Methodology · How TACTIX Predicts
Every football prediction TACTIX publishes comes from one model. This page explains how the TACTIX AI football prediction model turns team data into probabilities, how each market is calibrated and checked against real results, and the limits of what any prediction model can tell you about the next match.
The TACTIX AI football prediction model rates each team's strength from recent form and expected goals, turns the two ratings into a grid of likely scorelines, and reads goal markets such as both teams to score straight from that grid. Since the record began, its mean calibration error has been ±0.1% across 6M+ settled predictions (all markets, player markets included), losing picks included. Each market is calibrated separately before it is published, and the record is graded in public on the model scoreboard.
The output is a probability, and that word is the one most often misunderstood. A model that says "Liverpool 64% to beat Brentford" is not predicting that Liverpool will win. It is predicting that, if Liverpool and Brentford were to play the same match under the same conditions 100 times, Liverpool would win roughly 64 of them. Bookmakers' implied odds work the same way; they are probability estimates dressed in fractional or decimal odds. A well-calibrated model and a well-calibrated bookmaker disagree on individual matches but rarely on the long-run distribution of outcomes.
The TACTIX model is a transparent statistical model with machine-learned layers on top. The base model, version tag rule-based-v2, applies an openly weighted formula to six team statistics rather than learning hidden patterns match by match. The learning happens above it: calibration curves are fitted to settled results for each market family and refitted every night, a stacking model trained on settled results refines two goal markets, and a three-way match-result model, also trained on settled results, sets the home / draw / away probabilities. TACTIX calls that combination AI, and this page shows where each part begins and ends.
What no model can do is read the dressing room. It cannot know that a key player has fallen out with the manager, or that the away team's coach broke down on the M6. A model is always behind reality for events that have not yet been observed, which is why a model output is a starting point, not a final answer.
Every TACTIX football prediction passes through the same seven steps, in the same order, for every fixture. Each step is a separate piece of code that can be tested on its own.
The model scores each team from six inputs. By default, recent form carries the most weight, followed by expected goals (xG), the quality of the chances a team creates and concedes, and then shots on target. Possession, goal difference and injuries contribute smaller shares.
Why so few inputs: form and chance quality are the steadiest signals of current strength, and a short, openly weighted list is easier to test and harder to overfit.
Each side's attack is set against the other side's defence to give an expected-goals figure for each team: the average number of goals it would score if the match were played many times.
A team with too few recorded matches starts from a neutral rating, so the model never invents a strength it has no data for.
A Poisson distribution, the standard way to model how often a rare event such as a goal happens in a fixed time, turns the two expected-goals figures into a probability for every scoreline.
A Dixon–Coles correction then adjusts the low-scoring results, such as goalless and one-all draws, which a plain Poisson model misprices. The size of that correction is fitted league by league where the data supports it.
Goal markets come straight from the matrix: double chance, both teams to score, total goals, correct score, handicaps and half-time markets. Corners, cards, shots and player markets come from companion count models built on the same team and player data.
The cards markets are also adjusted for the appointed referee's card record against the league average.
Each market family is calibrated: its raw probabilities are mapped onto how often those outcomes actually happened in recently settled matches. The curves are refitted every night, so a stated 60% keeps meaning roughly 60% in every market.
For both teams to score and over/under 2.5 goals, a stacking layer, a second model trained on settled results that re-weighs the first model's outputs, supplies the published probability.
It is used only because it beat the base model on matches neither had seen. In every other market the calibrated base model is served unchanged.
The headline home / draw / away probabilities come from a dedicated three-way match-result model: a multinomial model, trained on settled results, that weighs the base model's own outputs, such as its match-result probability and each side's expected goals, together with Elo ratings (a running team rating that rises and falls with every result) and, where the active version uses them, recent form, rest days and rating trend.
It is fail-safe: if no trained version is loaded, the base model's own call is served instead, built from the two strength scores with a home-advantage allowance and a small head-to-head adjustment.
The TACTIX model uses six team statistics for its strength score: recent form, expected goals (xG), shots on target, possession, goal difference and injuries. The headline match result is set by a three-way model that adds Elo ratings and, where its active version uses them, recent form, rest days and rating trend to the base model's outputs. The cards markets also use the appointed referee's card record. The event markets draw on each team's and each player's recent counts of shots, corners, cards, fouls and similar events, pulled towards the league average when the sample is small.
All of it comes from historical fixtures and match statistics that are reconciled against final results. The open data page lists the datasets TACTIX publishes, and the methodology whitepaper describes each source in full.
