About the Soccer Model

The published model is market-free: a Dixon-Coles bivariate Poisson core with time decay, priced with club Elo, anchored into one coherent scoreline distribution that every market (1X2, totals, BTTS, Asian handicap, correct score) is priced from — never separate per-market models that could disagree with each other.

The market track is a comparison overlay: de-vigged prices where a real book quote exists (Bovada, FanDuel), plus a market-implied scoreline distribution for markets books don't quote directly (correct score, BTTS, team totals). Divergence — model minus market — is precomputed by the backend and shown as-is; it is never recomputed in the browser from mixed conventions. Divergence is the product here, not a hidden step toward a pick.

Every market row carries a tier, computed by the launch verification run — nobody edits it by hand:

• Calibrated — passed every standing gate, including null-aware calibration, at the stamping run. Only here does “calibrated probability” language apply.
• Receipts — a published probability with a public calibration receipt, but some bar is missed or still pending. Neutral wording; the receipt is linked inline.
• Display only — shown with an explicit “uncalibrated” chip (BTTS today, a known calibration violation).
• Unmeasured — not enough data yet to measure a claim (deep Asian handicap lines). An abstain reason shows instead of a probability emphasis.

Every market row carries ev, stake, and abstain fields — all null at launch except abstain, which always explains why. A dedicated study raced the model against the opening line across every primary market cell and every cell lost — see the backtest page for the numbers, stated plainly rather than buried. EV stays dormant until one of two pre-registered 2026-27 forward tests actually passes on the live CLV scoreboard — never a hand flip, and never quietly turned on without the receipts to back it.

The rendering rule for a pick chip is built and tested now, for whenever that day comes: a chip requires a non-null stake AND a null abstain — no v1 payload satisfies that. Every abstain reason renders as a plain-language caption; a row is never hidden just because EV is off.

The model is team-level. It does not see today's starting XI, a last-minute injury, or a transfer that lands a new signing in the squad. The market prices those things within minutes; we don't. Especially early in a season, part of any divergence you see between our number and the market's is genuine information the market has and the model structurally lacks — not an edge. Read divergence as a starting point for your own handicapping, never as a pick.

Every published claim is walk-forward: a season is predicted only with models trained on seasons before it, and nothing is scored in-sample. The backtest page carries the full gate state, calibration receipts (ECE alone misleads — always read alongside the null-q95 baseline and the ratio between them), and the live closing-line-value scoreboard once it has enough settled picks to mean something.