TACTIX Model Methodology (v2)

1. What is the TACTIX model methodology?

The TACTIX model methodology is the versioned, plain-English account of how TACTIX turns football data into calibrated match probabilities, tests them and grades them in public. Since the record began, the model's mean calibration error has been ±0.6% across 8M+ settled predictions, losing picks included. This document explains where the data comes from, how the model works and is calibrated, how picks are tiered, settled and pre-committed before kickoff, and where the approach has real limits, so that a reader, a journalist or a regulator can check the process rather than take a headline number on trust.

This is Methodology v2, published 9 October 2026. It replaces v1 (11 July 2026), whose description of the served model was wrong: v1 described an ensemble of machine-learning models combined by a meta-learner, and the code serves one base model with learned layers on top: calibration, a stacking layer and a three-way match-result model. Section 3 now describes what actually runs, and the changelog keeps v1 on the record, marked as superseded. Models are living things — features change, calibration is refit, new markets are validated and old assumptions are retired. When any of that changes materially, this document is updated and the version is incremented. Nothing here is a forecast of your results; it is an honest description of a process.

A note on language throughout: TACTIX publishes calibrated probabilities and statistics. That is model-informed analysis, not betting advice, not tips sold as certainties, and never a promise of profit. A probability is a considered estimate with error attached — not a guarantee.

2. What data does the TACTIX model use?

The TACTIX model scores each team from six statistics: recent form, expected goals (xG), shots on target, possession, goal difference and injuries. The headline match result comes from a three-way model that adds Elo ratings and, where its active version uses them, recent form, rest days and rating trend; the cards markets add the appointed referee's card record. Those inputs, and the event counts behind the corners, cards and player markets, are drawn from multi-season historical data and refreshed for live fixtures. The sources are:

• Historical fixtures, results and match events — goals, cards, substitutions and final scores across many seasons and competitions. This is the training base and the ground truth that every prediction is ultimately graded against.
• Team- and player-level match statistics — shots, shots on target, possession, corners, cards, fouls, tackles and more. Shots on target and possession feed the strength score; the rest feed the count models behind the corners, cards, shots and player markets.
• Expected goals (xG) — the quality of the chances a team creates and concedes, and the model's core signal for underlying attacking and defensive strength.
• Bookmaker odds — converted to implied probabilities and used two ways: as a market-sanity gate that suppresses any pick whose model probability wildly disagrees with the market, and as an internal quality signal by comparing our published price with the market's closing price (closing-line value). Odds are a benchmark and a guardrail — they are never treated as ground truth. The settled result is.

Not used by the published model: weather, the content of confirmed line-ups and fixture congestion. Adjustments for each are built, and 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 for the cards markets only, plus one card-tendency input to the learned layers. Player ratings and motivation are not inputs.

Feeds are refreshed continuously and reconciled against final results, and the open data page lists the datasets TACTIX publishes. We are also honest about what we do not have: our current feeds provide aggregate match statistics, not granular positional tracking, so we do not publish shot-location maps or heatmaps we cannot source accurately. We would rather show a real number than a decorative one.

3. How does the TACTIX model work?

The TACTIX model is one published base model, version tag rule-based-v2, that turns a weighted team-strength score into a Poisson / Dixon–Coles goal matrix and derives each market from it; every market is then calibrated, a stacking layer refines two goal markets, and a three-way model sets the headline match result. A single model with open weights is easier to test, explain and correct than a stack of opaque ones, and every layer on top of it has to earn its place out of time. The plain-English walkthrough, with a worked example, is on the AI football prediction models page.

What serves your predictions today

• Served  A team-strength score. Each team is scored from recent form, expected goals, shots on target, possession, goal difference and injuries. By default form carries the most weight, then xG, then shots on target; the other three contribute smaller shares. A team with too few recorded matches starts from a neutral rating.
• Served  A goal matrix. Each side's attack is set against the other's defence to give an expected-goals figure per team. A Poisson distribution — the standard model for how often a rare event such as a goal occurs in a fixed time — turns the two figures into a probability for every scoreline, and a Dixon–Coles correction adjusts the low-scoring results that plain Poisson misprices, using a value fitted per league where the data supports one.
• Served  Market derivation. Goal markets — double chance, both-teams-to-score, totals, correct score, handicaps and half-time markets — are read from the matrix. Corners, cards, shots and player markets come from companion count models built on the same team and player data, and the cards markets are adjusted for the appointed referee's card record against the league average.
• Served  Market-by-market calibration. Each market family has its own calibration curve, refitted nightly from settled results, so a probability means the same thing whether it is a 1X2 price or a corners line (section 4).
• Served  A validated stacking layer — a second-stage model trained on settled results — applied only to the markets where it measurably improves out-of-time results. Today that is both-teams-to-score (BTTS) and over/under 2.5 goals. Where stacking does not help, we do not use it.
• Served  A three-way match-result model for the headline home / draw / away probabilities: a multinomial model trained on settled results, whose inputs are the base model's own outputs (its match-result probability and each side's expected goals), 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 call — the two strength scores plus a home-advantage allowance and a small head-to-head adjustment — is served instead.

