Model Card

Key facts. LiveGoalz is a football prediction, live-score and odds-comparison platform that publishes deterministic pre-match 1X2 probabilities and explainable forecast pages. This card covers the model behind those forecasts: LiveGoalz Football 1X2 Forecaster, version det-2026-06-17 (deterministic). Intended use: informational pre-match 1X2 probabilities. Last updated 2026-06-17. No performance claim is made here — the production engine has no published historical backtest yet, and measured metrics are published on the performance page once a settled sample exists.

Model name and version

LiveGoalz Football 1X2 Forecaster, version det-2026-06-17. A deterministic model: the same inputs always yield the same probabilities, and each version is pinned to a set of golden regression fixtures.

How it works (summary)

The model takes the de-vigged bookmaker 1X2 odds as its prior and applies small adjustments from the available signals — recent form, head-to-head, non-neutral home advantage, season goals per game, season shot volume and possession, and provider-supplied absentee lists — then renormalises. Over/under odds are not part of this 1X2 blend; they feed the separate expected-goals, scoreline, both-teams-to-score and total-goals outputs only. With only odds available the model still returns a full pick at reduced confidence. Full detail is on the methodology page.

Intended use

• provide informational probability estimates for football match outcomes;
• let readers compare model probabilities with bookmaker-implied probabilities at the same horizon;
• support research and analysis of match forecasting.

Non-intended use

• not financial, investment or betting advice;
• not a guarantee of any result;
• not for anyone at risk of gambling-related harm;
• not a replacement for a reader’s own judgement.

Backtest and forward-test status

The production engine (det-2026-06-17) does not yet have a published historical backtest. It is protected by four golden regression fixtures (wc-neutral-fav, league-home, odds-only, season-pergame) that catch accidental code changes but do not prove forecast quality.

A separate Dixon-Coles statistical baseline has been backtested over a large set of historical matches, but that baseline is a comparison reference — it is not the production model and its results do not describe this engine.

The system supports a closed forward-test ledger. Public forward-test results will be reported on the performance page after capture and settlement have produced a verified sample. No result is reported here until then.

Data-quality tiers

Each forecast carries a data-depth indication — odds-only, moderate, or rich — based on how many inputs were available. Confidence reflects this directly: an odds-only forecast is labelled and carries lower confidence than one backed by form, head-to-head, provider-supplied absentee lists and season statistics, so a low-information estimate is never presented as a high-confidence one. Confidence is a separate measure and is not the probability of the pick.

Evaluation metrics

These are the metrics that will be reported on the performance page once a settled sample exists. Definitions:

• Brier score — mean squared error between the predicted probabilities and the actual outcome; lower is better.
• Log loss — penalises confident wrong predictions more heavily; lower is better.
• Calibration — agreement between a stated probability and the observed frequency of that outcome.
• Market benchmark — the same metrics computed on bookmaker-implied probabilities at the same horizon, for comparison. No superiority over the market is claimed unless it is statistically supported.

Limitations and failure modes

• no published historical backtest of the production engine yet — quality is not demonstrated;
• missing-player handling is a blunt count-based nudge (two or more absentees), reliant on the provider’s list, with no per-player weighting and no independent team-news checking;
• new teams or sparse data → odds-only depth, capped/labelled confidence;
• unusual fixtures (neutral venues, fixture congestion, heavy rotation) may not be fully captured;
• high variance: even well-calibrated probabilities yield many wrong single-match calls.

Informational only. 18+. Please gamble responsibly.