Prediction Methodology

Key facts. LiveGoalz is a football prediction, live-score and odds-comparison platform that publishes deterministic pre-match 1X2 probabilities and explainable forecast pages. For each covered fixture the model publishes probabilities for the three full-time outcomes — home win (1), draw (X), away win (2) — that sum to 100%, plus a single pick and a confidence level. Model version det-2026-06-17. Forecast horizon: pre-match. Last updated 2026-06-17. The model is newly deployed and has no published historical track record yet; measured results will appear on the performance page.

How a forecast is built

The model produces probabilities in three steps:

1. Start from the market. When a full 1X2 line is available, the bookmaker odds are converted to probabilities and the bookmaker margin is removed (de-vigging). These de-vigged market probabilities are the model’s starting point — its prior.
2. Adjust with signals. The model then applies small adjustments based on the data available for the fixture: recent form, head-to-head record, non-neutral home advantage, venue season scoring (goals scored per game), season shot volume, season possession, and provider-supplied absentee lists. Each signal nudges the probabilities toward one outcome, and the result is renormalised to sum to 100%.
3. Fall back gracefully. If only odds are available — no form, head-to-head or season data — the model still produces a full pick and probabilities from the de-vigged odds. Less data does not block a forecast; it lowers its confidence (see below).

Because the prior already is the de-vigged market, the model does not add a separate “the market favours X” signal — that would double-count the odds.

Over/under (goal-total) odds are not an adjustment to the 1X2 probabilities. They are used only to shape the separate expected-goals, scoreline, both-teams-to-score and total-goals outputs — they never move the home/draw/away probabilities.

Missing players

If the data provider supplies a list of absentees, the model applies a small adjustment in favour of the opponent when a side has two or more listed missing players. It does not gather medical or team news itself, and it does not judge how important any individual player is — the adjustment is a blunt, count-based nudge that depends entirely on the provider’s list.

Pick and confidence

The pick is simply the most likely outcome. Confidence is not the probability of the pick. It is a separate measure that reflects how far ahead the top outcome is over the next one, how much the signals agree with the pick, and how much data was available for the fixture. A pick based on odds alone carries lower confidence than one also backed by form, head-to-head, provider-supplied absentee lists and season statistics.

Data inputs

Inputs that adjust the 1X2 (home / draw / away) probabilities:

• de-vigged bookmaker 1X2 odds — the prior;
• recent team form;
• head-to-head history between the two teams;
• non-neutral home advantage (skipped at neutral-venue tournaments);
• season goals scored per game;
• season shot volume;
• season possession;
• matches played — the minimum-sample gate for the season inputs;
• provider-supplied absentee lists.

Separate inputs used only for the goals / scoreline outputs (expected goals, scoreline matrix, both-teams-to-score and total goals). These do not change the 1X2 probabilities:

• over / under (goal-total) odds.

Season conceding rates are not a current input: the stats feed reports shots conceded, not goals conceded, so goals-against is intentionally omitted until a real goals-conceded feed exists.

Forecast horizon

Forecasts are pre-match: each published probability set is the model’s view captured before kick-off. Live in-play probabilities are not part of this model.

Model version and reproducibility

The current model is det-2026-06-17. The leading det marks it as deterministic: identical inputs always produce identical outputs. Each release is pinned to a small set of golden regression fixtures so the code cannot change unintentionally between versions. These fixtures protect reproducibility — they do not, on their own, prove forecast quality.

Probability calibration

Outputs are probabilities that sum to one. Calibration — whether a stated probability matches the observed frequency of that outcome over many forecasts — is measured against settled results and reported on the performance page once the settled sample is large enough to be meaningful. Until then we make no calibration claim.

Settlement rules

A forecast is settled by the official full-time result (90 minutes plus stoppage time; extra time and penalties are excluded unless a specific market states otherwise). Postponed or abandoned fixtures are voided rather than scored.

Known limitations

• the production model has no published historical backtest yet, so its forecast quality is not yet demonstrated;
• confidence is not a probability, and a low-data pick has its confidence reduced and labelled;
• missing-player handling is a blunt, count-based adjustment (two or more absentees), not a per-player importance model, and depends on the provider’s list;
• the model reads no team news beyond that provided absentee list;
• football has high variance: even correct probabilities produce many “wrong” single-match calls.

Forecasts are informational only and are not betting advice. 18+. If gambling is affecting you or someone you know, please seek support.

Changelog

• 2026-06-17 — methodology aligned to the production engine: de-vigged market prior; 1X2 signal adjustments (form, head-to-head, non-neutral home advantage, season goals per game, shot volume, possession, provider absentee lists) separated from the goals / scoreline outputs; over/under odds disclosed as a goals-output input only; conceding-rate input removed (goals-against not in the feed); model det-2026-06-17 live.