September 20, 2026
Twenty-seven seasons, every number recomputed
The site said 24 seasons while it had tested 27. The fix was not to change the number but to rerun every study on the full replay. Here is what moved, what did not, and the one finding that flipped.
For two days the site quoted its backtest as 24 seasons, 1999 to 2022: FiveThirtyEight's published quarterback-adjusted Elo against closing spreads. That was the set the rules were fitted on. It was also wrong as a description of what the site had done, because Delta Sunday replays its own seasons from FiveThirtyEight's final ratings, and 2023, 2024 and 2025 had been in the player-system backtest since Friday. The builder put it plainly: this is Delta Sunday now, not 538. Tested seasons count.
The right fix was not to edit "24" to "27". It was to rerun every study the site quotes over the full set and print whatever came out. So that is what happened: one script, both tables, 6,918 decided games, every number on the site rewritten from its output.
The race
Points per season, all 27:
- The System 1,011
- The blend 1,003
- The closing market 997
- Elo 892
The shape is the one the site has always claimed: the market is the benchmark, Elo trails it by about a hundred a season, the blend edges the market, and The System's extras (the tight-agreement nudge, the lock press, the weather rule) add another eight on top of the blend. The gap between The System and the market is 14 points a season now. Over 24 seasons it was three. The three replayed seasons were good ones for the rules, which is worth saying out loud, because three good seasons is exactly the sample size that fooled us once before.
The rules, retested
Every rule survived. The blend-weight grid still peaks at a quarter Elo. The tight-agreement bump still wants two points, not eight: the eight-point version that looked wonderful on 2023 to 2025 loses 30 a season over 27, same as it lost over 24. The lock rule went 434 and 76, 85.1%, against a 79% forecast. The bold call went 246 and 264 while Elo claimed 65% for its side; the contrarian card stays a coin flip. When the models disagree by ten or more, Elo's side wins 50.3% of 1,102 games. The weather rule is worth eight points a season, with the same three effects at the same sizes.
What moved
Small things, in the direction you would expect from adding three seasons.
- Tight agreement: favorites won 69.8% against a 67.6% forecast over 2,607 games, up from 69.2% against 67.7% over 2,312.
- Byes. The published set had no rest data, so the bye study rested on 502 games from another source. The replay table now carries rest days from nflverse for every season, and with ten or more days of rest as the bye, home teams off a bye won 57.7% against a 60.3% forecast over 582 games; away teams 46.3% against 44.7% over 643. "About right" rather than "well calibrated". The 25-point bonus stays.
- The slices did not change their verdicts. Monday night: nothing there, 448 games. Thursday night: nothing, 267. The last regular week: not Elo's worst slice, 411. Divisional games: the home edge is about two points smaller, 2,198 games, still watching.
The one that flipped
The playoff finding read, over 24 seasons, that Elo and the market were level in January, 1,610 points to 1,617 over 248 games. Over 287 games it is not level: Elo scored 2,038 to the market's 1,917. Three recent postseasons went Elo's way by a lot. The blend at its normal weight scores the same 2,038, which means the quarter of Elo it already carries captures the whole effect, and the 75% playoff weighting that was tried and removed in September would still not be needed. The nudge is reworded; the rule is unchanged.
Why it matters
A model that quotes a number its own data does not match is a model that has stopped looking. The site's claim is that everything on it is computed and accountable. The 27-season figures are on the System page, the Nudges, the FAQ, and in every place an agent reads them, and the script that produced them is in the repository. When the 2026 season ends it becomes 28, and the numbers move again, and the page will say so.