Season tracker
Locked picks only, against the closing favourite on the bouts where closing odds were captured. Odds are a yardstick here and nowhere else — never a model input.
50% on 14 locked picks across 1 event(s), log loss 0.7747; the closing favourite went 86% on the 14 of those with odds
Cumulative
One graded card so far. The cumulative lines start at the second — a single point on an empty grid is not a trend.
Calibration
Does a 70% pick win about 70% of the time? Each point is a band of
locked picks: where it sat on the dashed line, the model meant what it said. Above the line
it was too cautious at that price, below it too confident. The blue bar is a 95% interval.
This rests on 14 graded locked picks. Every bucket here is below the 10 picks it would take for its position to mean anything on its own — read the intervals, not the dots. The chart is here so the record builds in the open, not because it is yet evidence of anything.
| Band | Picks | We said | Happened | 95% interval | |
|---|---|---|---|---|---|
| 50–60% | 8 | 56% | 62% (5 of 8) | 31–86% | too few to read |
| 60–70% | 4 | 64% | 25% (1 of 4) | 5–70% | too few to read |
| 70–80% | 2 | 71% | 50% (1 of 2) | 9–91% | too few to read |
Event by event
| Date | Event | Us | Closing line | ||||||
|---|---|---|---|---|---|---|---|---|---|
| n | Hit | Log loss | Brier | n | Hit | Log loss | Brier | ||
| 2026-10-24 | UFC 333: Volkanovski vs. Evloev | 14 | 50.0% | 0.7747 | 0.2892 | 14 | 85.7% | 0.6741 | 0.2371 |
| All locked picks | 14 | 50.0% | 0.7747 | 0.2892 | 14 | 85.7% | 0.6741 | 0.2371 | |
Footnotes
- UFC 333: Volkanovski vs. Evloev (2026-10-24): picks were formally locked 2026-10-25, after the event. The checksum verifies and the picks are unchanged, but the lock itself was applied late, so this row is weaker evidence of a forecast than the others.
- UFC Fight Night: Cortes-Acosta vs. Walker (2026-11-07): the picks were locked 80 minutes after their deadline (deadline 2026-11-06 15:00 UTC, locked 2026-11-06 16:20 UTC). The picks are identical to the draft saved the day before, so the numbers did not move; only the timestamp is late. The 09:00 CT run started at 09:01 but the Mac went back to sleep (battery, lid closed, Maintenance Sleep) 17 seconds in, mid-collect, and the run stayed frozen until the lid opened at 10:18. It resumed and locked at 10:20. Wake-before-run is now scheduled and the lock runs first.
Retroactive walk-forward picks are not counted here. They are out of sample but were produced after the card, so counting them would turn a backtest into a forecasting record; they appear on their own event’s review page, labelled.