The published TACTIX model does not use the weather or the content of confirmed line-ups. TACTIX has built adjustments for both, and for fixture congestion, but they stay switched off unless out-of-time testing shows they improve the record; the learned layers note only whether line-ups have been announced, not who is in them. The referee is used narrowly: the referee's card record adjusts the cards markets, and the same card-tendency figure is one input to the learned layers, but it does not move the goal markets directly. Player ratings, motivation and dressing-room news are not inputs, so late team news is not reflected in the published probability.
Calibration is the step that makes a TACTIX probability mean what it says: when the model says 60%, the outcome should happen about 60% of the time. The base model produces a raw probability for each market, and two measurement-driven layers adjust it before it is published. Neither layer changes anything unless it helps on fixtures the model has never seen.
A model can rank matches well and still be mis-calibrated, saying "60%" when the real frequency is 65%. TACTIX fits a separate calibration curve for each market family (match result, both teams to score, totals, corners, cards, player props and the rest), so every market is corrected to its own historical reliability rather than sharing one blunt, pooled adjustment. The curves are refitted every night from recently settled predictions.
The headline 1X2 probabilities (home / draw / away) are set by the three-way match-result model in step 7, a multinomial model trained on settled results. If no trained version is loaded, the base model's own call is served instead.
For some markets a second-stage "stacked" model squeezes more signal out of the same inputs; for others it does not. TACTIX decides this by measurement, not assumption: a market is served from the stacked model only when that model beats the base model on a held-out, out-of-time sample. Where it does not, the base model is served unchanged. Today both teams to score and over/under 2.5 goals are served from the stacked layer on that basis.
TACTIX measures accuracy first as calibration, how closely its stated probabilities match what actually happens, and publishes the result live: a mean calibration error of ±0.1% across 6M+ settled predictions (all markets, player markets included). Hit rates are published separately, by market and by confidence tier, with the number of settled picks beside each, on the model scoreboard and the How accurate is TACTIX? page.
Each published pick is written, with its publication time, to the public, hash-chained pick ledger, so the record cannot be edited afterwards and anyone can check whether a pick was published before kick-off.
A single "accuracy %" is easy to quote and easy to mislead with: a high number can simply reflect a league where favourites usually win. That is why the headline measure is calibration, measured out of time, by market, with sample sizes shown on the track record. The methodology page explains how predictions are tested and corrected.
A TACTIX confidence score is the model's calibrated probability for its selection, so a 60% pick is one the model expects to land about three times in five. Picks are banded into Bronze, Silver and Gold tiers by that probability, and below the Bronze floor the model shows analysis but does not rate a pick. The thresholds are set out in the methodology whitepaper.
Take a fixture where the model gives the home side 1.6 expected goals and the away side 1.1, the illustrative inputs used in the TACTIX goal-matrix code. The Poisson / Dixon–Coles grid makes 1–1 the single most likely score, at about 13%. Read off the same grid, before calibration, both teams to score comes out at about 55% and over 2.5 goals at about 51%. Neither clears the rating floor, so neither would be published as a rated pick: the model is saying these are close calls, which is useful information in itself.
A Gold pick has cleared the highest bar and a Bronze pick the lowest, and each tier's hit rate is published separately with its sample size, because blending them into one number would flatter the model.
One consequence of honest calibration: low-confidence picks are not "worse" picks — they are matches where the model is appropriately uncertain. Reading a low-confidence call as a recommendation to bet the opposite outcome would be a misuse of the data; the right interpretation is "this is a coin-flip and the model knows it."
League fixtures with stable line-ups in well-covered leagues (Premier League, Bundesliga, La Liga, Serie A, Ligue 1, Championship, Eredivisie). Recent form and expected goals, the model's two heaviest inputs, are most informative when a team is picking from a settled squad, and these leagues have enough fixture density to keep the rolling figures fresh. The scoreboard shows, market by market, where the record is strongest.
Cup ties in the first knockout rounds (rotation chaos, motivation gap), early-season fixtures (last season's data is partially obsolete), matches involving newly-promoted or newly-relegated clubs (the model's prior is wrong by definition), and matches where both teams have nothing to play for in the final two weeks of the season (motivation evaporates, results become quasi-random).
Predict individual freak events (red cards in the first ten minutes, missed penalties, refereeing errors that the model cannot weight in advance). Account for team news: the published model does not read confirmed line-ups, so a late absentee is not in the probability. Predict outcomes in matches where one team has shipped a manager in the last 72 hours (the model has insufficient data on the new system). Beat well-priced bookmaker odds systematically over the long run after the bookmaker's margin — that is a separate, much harder problem that requires not just a good model but better information than the bookmaker has.
