EDGE CALIBRATION
ADVERSE SELECTION: REALIZED RESULTS BY CLAIMED EDGE
Graded bets grouped by the edge the system claimed at bet time. If real edge exists, realized ROI should hold or rise with claimed edge. If it falls, the biggest apparent edges are model error (the winner's curse).
| Edge bucket | Bets | Hit rate | Model prob | Market prob | ROI | Mean CLV |
|---|---|---|---|---|---|---|
| 0-3% | 2low n | 50.0% | 60.0% | 16.9% | +225.00% | n/a |
| 3-5% | 3602 | 51.3% | 50.4% | 48.7% | +4.59% | +2.80% |
| 5-8% | 1999 | 48.8% | 49.4% | 46.8% | +4.18% | +10.10% |
| 8-12% | 960 | 48.6% | 47.7% | 44.1% | +12.07% | +3.76% |
| 12%+ | 696 | 50.3% | 47.6% | 39.2% | +28.80% | +3.75% |
Mean CLV is the average of each bet's price against the closing price, and a few extreme closes can move it a lot; the CLV Tracker shows the full spread and the median. Buckets under 30 bets are marked low n and should not be read for ROI.
HOW MODEL PROBABILITIES ARE CONTROLLED
Model systems pass every probability through an isotonic calibrator fit on realized outcomes, then skip any bet whose post-calibration edge still exceeds an edge cap. Both exist because early diagnostics showed the largest claimed edges were mostly overconfidence, not value. Calibration is fit on past bets, so treat the chart above as the ongoing test of whether it works.
Computed from graded bets in the public ledger. All results are paper: flagged price at flag time, voids excluded from ROI, no slippage or limits.