How our models work.
A clear boundary between what the data says, what a model assumes, and what an explanation can reasonably claim.
Data
Production will begin with licensed football results and events. In v0.1, all public match and team data are isolated, versioned demo fixtures. Team initials are neutral identifiers, not official club crests. No live vendor or club affiliation is implied.
Feature engineering
Planned inputs include team strength, attack and defence, home venue, recent form, opponent strength and schedule congestion. Missing squad or injury information remains explicitly unavailable. No imputation is silently presented as an observed fact.
Models
Elo, Poisson, xG and form are represented with illustrative comparison rows. Only the displayed Poisson score matrix is calculated at runtime, using two synthetic scoring rates and independent goal counts. Bayesian and machine-learning models are planned.
Ensemble
The displayed NQ Ensemble distribution is a versioned demo fixture, not a fitted or calibrated production ensemble. The current comparison rows are constructed examples, not empirical model disagreement. Production weights require forward validation.
Probability
Home win, draw and away win probabilities sum to 100%. Database snapshots store probabilities between 0 and 1; the interface formats them as percentages. Results refer to 90 minutes plus stoppage time, excluding extra time and penalties.
AI explanation
Explanations must describe structured quantitative outputs. The current site uses deterministic templates. A future language model may explain verified inputs, but it must not invent numerical forecasts, squad news or tactical observations.
Confidence and uncertainty
The demo confidence score is a synthetic index from 0 to 100. It has no empirical calibration and must not be interpreted as the probability of a correct prediction. Model agreement is not proof of reliability.
Forecast provenance
Every prediction record carries a match, model, model version, generation timestamp and input-data version. New forecasts create new records. Database constraints prevent editing or deleting predictions and their model-version definitions, preserving the basis of future audits.
Update frequency
There is no automated refresh in v0.1. The demo slate is fixed at 20 September 2026. Future update cadence depends on provider licensing, ingestion checks and operational verification.
Evaluation protocol
Forward evaluation will compare pre-kickoff snapshots against verified full-time results on identical cohorts. It will publish sample counts, the forecast horizon, multiclass Brier score (unscaled sum, range 0–2), log loss and class-level calibration. There are no historical performance claims in v0.1.
Known limitations
The independent Poisson illustration ignores goal-count dependence, game state and lineup changes. Its score matrix covers zero to five goals per team and reports the omitted probability mass. It will not reproduce the separate illustrative ensemble forecast.