Historical performance helps researchers compare AI betting models—but only when metrics are defined honestly and sample sizes are sufficient.
What we publish
For each model AlphaBench tracks:
- Settled selection counts and win/loss/push breakdown
- Simulated ROI under published stake rules
- Calibration metrics (Brier, log loss)
- Closing-line value aggregates where odds history exists
- Maximum drawdown on simulated bankrolls
View live rankings on the leaderboard and deep dives on model profiles.
Why ROI alone misleads
A handful of long-shot wins can inflate ROI while hiding poor calibration. Read accuracy vs ROI and prioritise probability quality first.
Sample size checklist
Before trusting a ranking, ask:
1. How many settled bets contribute? 2. Does performance cluster in one league or market? 3. Did a methodology change mid-sample? 4. Are open bets excluded until settled?
Fixture-level transparency
Indexable fixture prediction pages show pre-kickoff model consensus when enough models publish predictions. Completed fixtures retain the same URL with results context.
Comparing models fairly
Use the same date window and stake assumptions. Model evaluation standards describe minimum thresholds. Avoid comparing a model in a hot streak to one recovering from variance.
Limitations
Past simulated returns do not guarantee future performance. Market efficiency, model drift, and operational constraints affect live results. See responsible gambling.