Record

Watching

10 patterns under test, with the rule to promote or retire each. Nothing here is a pick.

Updated Sep 15


	NFL	Running	Low	Loading the record…	Wins above market expectation clears +1.96 standard errors on 100+ graded calls	After 100 graded calls, wins above market expectation is below -1.96 standard errors	Weekly, as each NFL week grades; the stopping rule applies at 100 graded calls
The five-slot NFL sheet (one value dog, up to three leans, one key-number call, one unit each) is a public record that measures something: wins above what the prices expected is zero, on average, for a sheet with no edge, whatever mix of favorites and underdogs it runs. If the sheet has an edge, that number drifts positive as the calls add up; if it doesn't, it hovers around zero and the stopping rule stays quiet. Promote if Wins above market expectation clears +1.96 standard errors on 100+ graded calls The average against the closing line is positive with a 95% interval that excludes zero Retire if After 100 graded calls, wins above market expectation is below -1.96 standard errors The average against the closing line is negative with a 95% interval that excludes zero Week 1 was graded under the August rule it was published with; the five-slot rule applies from Week 2. The stopping rule does not apply until 100 graded calls. The full sheet, with wins above market for every week, is on /calls. Next check: Weekly, as each NFL week grades; the stopping rule applies at 100 graded calls
	NFL	Running	Low to medium	sample of 2,127, break-even 52.4%, as of 2026-09-15	In 2026, the model's side in the lean band wins above break-even, with the low end of its interval over 50%, on 100+ games	Outliers outperform leans on 100+ 2026 games, which would mean the band structure is wrong, not the market	Reviewed at the NFL midpoint (after week 9) and again after week 18
Every NFL game is sorted by how far our number sits from the frozen line: agree (under 1.5 points), lean (1.5 to 3), outlier (3 or more). The bands exist because on seasons already played our own error exceeds the market's from about the 3.5-point cut on. The forward question is whether 2026 reproduces that shape: leans roughly break-even, outliers worse than the market. Promote if In 2026, the model's side in the lean band wins above break-even, with the low end of its interval over 50%, on 100+ games Outliers keep losing to the market on margin error, confirming the cut belongs where it is Retire if Outliers outperform leans on 100+ 2026 games, which would mean the band structure is wrong, not the market The lean band's hit rate sits at or below the agree band's on a comparable sample Backward-looking, 2018-2025 seasons tested in order: the size of our disagreement with the line tells you nothing reliable about which side wins (the effect's 95% interval includes zero, over 2,127 games), while our extra error over the market becomes clear and repeatable once we are 3.5 points or more off the line. That is why big disagreements are not called. The 2026 games add up on the game projections board and the scorecard. Next check: Reviewed at the NFL midpoint (after week 9) and again after week 18
	NFL	Running	Low	sample of 160, won 57.5%, 95% interval 49.8% to 64.9%, break-even 52.4%, as of 2026-09-14	The low end of the interval is above 52.4% on 2018-2025 plus 2026 combined (about 600 games, three to four seasons)	2026 point estimate at or below 50% on 30+ games	Weeks 2 through 18, 2026; read-out after Week 18, not before
Over 2018-2025, leans whose model number and market number sat on opposite sides of 3 or 7 won 57.5% of the time, against 51.0% for the other leans. It is one cut found after the data had already been sliced fifteen ways, and the low end of its interval is under break-even, so it proves nothing. It is the only split with an estimate clearly above break-even, which is why it gets a forward test and nothing else does. Written down on September 14, before any Week 2 game. Promote if The low end of the interval is above 52.4% on 2018-2025 plus 2026 combined (about 600 games, three to four seasons) The average against the closing line is positive alongside the hit rate, so it is the model and not the line drifting toward 3 or 7 Retire if 2026 point estimate at or below 50% on 30+ games The average against the closing line is negative with a 95% interval excluding zero A 2018-2025 figure, not a 2026 result. The comparison group is 51.0% (46.1% to 56.0%) on 390 leans that cross neither number. Expect 2 to 3 qualifying leans a week, 35 to 50 by Week 18: the season answers whether the estimate stays above break-even on new games, not whether it is significant. The cutoffs (1.5, 3.0, the numbers 3 and 7) cannot be changed during the test. Next check: Weeks 2 through 18, 2026; read-out after Week 18, not before
