Project TD Tuning — Version 6

# V6 Context V5 ran for nine weeks and hit 71% in Week 8 — the best single-week result we’ve seen. Then Week 9 came in at 36% and the model stopped. That’s the part that needs fixing. The single-week floor existed to cut off versions that weren’t working at all. But V5 was working. A 57% rolling average with a 71% peak doesn’t describe a broken model. It describes a model that hit a rough week. The floor was a safeguard against failure, and it fired on success. V6 lowers that threshold from 50% to 40%. One bad week can no longer end a run that’s otherwise performing well. The rolling 4-week average is still in place — that’s the real signal. A single outlier week no longer outweighs it. The second change is how the model measures team strength. Until now, “power differential” was a shortcut: subtract each team’s average scoring margin over recent games and compare. It’s a reasonable proxy, but it’s blind to schedule. A team going 6-2 against bad opponents looks identical to a team going 6-2 against strong ones. V6 replaces that proxy with a Simple Rating System. SRS works by iterating: a team’s rating equals its average scoring margin adjusted for the ratings of the teams it played. Beat a strong team by a lot, and your rating goes up more than beating a weak team by the same margin. The math converges quickly — ten iterations is enough — and it produces numbers that actually mean something about how good a team is relative to everyone else in the league. The two changes work together. A more accurate picture of team quality means better predictions on the weeks that matter. A lower single-week floor means we get to run long enough to see whether they matter.