I_PREDICT_SPORT
History — live, settled, audited
72 settled predictions since 2026-07-18 · 55 open · accuracy 62.5% · Brier 0.2220 vs Kalshi's 0.2774 (model better by 0.0555) · avg CLV -1.6pp
Proposed bets: 5 proposed · 4 settled · 3 won · total P&L +0.120 units on 0.398 staked (1 unit = the bankroll fraction staked on that bet) · avg bet CLV +1.7pp
ROI (settled bets): overall +30.1% · average per bet +29.3% · average per betting day +13.9% (return on capital staked; per-day averages weight each betting day equally)
Every number here comes from predictions timestamped BEFORE the match and settled automatically against official results — a live audit, not a backtest. Samples are still small; treat every split as provisional until n grows. CLV (closing-line value) is how far the market moved toward our side between logging and match start — the standard test of whether an edge is real before results accumulate.
Where the model is strongest
| Slice | n | Accuracy | Our Brier | Kalshi Brier | Avg CLV (pp) |
|---|---|---|---|---|---|
| confidence: medium | 9 | 66.7% | 0.2067 | 0.3175 | +3.4 |
| confidence: low | 63 | 61.9% | 0.2242 | 0.2717 | -2.7 |
| disagreement ≥10pp | 38 | 55.3% | 0.2359 | 0.3403 | -0.3 |
| disagreement <10pp | 34 | 70.6% | 0.2064 | 0.2071 | -3.2 |
All proposed bets
| Date | Model | Match | Pick | Our prob | Market at log | Confidence | Result | Stake | P&L (units) | ROI | CLV (pp) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2026-07-23 | — | Yuta Shimizu vs Johannus Monday | Johannus Monday | 70% | 54% | low | open | 0.021 | — | — | +0.0 |
| 2026-07-21 | — | Arthur Rinderknech vs Sebastian Baez | Arthur Rinderknech | 64% | 47% | medium | lost | 0.098 | -0.101 | -103.7% | +2.0 |
| 2026-07-19 | — | Facundo Diaz Acosta vs Federico Cina | Facundo Diaz Acosta | 67% | 58% | low | won | 0.100 | +0.070 | +69.5% | +2.5 |
| 2026-07-19 | — | Jurij Rodionov vs Joel Schwaerzler | Jurij Rodionov | 77% | 54% | low | won | 0.100 | +0.082 | +82.0% | +0.5 |
| 2026-07-18 | — | Taro Daniel vs Juan Carlos Prado Angelo | Taro Daniel | 74% | 58% | low | won | 0.100 | +0.070 | +69.5% | — |
All predictions — chronological
| Date | Match | Model pick | Our prob | Market | Confidence | Status | Result | Bet | Stake | P&L (units) | ROI | CLV (pp) |
|---|
Lower Brier is better (0 = perfect, 0.25 = coin flip). Machine-readable: track_record.json · predictions.json · for robots.