	College football	Running	Low to medium	sample of 6,887, as of 2026-09-14	After the September 26 games, the 2026 data favors the blend, with the blended rating beating the preseason rating on margin error	After Week 4 the 2026 data still favors the preseason rating at every weight, which would mean the season disagrees with the backtest	After the September 26 games (Week 4), with the same test run unchanged
Since September 14 the public college ratings blend the preseason rating with a rating fit on every completed FBS game so far. The in-season part starts near zero weight in Week 1 and reaches half around Week 4 or 5. Over 2016-2025 that blend beat the preseason rating alone by 0.82 points of average margin error, all of it from learning during the season. The question is whether 2026 agrees, or keeps favoring the preseason rating. Promote if After the September 26 games, the 2026 data favors the blend, with the blended rating beating the preseason rating on margin error The weight the 2026 data implies crosses one-half around Week 4 or 5, as the backtest did Retire if After Week 4 the 2026 data still favors the preseason rating at every weight, which would mean the season disagrees with the backtest Blending makes the game predictions worse than the preseason rating on 200+ 2026 games Backtest over 2016-2025, 6,887 games between FBS teams: the blend beats the preseason rating by 0.82 points of average margin error (95% interval 0.68 to 0.94). The in-season rating on its own is 1.55 points worse than the preseason rating in Weeks 1-3, which is why it is blended rather than swapped in. First used entering Week 3, at a weight of 0.26 after 100 games. Next check: After the September 26 games (Week 4), with the same test run unchanged
	NBA	Offseason	Low to medium	sample of 81, won 71.6%, 95% interval 61.0% to 80.3%, break-even 52.4%, return +45.2%, as of 2026-06-14	The playoff-only sample reaches 150 bets with a win rate of 55% or more	The win rate drops below 55% on the next 50 picks	Resumes with the 2026-27 NBA season, late October
Our NBA totals model is well calibrated in the middle and overconfident at the extremes. Bets where it saw a 5 to 10 percentage point edge are the real edge; bigger claimed edges are noise. Promote if The playoff-only sample reaches 150 bets with a win rate of 55% or more The low end of the interval stays above 52.4%, the break-even at DraftKings prices Pattern replicates in regular-season post-fix data Retire if The win rate drops below 55% on the next 50 picks The low end of the interval falls below break-even Playoff-only cohort from the 2026-04-07 playoff model launch. The clock stopped with the season on 2026-06-14; the season audit confirmed totals as the model's one repeatable edge, and the cell resumes when the 2026-27 season tips in late October. Next check: Resumes with the 2026-27 NBA season, late October
	NBA	Offseason	Low	sample of 276, won 55.1%, 95% interval 49.2% to 60.8%, break-even 52.4%, return +10.5%, as of 2026-06-14	30 more days of in-season data move the low end of the interval above 52.4%	The low end of the interval drops below 48% after 100 more picks	Resumes with the 2026-27 NBA season, late October
Bets where the model claimed an edge of 10 points or more are the ones it is most confident in, but 60 days of results put the low end of the interval just below break-even. Either it is noise that fades, or more data separates it. Promote if 30 more days of in-season data move the low end of the interval above 52.4% The same pattern shows up in regular-season games alone Retire if The low end of the interval drops below 48% after 100 more picks Big claimed edges keep doing worse than moderate ones, which supports the 5 to 10 point entry above instead The low end of the interval, 49.2%, is 3.2 points below break-even. Don't bet it: the +10.5% return is real for the period, but the interval says noise is plausible. No new data until the 2026-27 season. Next check: Resumes with the 2026-27 NBA season, late October
	College baseball	Offseason	Low to medium	sample of 225, won 73.3%, 95% interval 67.2% to 78.7%, break-even 65.5%, as of 2026-06-21	Three weeks of DraftKings prices in 2027 show the cover rate holding at 70% or more	The cover rate on new games drops below 65%	Opening weekend 2027, mid-February
