Track Record
Transparency
Model track record
Every pick is locked before kickoff and scored against the real result — no cherry-picking. Below is how well the model's probabilities actually held up over the last 180 days, shown with sample sizes so you can judge what's signal and what's noise.
Measured on the current production models only — retired and experimental versions are excluded, so the sample is intentionally small while the live models build a record. A sport earns a “Calibrated” badge only once it has enough settled games and its probabilities actually match outcomes.
How we treat “value” edges
We only present model-vs-market edge as a validated value signal once a sport's model is independently calibration-checked. Until then, edge is shown as informational context, not a bet recommendation. No sport is validated for edge yet — so no edge is sold as alpha.
Graded picks
Win rate
Brier score
ROI (flat 1u)
Calibration
Each dot is a probability band; dot size is its sample count. The dashed line is perfect calibration — points above it mean the model was underconfident, below means overconfident.
Closing line value (CLV)
CLV measures whether the market moved toward our pick after we locked it — the sharpest single proof of edge. It needs a captured closing line, which we only have for a subset of picks so far, so we show it with its sample size and flag when it's still preliminary.
No closing-line data captured for this view yet. CLV appears as the closing-line snapshots accumulate.
How to read this
- Locked before kickoff. Picks are snapshotted at lock time and graded against the final result — the page can't be retroactively flattered.
- Calibration over bravado. A model that says 60% and wins 60% is trustworthy even if it's rarely loud. We'd rather show honest calibration than a hot streak.
- 1-unit hypothetical tracking. ROI uses flat 1-unit notionals at the best listed price — model evaluation, not betting advice. Variance is real. Not gambling guidance.