from year to year than most leagues', and because a site whose whole claim is "our calibration curve is public" should keep that curve automatically straight rather than relying on someone remembering to refit it. The calibration layer is required to prove it does no harm before it is allowed to ship (a pre-registered "do-no-harm" gate on a held-out period), never to prove it helps.

Validation protocol

The period split was fixed before any test data was looked at: 2000–2011 to choose the decay rate, ridge strength and promoted-team prior; everything from 2014 onward held out and untouched until a single confirmation run. Pass/fail thresholds for that confirmation run were written down and frozen before it was executed, alongside a git tag over the exact frozen parameters and thresholds used, so nothing was adjusted after seeing the result. The confirmation run itself was done in a fixed order — fifteen other, previously-untouched leagues first, to catch implementation bugs cheaply, and only then the one J1 run this project gets to spend. A params_hash is embedded in every published prediction file so the parameters that produced it can be checked against that pre-registration tag after the fact. Each week's predictions must also clear a publish gate (training-data floor, fixture-count check, hash match) before anything is written — a run that fails the gate produces no page rather than a stale or empty one.

The one confirmation run, in full

This is the single most important evidence on the site: a one-shot, pre-registered walk-forward backtest against 3,882 J1 matches (2014–2025-26), scored match-by-match against Pinnacle closing price (Shin de-vig) and against a "knowing nothing" climatology baseline. The result:

The betting market is still better than this model, by 0.0044 RPS (95% CI [0.0026, 0.0060], paired, block-bootstrapped by matchweek). Put differently, this model closes about 74% of the gap between guessing from long-run rates and the sharpest available price. Decomposing the remaining gap (a Murphy/Brier decomposition): essentially none of it is bad calibration (-0.00001) — almost all of it (0.00290) is the model simply not having information the market has. When we say a team has a 25% chance to win, it wins close to 25% of the time: the mean predicted probability of a draw across the whole set was 0.2501 against an observed draw rate of 0.2499 (a +0.0003 difference).

Verdict: PASS, 12 of 3,894 matches excluded from the head-to-head comparison for lacking a market close (they remain in the model-only figures). Reproduce this exact run yourself with ./bin/footy backtest --league jpn1 --from 2014 --to 2025.