Record

Calibration

10 models, live grade or dated backtest.

RecordLoading the record…


	In season	n/a	Loading...	n/a
Every completed 2026 game with a model number. The spread record grades the model's own line, set before kickoff, against the closing spread; winner counts outright wins for the side the model had winning; margin error is the line against the final margin. Small samples early in the season. The NFL model does not beat the market. Source: 2026 NFL game payloadsModel vs the closing line, since 2018
	In season	n/a	Loading...	n/a
Same grading over every completed FBS game with a model number and a closing line. Through Week 2 the predictions came from frozen preseason ratings; from September 14 the ratings also learn from 2026 games, with more weight as the season goes on, and Week 4 is the first real check. Source: CFB in-season rating test, Sep 9 2026Build a scorecard from any set of games
	Season complete	n/a	n/a	n/a
A calibration curve sits on the model and was refit weekly; it is weakest at 10-20%, where there is little training data. The 2026 season ended with the College World Series (Oklahoma won), so these counts are final. The model returns in February. Source: Weekly isotonic refit, 2026 season
	Season complete	n/a	n/a	n/a
2025-26 is complete. The moneyline was overconfident at the extremes: where the model said 90% or more, those bets won about two in three. A smooth curve replaced a broken calibration table mid-season, and a fix that pulls extreme numbers toward the market is built and tested for 2026-27 (late October). Game totals were the one repeatable edge. Source: NBA season audit, 2026
	Paused Sep 14	n/a	n/a	n/a
The moneyline, especially the No side, is the only MLB edge that held up. The totals edge tested backwards (the bigger it looked, the worse it did) and was pulled. Brier covers every resolved bet, totals included. Paused September 14 when the odds feed stopped, so the counts stop there too. Source: MLB exit study, 2026-06-10
	Retired	2,387 settled predictions	26.6% win rate vs 23.1% break-even	Backtest
Retired August 28; directional trading closed July 21 after 11 approaches failed out-of-sample. The paper record showed an edge (26.6% vs 23.1% break-even) but only 11.6% of live orders filled, so the edge sat at prices nobody would trade. The model was overconfident above 30% predicted, which was capped before close. Source: 12,000-game temperature backtest (2026-05-01); live fill audit (2026-08-28)
	Backtest	1,220 held-out games	Average miss 18.2 to 11.2 points	Backtest
The new in-game totals model cut the average miss on held-out games from 18.2 to 11.2 points (39% smaller). The pregame total is blended in to steady early projections. Source: Live totals overhaul backtest, 2026-04-30
	Backtest	166 playoff games (2024-25)	Missed the champion	Backtest
Two Elo ratings plus the state of each series. The Knicks won the 2026 title and the pre-Finals board favored San Antonio. The June audit found favorites overrated and underdogs underrated in the playoffs; the same pull-toward-the-market fix targets both. Source: NBA season audit, 2026The saved bracket run
	Grades in January	32 teams	25 of 32 within 1.5 wins of the market	Backtest
Preseason ratings against DraftKings' 2026 win totals: 25 of 32 teams within 1.5 wins, median 0.8. The full table is frozen and public, and grades in January. Source: Preseason win totals vs DraftKingsThe 32 frozen win totals
	Backtest	1,139 NFL / 3,174 CFB held-out games	80% bands covered 81.1% (NFL and CFB)	Backtest
A range, not a pick: 50/80/90% bands covered 50.1/81.1/89.9% of NFL games and 50.5/81.1/91.4% of college games, college slightly wide at 90%. The market's total is more accurate (average error score 7.29 vs our 7.45 on NFL) and blending ours in adds almost nothing. The 2026 coverage check starts at 250 graded games per sport. Source: Totals range rebuild, held-out games (2026-08-28)The totals range, as published

Total range bands covered 81.1% against an 80% target.

McConnell's Calls

Notes

Model against the spread. On every game, the side the model's own line likes better than the Vegas closing spread, graded against that spread. A coin flip gets about 50%; you need about 52.4% to beat the sportsbook's cut.

Brier. The average squared gap between the predicted probability and what happened. Lower is better; always guessing 50% scores 0.250.

Backtest. A completed test on games already played, with a date and a source. Not a live record.

The spread record counts every game. A separate tally of only the games where the model sat 1.5 to 3 points off the line is a smaller sample and is not shown here.