Methodology

1. Data and model pipeline

Scorestatix is designed around a layered pipeline. Available real match, team, player and odds data are first validated for readiness. Specialist models then run only where the required inputs are present.

• BB 5.0 PRO — bookmaker-intelligence / market-context layer.
• BB 5.1 PRO — statistical goal and probability modelling.
• Monte Carlo PRO — repeated simulation of match outcomes.
• Corner Model PRO — specialist corner markets.
• Cards Model PRO — specialist cards markets.
• Goalscorer Model PRO — player goalscorer markets where player data is available.
• Specials Model PRO — supported special markets.
• Consensus PRO v2 — combines supported real model outputs without imputing missing models.

2. Calibration policy

A calibrator is not activated merely because it improves an aggregate historical metric. Production activation requires sufficiently robust out-of-sample evidence and stability across relevant segments. A candidate may therefore remain in HOLD or SHADOW status even if pooled results improve.

Current production policy may use raw/native probabilities where a proposed calibrator has not demonstrated sufficiently robust improvement.

3. Candidate selection and AI decision layer

Candidate Builder applies deterministic eligibility filters such as real odds availability, model validity, minimum probability, expected value and positive Kelly criteria where applicable. Candidates that do not pass these requirements are not promoted as premium selections.

Decision Calibration uses historical evidence as a risk gate. It may classify a candidate as TRUST, REVIEW or BLOCK. A BLOCK candidate may still be analysed but cannot become an official Best Single.

Official Best Single and Global Best Single outputs are intended to require the AI Decision Agent. The AI layer does not create missing probabilities, odds, expected value or Kelly values; it evaluates candidates already produced by the deterministic pipeline.

4. Known limitations

• Football is inherently uncertain and affected by events that cannot be modelled perfectly.
• Data providers may correct, delay or omit data.
• Odds can change quickly and may differ between operators and jurisdictions.
• Historical performance does not establish future profitability.
• Small samples can produce unstable or misleading apparent edges.
• Model errors, implementation errors and infrastructure failures remain possible.