What is built but switched off

• Off  Weather, confirmed line-up and fixture-congestion adjustments. Each exists in the code behind its own switch, and each stays off unless out-of-time testing shows it improves the record. None of them shapes a published probability today.
• Off  A legacy multi-model system. An earlier machine-learning model with several sub-models remains in the codebase for reference. It does not run and produces no published prediction.

What we are still testing

• In shadow  Challenger models run in shadow: alongside the live model on real fixtures, without serving any pick. A challenger is promoted only if it beats the served model on both the Brier score and calibration, on an adequate out-of-time sample. We test before we ship — a model in shadow is a valid, ongoing outcome, not a launch.

Two rules govern every change. First, nothing ships unless it is validated out-of-time — tested on matches the model never saw during training, not on the data it learned from. Second, the market-sanity gate above suppresses any pick whose model probability wildly disagrees with the bookmakers' implied price, so a thin read on a data-poor competition cannot surface as a confident, contradicted tip. The live, measured performance of everything served here is on the Model Scoreboard.

4. How is the TACTIX model calibrated?

The TACTIX model is calibrated market by market: each market family has its own curve, fitted to settled results and refitted every night, that maps the raw probability onto how often such outcomes actually happen. Calibration is the property that makes a probability trustworthy. A model is well-calibrated when its stated probabilities match how often things actually happen over the long run: of all the outcomes it rates at 70%, close to 70% should occur; of all those it rates at 30%, close to 30% should occur. In plain terms — when TACTIX says 70%, it should happen about seven times in ten, including the three it does not. That is the number that actually matters when you are weighing a decision, and it is measured over winners and losers alike.

We quantify it two ways, both in plain sight:

• Mean calibration error (a form of Expected Calibration Error, ECE). We group every settled prediction into probability bands, compare the average predicted probability in each band with the actual hit-rate in that band, and average the gap. A smaller number means the stated probabilities track reality more closely. It is a whole-of-model honesty score — not a win rate, and not something you can inflate by cherry-picking easy markets.
• Brier score. The average squared difference between the predicted probability and the actual result (1 if it happened, 0 if not). It rewards being both confident and correct and punishes confident misses; lower is better. Where mean calibration error asks “are the probabilities honest on average?”, the Brier score also captures how sharp they are.

Both are computed on settled results — losing picks included — and published live on the scoreboard, not asserted in marketing. As of this publication, TACTIX's live figure is ±0.6% mean calibration error across an 8M+-prediction track record. Because it updates from settled fixtures, the number on this page tracks the live record rather than a frozen claim.

5. What do the Gold, Silver and Bronze tiers mean?

A TACTIX confidence tier is a band of the model's calibrated probability for its selection. Not every probability is worth surfacing as a pick: TACTIX bands its rated picks by that probability, and is explicit that below the rating floor it does not tip at all:

This makes two different numbers on the site easy to keep straight, because they answer different questions:

• Accuracy / hit-rate figures anywhere on TACTIX count rated (≥70%) picks only, reported by tier. A Gold hit-rate and a Bronze hit-rate are not the same claim, and we never blend them into one flattering headline.
• Calibration is the separate, full-range reliability metric. It uses every probability we publish across the whole range — not just rated picks, and not just winners — which is exactly why it cannot be gamed by only counting confident calls.

6. How are TACTIX predictions settled?

• Frozen before kickoff. Every pick is committed and locked before the match starts. Once the whistle blows, we do not touch it.
• Graded after full-time. The pick is settled against the official result. The settled probability stays public.
• Misses are shown, not hidden. A losing pick remains on the record and inside the calibration maths. There is no quiet deletion of bad weeks.
• Voids and postponements follow the relevant market's standard rules; a fixture that does not complete is not recorded as a win.

Settlement is deliberately boring and mechanical — that is the point. The record you see on the scoreboard is the sum of these settled results, winners and losers together.

7. How can you prove a pick was made before kickoff?

Freezing a pick before kickoff only means something if you can prove it was frozen. So every committed pick is written to a public, append-only pick ledger at /transparency.

The ledger is hash-chained: each entry is hashed together with the previous entry's hash, so the whole history links into a single chain. Change any past entry — a probability, a selection, a timestamp — and every subsequent hash breaks, making the tampering self-evident. Anyone can recompute the chain and confirm that a given pick existed, unchanged, before the match kicked off. That is the definition of pre-commitment, and it is verifiable rather than asserted.

