Methodology
11Stat is a football data-analytics platform that models matches with three independent probability engines, transparent data sources and calibration checks; every output is an analytical model signal, never betting advice or a guaranteed outcome.
The models and engines we use
Instead of a single 'magic formula', 11Stat runs three independent engines that cross-check each other. Whether or not those engines agree on a market is, for us, a signal as important as the probability itself — we call it model agreement.
- SHARP engine (Poisson + Dixon-Coles): Derives expected goals (xG) as a weighted blend of season performance, recent form and home/away-specific form (roughly 40% season, 40% general form, 20% specific form), builds a Poisson score matrix, then applies the Dixon-Coles correction so low scores match their real-world frequency.
- 888 engine (Elo + form): Models team strength with an Elo rating and momentum with a points-per-game (PPG) measure that weights recent matches more heavily.
- Advanced engine (bivariate Poisson): Uses a two-variable Poisson distribution with a shared dependency parameter (rho) plus low-score shaping, so home and away goals are not treated as fully independent.
Each engine's output is converted into markets such as match result (1X2), goal lines (over/under 1.5/2.5/3.5), both teams to score (BTTS), half-time/full-time and total-goals bands. The three engines are compared per market, and how many agree feeds directly into the confidence label of the output.
Data sources and freshness
Model inputs come from a live sports-data provider: fixtures, finished match results and bookmaker odds. We do not use synthetic or hand-invented odds; if no real market price exists for a match, that market is excluded from any value calculation.
- Match history: For each team we use recent matches (typically the last ~12 fixtures).
- Point-in-time, leakage-free design: When modeling a match we only consider games played before that match's kick-off. This prevents look-ahead data leakage.
- Real odds: Value and edge are computed against genuine bookmaker odds wherever possible. For markets whose odds can't always be fetched (e.g. half-time/full-time and total-goals bands), the price used is an estimate and is clearly flagged in the output, so that market's 'value' is treated as indicative only.
When a match lacks enough history (for example fewer than 3 prior games) or has no odds, the system deliberately returns 'no signal'. We do not fill gaps to manufacture a forced prediction.
Calibration: bringing probabilities closer to reality
Raw model probabilities are usually over-confident. Calibration is the step that corrects them, with the goal that 'when we say 70%, over the long run it really happens about 70% of the time.'
- Logit-space shrinkage: For each market, probabilities are gently pulled toward 0.5 in logit space using a Platt-style transform with a slope just under 1. This reduces over-confidence while preserving the ranking.
- Learned market factor (Bayesian shrinkage): The gap between observed hit-rate and predicted probability builds a correction factor over time. With a small sample the factor is pulled toward neutral (1.0); as data accumulates it approaches the observed rate. The factor is bounded by a symmetric band (≈0.85–1.15), so it can correct both over- and under-confidence.
This design replaces an older, flawed implementation that silently shaved ~5% off every probability. The aim is honest probability, not manipulation.
Backtesting and simulated ROI (not real profit)
We measure model quality with a leakage-free backtest over historical matches. The ROI and win-rate here are simulated, paper-only quality metrics — they are not real-money returns and do not guarantee future performance.
- Brier score: Measures probability calibration (0 = perfect, 0.25 = uninformative). A value above 0.25 triggers a calibration warning.
- Sample threshold: A market needs at least 20 selected signals before its ROI/win-rate is trusted; totals under 80 are flagged as 'not yet reliable'.
- Concentration check: If a single market accounts for more than 55% of selections, it is flagged as an overfitting risk.
- Market health: Markets with negative ROI over two consecutive backtests, or high calibration error, are automatically disabled on the next run (for example, the total-goals market is off by default until it proves value).
All of these checks exist to expose the model's weaknesses rather than to make results look better than they are.
Known limitations and uncertainty
No football model can know the future. These are the limitations 11Stat keeps deliberately transparent:
- Football is inherently high-variance; over small samples, luck can matter more than the model.
- Context variables such as injuries, squad rotation, weather, refereeing and motivation can't always be fully captured.
- Because odds for some markets (half-time/full-time, total goals) aren't available for every match, their value calculations are indicative only.
- Bookmaker odds usually price the market efficiently; genuine long-run edge is rare and should not be claimed unless measured against the closing price.
That's why every output ships with a risk level and a confidence label — we surface uncertainty rather than hide it.
Product boundary: data analysis, not betting
11Stat is a football data-analytics and model-evaluation platform; it is not a sportsbook, coupon seller or gambling automation tool. This is exactly where the difference between 'football predictions' and 'data-driven analysis' becomes concrete.
- Our outputs are analytical signals — probability models, expected goals, model agreement and risk level — not instructions to place a wager.
- We never present any outcome as 'guaranteed', 'certain' or a 'lock', because probability by definition carries uncertainty.
- ROI and win-rate are simulation quality metrics, not a promise of real-money winnings.
In short: we don't tell you what to bet. We transparently show you what the data says, how confident the model is and how accurate it has been historically. The decision — and the responsibility — remain entirely with the user.
Frequently Asked Questions
Is 11Stat a betting site?
No. 11Stat is a football data-analytics and model-evaluation platform. It does not accept bets, sell coupons or offer gambling automation; it only produces probability models and analytical signals.
Is the ROI shown real profit?
No. ROI and win-rate are simulated, paper-only quality metrics from backtests over historical matches. They do not represent real-money returns and do not guarantee future performance.
Which models do you use?
Three independent engines: a SHARP engine based on Poisson + Dixon-Coles, an 888 engine based on Elo + form, and an advanced engine based on a bivariate Poisson distribution. Outputs are compared per market.
How do you calibrate probabilities?
Probabilities are gently shrunk in logit space and adjusted by a learned market factor that reflects observed hit-rate and is pulled toward neutral when samples are small. We measure calibration with the Brier score.
How fresh is the data, and is it leakage-free?
Data comes from a live sports-data provider, and each match is modeled using only games played before its kick-off (point-in-time). This prevents look-ahead data leakage.