Track record
Walk-forward backtest
Every prediction the model has ever made, settled. $100 staked per pick at fair odds (1/argmax_prob). This is the optimistic bound — a real book charges vig that drags ~5-8% per leg, so flat-positive here is required just to break even at a sportsbook.
Sample: 2022-01-03 → 2026-10-31 · 30,000 predictions
Walk-forward floor vs live model
Walk-forward is Elo-only — the lowest defensible model. The live model adds Platt calibration, news features, and market-blended priors. This is a true same-match comparison: BOTH rows are computed over the 1,331 matches where each model made a graded prediction — so the Δ is the signal the extra stack actually adds, not an artefact of different match sets.
Shared matches- Elo only
- 1,331
- Live model
- 1,331
- Δ
- —
Argmax hit- Elo only
- 56.5%
- Live model
- 56.3%
- Δ
- -0.2pp
ROI (fair odds)- Elo only
- 0.65%
- Live model
- -2.69%
- Δ
- -3.34pp
Overall (fair odds)
Predictions
30,000
argmax avg 62.6%
Argmax hit rate
61.6%
18,485 of 30,000
Net P/L @ $100/pick
−$40,282
staked $3,000,000
ROI
-1.34%
fair-odds — book vig drags ~5-8% per leg
By sport
Tennis-1.76% roi
- Predictions
- 23,391
- Argmax hit
- 62.9%
- Avg prob
- 64.0%
- Staked
- $2,339,100
- Returned
- $2,297,816
- Net
- −$41,284
Football3.74% roi
- Predictions
- 4,221
- Argmax hit
- 51.7%
- Avg prob
- 49.4%
- Staked
- $422,100
Reading this honestly
- Fair-odds backtest uses 1/argmax_prob as the decimal price. Real books charge vig — a fair-odds ROI of +5% corresponds to roughly -3% at a real sportsbook. So the realised number here is the optimistic bound, not what you'd actually win.
- Walk-forward = Elo only. No Platt blending, no news, no market signal — predictions captured BEFORE the Elo update on each match. This is a strict lower bound on the live model's quality.
- Argmax bias. The model picks its most-likely outcome. In football that's usually the home team at short odds — the market's most efficient band. ROI ≈ 0% there is honest calibration; meaningful upside would require beating the close in those bands — which is measured directly here.
- Why this beats small-sample. The daily-tickets track record needs ~30 graded tickets per tier before signal stabilises. This page covers thousands of predictions across multiple seasons — the variance question is settled here, not on a 5-ticket sample.