Football Prediction Methodology and Editorial Standards

How Does the TACTIX Prediction Model Work?

TACTIX’s football prediction methodology, in brief: every match forecast starts from the TACTIX model, which turns team strength into probabilities for every market and is then calibrated market by market against settled results. Its record is measured in public, with a mean calibration error of ±0.2% across 6M+ predictions settled since the record began (all markets, player markets included). The model rates each team from recent form and expected goals (xG), backed by shots on target, possession, goal difference and injuries, and converts the two ratings into a grid of likely scorelines (a Poisson / Dixon–Coles goal matrix). A dedicated three-way match-result model, trained on settled results, sets the home / draw / away probabilities from the model’s own outputs and Elo ratings, and a validated stacking layer is used only in the markets where it measurably improves results: today, both teams to score and over/under 2.5 goals.

A market-sanity gate suppresses any pick whose model probability wildly disagrees with the bookmakers' implied price, so a thin read on a data-poor competition can't surface as a confident, contradicted tip.

The full technical explainer lives on the AI Prediction Models page, and the live, measured performance is on the Model Scoreboard.

Why Do We Lead With Calibration, Not Accuracy?

Most prediction sites quote an "accuracy %". We do not lead with one, because a single accuracy number mostly reflects how predictable a league or market is, and it is easy to inflate by choosing easy markets.

Our headline is calibration instead: across 6M+ settled predictions (all markets, player markets included), the mean calibration error is ±0.2%. In plain terms, when the model says 60%, the outcome should happen close to 60% of the time, and the figure tells you how close it gets. That is the property that matters when you use a probability to make a decision. It updates live, losers included, on the scoreboard.

The TACTIX Rating — A Transparent Player Score

Most player ratings are black boxes: a number appears and you're asked to trust it. The TACTIX Rating is the opposite. It's a per-match contribution score built from on-pitch actions with published weights, so anyone can reproduce it. Attacking output and creation add value; defensive actions are rewarded; costly actions subtract.

See the live TACTIX Rating leaderboards by league. Because every weight is public, you can check exactly why any player earned the score they did.

Where Our Data Comes From

Match results, fixtures, line-ups, league tables and event data (goals, cards, substitutions) come from established sports-data feeds, refreshed continuously and reconciled against settled results. Team- and player-level match statistics — shots, shots on target, passing, duels, tackles and more — drive both the model's features and the TACTIX Rating.

We're also clear about our limits. Our current feeds provide aggregate match statistics, not granular positional tracking, so we don't publish shot-location maps or player heatmaps that we can't source accurately — we'd rather show a real number than a decorative one. Expanding to a positional-tracking feed is on our roadmap.

How We Present Predictions Honestly

• Probabilities, not certainties. We write about the model's confidence in each outcome, never that a result "will" happen.
• No inflated accuracy. We lead with calibration and show our working, including losing predictions.
• Honest accumulators. Suggested accas are built from favourites with a hard per-leg odds cap and correlation-correct pricing — not longshot lotteries padded to hit a headline price.
• Value is model-vs-market. When we flag "value", it means the model's probability beats the bookmakers' implied price — and we say so.

Editorial Standards & Corrections

Content on TACTIX is built around data and modelling, not original reporting. Where we reference news — transfers, manager changes, injuries — we link to and credit the original source. We don't accept payment to influence editorial coverage or bookmaker reviews, and affiliate relationships are disclosed in context and never affect the model.

We correct mistakes. If you spot a factual error — a wrong stat, a mislabelled fixture, a broken claim — please contact us and we'll fix it and note material corrections. Articles and data pages carry a named author or the TACTIX Model Desk, and a visible last-updated date.

Verify It Yourself

None of this asks for trust. Every claim on this page is checkable — live, with losing predictions included.