How football predictions work

When market picks become final

Market 1X2 and double-chance picks freeze one hour before kickoff. We use the most recent quote captured at or before that cutoff, at most 20 minutes old. Later price changes never alter the saved picks. Results show green Correct or red Wrong against that record. Missing or stale quotes are disclosed and excluded, not filled in later. Changed schedules are flagged and excluded. Market-only matches may use a fully resolved 90-minute market outcome; that is labelled separately from a verified football score.

Read the match in two ways.

Kickstatix publishes probabilities for a home win, a draw and an away win. The compact match board shows a 1X2 pick and a double-chance pick. Click either to see the full probability breakdown, recent form, head-to-head results and the evidence behind the prediction.

These markets refer to full time, including stoppage time, and exclude extra time and penalties. Double-chance probabilities are sums of the original 1X2 probabilities. For example, a 50% home chance and a 28% draw chance give 1X a 78% probability. That covers two outcomes, so its hit rate must never be compared directly with a single-outcome 1X2 hit rate.

Confidence is about the evidence.

A 1X2 pick starts at High for a probability of at least 70%, Medium for at least 55%, and Low below that. Disagreement between methods, a gap of at least 20 percentage points, sparse team history or a team-news concern lowers it to Low. Missing reviewed news caps it at Medium. The final tier can therefore be lower than its headline probability suggests.

Double chance inherits at most the 1X2 confidence tier and is lowered to Low if its methods disagree. Its larger probability does not earn it a higher tier. Confidence tiers are evidence rules, not separately calibrated win probabilities.

Independent methods, visible disagreements.

The statistical model estimates scoring strength using historical goals. Elo tracks relative strength through sequential match results. Where available, bookmaker probabilities have their margin removed and supply the existing model's market-aware baseline. Each method's numbers remain visible. A strong disagreement lowers confidence.

Qualitative team news must have a source and a pre-kickoff observation time. If nobody has supplied a verified review, the page says so. Current score feeds, league tables and club artwork do not modify the frozen forecast.

The record is the product.

The first forecast is saved before kickoff and retained after the result. Incorrect predictions, low-confidence calls, missing forecasts and unresolved fixtures stay visible. Final-score corrections are appended with their prior versions retained. Older forecasts with newly derived double-chance views are excluded from the double-chance performance record.

We report log-loss, Brier score, calibration and sample sizes alongside accuracy. Model changes must clear chronological validation and a reserved holdout test. Read the model lab for the completed tests and rejected experiments.