The TACTIX model improves through measurement rather than constant rebuilding. Finished matches are settled against the published probabilities. Calibration curves are refitted every night from recently settled predictions; the stacking layer is retrained on settled results and kept only where it beats the base model out of time; and challenger models run in shadow, scored on real fixtures without publishing anything, until they prove better on both calibration and the Brier score (the average squared gap between a probability and the result).
Football changes, and calibration drift is monitored for that reason. Because each strength score is built from a team's recent matches, it moves with the team. If the model says "60% home win" across a sample of 200 matches and only 50% of those matches end in home wins, the calibration is off and the next refit corrects it. The aim is not to maximise raw accuracy but to keep the headline probabilities matching the observed long-run frequency. A well-calibrated 60% probability that loses three games in a row in a small sample is doing its job; an overconfident 75% probability that loses one game out of fifty is not.
Until 9 October 2026 this page described TACTIX as a multi-model machine-learning ensemble combined by a meta-learner. That was wrong, and I have corrected it: every published prediction comes from the model and the layers described above. The methodology whitepaper's version history records the change.
A prediction is a starting point, not a verdict. The most useful way to read a TACTIX prediction is to ask three questions: do the model's headline probabilities differ meaningfully from the bookmaker's implied odds; is the confidence tier high enough to take the disagreement seriously; and is there a piece of information you have that the model does not (a late team-news leak, a tactical pattern not yet in the data) that would shift your read further?
If the model and the bookmaker agree closely, the market is efficient and there is no edge to exploit. If they disagree at high confidence, that disagreement is interesting — but it is also exactly where the model is most likely to be missing something the bookmaker has priced in. Treat disagreement as a research prompt, not an action signal.
If you do choose to act on prediction data with money, treat it as entertainment, not as a financial plan. Set a budget you can afford to lose, stake the same percentage on each pick, and stop when you hit your limit. If gambling is causing you harm, free confidential help is available from the National Gambling Helpline on 0808 8020 133 (free, 24/7) or via BeGambleAware.org.
18+ | BeGambleAware.org
An AI football prediction model is software that turns football data, such as results, expected goals, form and other match statistics, into a probability for each possible outcome of a match. The TACTIX model rates both teams, converts the ratings into a grid of likely scorelines and calibrates every market against settled results. It is not a crystal ball: it is a calibrated estimate of how often each outcome would occur if the same match were played many times.
TACTIX publishes its accuracy as calibration first: a mean calibration error of ±0.1% across 6M+ settled predictions (all markets, player markets included), losing picks included. Hit rates are published separately by market and confidence tier, with sample sizes, on the scoreboard. No single accuracy number is honest across every league, market and period, so TACTIX does not quote one.
The TACTIX prediction model scores teams from six statistics: recent form, expected goals (xG), shots on target, possession, goal difference and injuries. The headline match result adds Elo ratings and, where the active three-way model uses them, recent form, rest days and rating trend; the cards markets add the referee's card record. Weather and the content of confirmed line-ups are not inputs to the published model.
No. Every TACTIX prediction starts from one base model. Learned layers sit on top of it: per-market calibration, a stacking model for both teams to score and over/under 2.5 goals, and a three-way model for the headline match result, each trained on settled results. Earlier versions of this page described a multi-model ensemble; that description was wrong and has been corrected.
The TACTIX confidence score is the model's calibrated probability for its selection. Picks are banded into Bronze, Silver and Gold tiers by that probability, and each tier's hit rate is published separately with its sample size, so a Gold record is never blended with a Bronze one.
Bookmaker prices are themselves model outputs, refined by trading desks and adjusted in response to the money placed against them. Beating the bookmaker on average — across a large sample, after the bookmaker's margin — is genuinely hard, and most prediction services that claim to do so are either selecting their results retrospectively or testing on data the model has already seen. TACTIX's prediction page is designed to help readers reason about matches, not to promise returns against bookmakers.
Football is a high-variance sport. A single goal can swing the outcome of a 90-minute match, and the underlying expected-goal probabilities for most shots are low. Even a perfectly-calibrated model will be wrong on individual matches frequently — its claim is only that its probabilities are right on average across a large sample. Anyone using predictions for betting or fantasy decisions should treat them as one input among several, not as a definitive forecast.
Every prediction comes from the TACTIX model described above, with its calibrated probability, confidence tier and reasoning shown for each match, and the record graded in public on the scoreboard.