How 11Stat Works

In short: the 7-step process

• The process is leak-free (point-in-time): only data from before a match is used to model it.
• Not one engine but at least 3 independent engines (Poisson/Dixon-Coles, Elo/form, advanced bivariate) model each match separately.
• Calibration corrects raw probabilities once; a '70%' label really means about 70%.
• The data-quality gate blocks false high signals from missing data; insufficient data = PASS.
• ROI and hit-rate are paper/simulation backtest metrics only, not real-money outcomes.

Overview: From Data to a Model Signal

11Stat runs as an auditable pipeline. Each stage passes to the next only when it clears specific quality thresholds; otherwise the system says PASS (no signal). The goal is to produce a small number of clean, explainable model signals rather than a flood of weak ones.

1. Data ingest: Fixtures, team form history and market odds.
2. Multi-engine modelling: Independent probability engines model the same match separately.
3. Calibration: Raw probabilities are reshaped in logit space.
4. Model agreement & confidence: How strongly the engines concur is measured.
5. Data-quality gate: Missing or thin data stops the signal.
6. Risk level: Variance is labelled low, medium or high.
7. Publish: A model signal if thresholds pass, a PASS if not.

Any ROI or win-rate figures shown on the platform are simulated, paper backtest quality metrics only — they are not real-money results.

1) Data Ingest and Multi-Engine Modelling

For every match the system first pulls three kinds of input: recent team form (goals scored and conceded, recent fixtures), fixture metadata (league, season, kickoff time) and market odds. If no odds are available, the match is dropped outright.

The same match is then modelled by independent engines, so no single method's blind spot drives the whole output:

• Poisson / Dixon-Coles engine: Estimates goal expectations (an xG-style lambda), builds a score matrix with a low-score correction, and derives market probabilities such as 1X2, both-teams-to-score and over/under from it.
• Elo / form engine: Provides a second perspective based on team strength ratings and current form.
• Advanced bivariate engine: A more detailed approach that models the relationship between the two teams' scoring and zero-inflation effects.

Each engine produces its own probabilities for the same match; the later stages combine and audit those outputs.

2) Calibration: Making Probabilities Honest

Raw model probabilities are often overconfident. 11Stat softens them with a single, logit-space calibration step. Each market uses a monotonic slope kept slightly below 1, which reduces over-confidence without changing the ranking — so a probability becomes more realistic rather than just smaller.

On top of that, a league- and market-specific calibration factor learned from historical performance is applied. A key design choice: that factor is applied exactly once; the double-application bug in older versions (which shrank probabilities twice) has been fixed. The result is that a "70% probability" label is tuned to genuinely reflect roughly 70% confidence.

3) Model Agreement, Confidence and Risk Level

Calibrated probabilities alone are not enough; 11Stat also looks at how strongly the engines point the same way. Model agreement is the share of engines that support a given market. High agreement raises confidence; when engines conflict, confidence is lowered.

The confidence score is not a single number but a weighted blend of components: calibrated probability, model agreement, value ratio versus the market, historical market accuracy, sample reliability and variance control. Signals are then labelled LOW / MEDIUM / HIGH / ELITE.

• Confidence floor: A long-odds, low-probability pick cannot earn the top label just because its price is high; upper tiers require a minimum calibrated probability.
• Risk / variance: Every match is flagged low, medium or high variance, and that directly affects confidence.

This is deliberately framed as a measurable risk level rather than language like "sure" or "lock."

4) The Data-Quality Gate and Market Health

11Stat's most important safeguard is the data-quality gate. If a fixture id, team id, kickoff time or odds feed is missing, the match is marked "DATA ERROR / NO PICK" and no signal is produced. This is what stops fake high-value picks manufactured from thin data from ever reaching publication.

The platform also tracks a rolling health history per market (the last few periods' simulated ROI, win rate and calibration error). Markets that keep performing poorly can be disabled automatically, while some volatile markets stay off by default until enough positive history accumulates. The system only speaks in markets it currently finds trustworthy.

5) Publishing: A Model Signal or a PASS

When every stage passes, the result is published as a model signal. The signal carries the market, the calibrated probability, a confidence tier (e.g. HIGH/ELITE), model agreement (e.g. "3/3 engines"), the risk level and a comparison against the market odds. None of this is a promise that a prediction will land; it is a probability-based analytical output.

If the thresholds are not met, the system honestly says PASS. That is precisely the difference between football predictions and data-driven analysis: 11Stat is not a coupon seller but an analysis tool that shows how confident the model is — and when it stays silent. All published content is for informational purposes and is not betting advice.