8. What are the model's known limitations?

An honest methodology names its own weak points. These are ours:

• Small-sample leagues. Less history and thinner data mean wider uncertainty. In these competitions the market-sanity gate and the rating floor bite harder, and some leagues are simply not tipped.
• One base model, a short list of inputs. Every probability starts from the same six-statistic strength score. That keeps the model testable, but anything outside those inputs — the weather, a confirmed team sheet, a manager's tactical change — is invisible to the published probability until it shows up in results.
• Lineup timing. Confirmed starting XIs typically arrive roughly 20–40 minutes before kickoff. A pick made before team news carries more uncertainty than one made after it — a key absentee can move a match materially, and the published model does not read confirmed line-ups at all.
• Model uncertainty. A probability is an estimate with error bars, not a fact. Football is high-variance; upsets are a feature of the sport, not a bug in the model.
• No guarantees. No model reliably beats the bookmakers' margin over time. TACTIX publishes calibrated probabilities and its full record; it does not sell certainty, and it never frames predictions as income.

18+. This document, and everything on TACTIX, is model-informed analysis and statistics — not betting advice, not a tipping service, and not a promise of returns. If gambling is affecting you or someone you know, free confidential help is available via BeGambleAware.org. Please gamble responsibly.

9. Changelog: what changed, and when?

Material model changes are versioned here and linked from the affected match verdicts, so a reader can always see which version of the model produced a given pick.

• v2— 9 October 2026 · Current
  Corrects the description of the served model. TACTIX serves one base model (rule-based-v2): a team-strength score from recent form, expected goals, shots on target, possession, goal difference and injuries, turned into a Poisson / Dixon–Coles goal matrix from which the goal markets are derived, with companion count models for corners, cards, shots and player markets. Every market is calibrated against settled results, a stacking layer serves both-teams-to-score and over/under 2.5 goals, and the headline 1X2 probabilities come from a three-way multinomial model trained on settled results (base-model outputs, Elo ratings and, where used, recent form, rest and rating trend), with the base model's call as the fail-safe. The referee's card record adjusts the cards markets. Weather, line-up and fixture-congestion adjustments are documented as built but switched off. The Dixon–Coles correction, recorded in v1 as under shadow evaluation, is now described as part of the served goal matrix. Tiers, settlement rules and the pick ledger are unchanged.
• v1— 11 July 2026 · Superseded
  As published: Initial publication of the TACTIX model methodology: the five-model ensemble plus meta-learner, the three-way xG/Elo match-result model, the validated stacking layer on both-teams-to-score and over/under 2.5, market-by-market calibration, the Bronze/Silver/Gold rating tiers, pre-kickoff settlement and the hash-chained public pick ledger. Dixon-Coles recorded as under shadow evaluation, not yet served.
  Superseded by v2 (9 October 2026). v1 misdescribed the served model: TACTIX does not run a five-model ensemble or a meta-learner, The three-way xG/Elo match-result model, the stacking layer, calibration, tiers, settlement rules and the pick ledger described in v1 stand.

Frequently asked

Is TACTIX an ensemble of machine-learning models?

No. Every TACTIX prediction starts from one base model, which turns team strength into a Poisson / Dixon–Coles goal matrix. Learned layers sit on top: 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. Version 1 of this whitepaper described a multi-model ensemble; Version 2 corrects that, and the changelog records the change.

What does “calibrated” mean?

A model is calibrated when its stated probabilities match how often things actually happen over the long run: of all the outcomes it rates at 70%, close to 70% should occur; of those it rates at 30%, close to 30% should occur. It is the property that makes a probability trustworthy for a decision, and it is measured over winners and losers alike.

What is ECE, and what is a Brier score?

Expected calibration error (which we report as a mean calibration error) groups every settled prediction into probability bands and averages the gap between the predicted probability and the actual hit-rate in each band — smaller is better. The Brier score is the average squared difference between the predicted probability and the actual result, so it rewards being both confident and correct. Both are published live on the scoreboard.

Are TACTIX picks pre-committed before kickoff?

Yes. Every pick is frozen before kickoff and graded after full-time, and each committed pick is written to a public, append-only, hash-chained ledger at /transparency. Altering any past entry breaks the chain, so anyone can verify a pick existed unchanged before the match. Misses are shown, not hidden — proof of process, never proof of profit.

Does TACTIX guarantee winning bets?

No. A prediction is a probability, not a certainty. Football is high-variance and no model reliably beats the bookmakers' margin. Everything on the site is model-informed analysis and statistics, not betting advice or a promise of returns. 18+.