Heavy road favorites in college baseball (a rating edge of 3.0 or more in PEAR's college baseball ratings) covered -1.5 on 73.3% of 225 held-out 2026 games (95% interval 67.2% to 78.7%), and the rate held across years. The unknown is the price DraftKings charges for heavy road -1.5: break-even is about 65.5% at -190 and 71.4% at -250. The DraftKings price captures needed to work out the real edge only started in mid-May, and the season ended in June. Promote if Three weeks of DraftKings prices in 2027 show the cover rate holding at 70% or more The price at the time of the bet usually falls between -180 and -220 (break-even 64% to 68%) Retire if The cover rate on new games drops below 65% DraftKings consistently prices these games at -250 or worse, which eats the edge The cover rate is real, and it rises steadily with the size of the favorite. The 2026 season ended with the College World Series (Oklahoma won it); the price captures were too short to settle the question. The model is paused for the rest of 2026. Next check: Opening weekend 2027, mid-February
	College baseball	Offseason	Low	sample of 527, as of 2026-06-21	A correction by projection range cuts the average miss on held-out games by 0.05 runs or more	The refit doesn't reduce the average miss, or the bias turns out to be a quirk of the sample	Offseason refit before opening weekend 2027
On 527 held-out games, the totals model runs 0.41 runs high when it projects 11 or fewer, and 0.93 runs high when it projects 15 or more. That is a candidate for a calibration fix, separate from the ballpark effect we already track. Promote if A correction by projection range cuts the average miss on held-out games by 0.05 runs or more In 2027, the average miss on high-total games shrinks to within 0.3 runs Retire if The refit doesn't reduce the average miss, or the bias turns out to be a quirk of the sample Not a win-rate question: this is a check on the model's bias. The fix would be a refit of the model, not a new bet. The refit is offseason work and needs no live games. Next check: Offseason refit before opening weekend 2027
	MLB	Offseason	Low	break-even 52.4%, as of 2026-08-27	Live picks where the model is at least a run off the market win 56% or more over 80+ picks	The live win rate drops below 50%, which would mean the backtest fit to noise	Parked; revisited when the MLB feed is restored
The first-five-innings totals signal looked real in a backtest that used information a bettor would not have had in advance (1,141 games; a 59.5% win rate when the model was at least a run off the market). We collected live Kalshi prices from May 7. The live test was the checkpoint: does live data confirm the backtest, or did we fit to noise? Promote if Live picks where the model is at least a run off the market win 56% or more over 80+ picks The live pattern matches the backtest with similar strength Retire if The live win rate drops below 50%, which would mean the backtest fit to noise No live betting was ever connected, by design. The MLB model was paused on September 14 when the odds feed behind it stopped, so this sits where the backtest left it until the feed and the season are both back. Next check: Parked; revisited when the MLB feed is restored
	Weather	Closed	Low	sample of 475, won 23.8%, 95% interval 20.0% to 28.0%, break-even 23.9%, as of 2026-07-21		Retired July 21; see the calibration page for the closing note	None. Program closed.
On May 13 we raised the lowest probability the model would bet on from 40% to 45%, because an audit of 475 predictions found it 8.1 points overconfident at that floor. The check was whether predictions after the change came true as often as they said. They did not, and neither did ten other approaches. Outcome Retired July 21; see the calibration page for the closing note Closed with the program. Directional weather trading ended July 21 after 11 approaches failed out-of-sample testing; the model was switched off on August 28. The paper record showed an edge at prices nobody would trade (only 11.6% of live orders ever filled). Kept here as the closed entry it is. Next check: None. Program closed.

Four patterns are in test this football season; six sit paused or closed. The Calls record is on McConnell's Calls.

Notes

Running: the sport is in season and results are coming in. Offseason: paused; the evidence stays. Closed: the program is retired and the entry is kept to show how it ended.

Confidence. Low: wide interval. Low to medium: the direction looks real and needs more games.

Typed numbers carry the date they were checked. Entries that use one of the site's records read it live from the same source as the record page.

An entry leaves when it graduates to a published pick or the data rules it out. A ruled-out entry stays, marked closed.