Transparent methodology separates research platforms from opaque tipster services. This page summarises how AlphaBench evaluates AI betting models on football markets.
Design goals
1. Pre-kickoff immutability — Predictions and probabilities are frozen before kickoff. 2. Public settlement rules — Wins, losses, and voids follow documented criteria in settlement rules. 3. Multi-model comparison — Independent models compete on the same fixture set. 4. Educational framing — Outputs support learning, not guaranteed profit claims.
Data pipeline
Fixture, odds, and team context feed each model through a controlled prompt or statistical template. Odds snapshots are timestamped; see odds snapshot collection and data sources.
Scoring dimensions
| Dimension | Question answered | |-----------|-------------------| | Calibration | Are probabilities honest? | | CLV | Did we beat the closing market? | | Simulated ROI | What happens under uniform stakes? | | Drawdown | How painful was the worst losing streak? |
Detailed definitions: how AlphaBench scores models.
Sample size and pilot scope
Pilot benchmarks use curated leagues and fixtures. Small samples exaggerate luck. Model pages may remain noindex until minimum prediction thresholds are met—see model evaluation standards.
Governance and corrections
Editorial and AI-generated content passes review described in editorial policy and AI content policy. Factual corrections are logged on corrections.
What methodology does not promise
Methodology cannot remove football randomness, account restrictions, or execution slippage. Models change; version shifts are noted in release notes.