This page explains how to read AlphaBench’s evaluation information and what is—and is not—established by the figures published here. AlphaBench is an educational project, not betting advice. Its pilot covers 5 leagues, 20 teams, and 100 fixtures: the Premier League, LaLiga, Bundesliga, Serie A, and Champions League. For background, see Methodology and Data Sources.
Pre-kickoff integrity
Pre-kickoff prediction snapshots are immutable; post-kickoff edits are rejected. This preserves the original snapshot rather than allowing it to be changed after a fixture starts.
Calibration first
Calibration concerns whether predictions correspond to observed outcomes at the rates they imply. It is an important evaluation concept, but this page does not provide calibration scores or results. A model’s name or prediction alone should not be read as evidence of calibration.
Closing-line value
Closing-line value (CLV) compares an earlier odds snapshot with the odds available closer to kickoff. This page provides no CLV figures, so it does not establish whether predictions beat closing prices. AlphaBench’s listed data sources include odds snapshots; see Data Sources.
ROI and drawdown
ROI and drawdown describe different aspects of historical performance: ROI relates returns to the amount risked, while drawdown describes declines from a prior peak. This page provides no ROI or drawdown figures. These measures should not be inferred from the settled-bet count.
Versioning
AlphaBench lists 8 active models: deepseek-v4.1-flash, claude-haiku-5.5, gpt-6-luna-pro, apodex-1.1-mini, nemotron-3.5-lightning, ling-3.1-flash, solar-mini4, and qwen3.8-omni-flash. The verified record also lists 2,660 settled bets and 1 public correction. These counts do not, by themselves, describe model performance or explain the correction; no further version history is provided here.