Project TD Tuning — Version 5

V4 answered the cold-start question. Starting with a full season of history, the model had something to work with. Week 2 hit 60%, which is the best single-week result we’ve seen. But it still stopped. One bad week brought the rolling average down far enough to trip the wire. Two things happen in V5. First, we gave the model more runway. Four weeks isn’t enough to know whether you’re dealing with a broken model or just a rough stretch. Eight weeks gives it room to prove itself — or not. Second, we turned on a set of factors that have been sitting dormant since Version 1. They were always in the model, always carrying weight, but they had no data behind them. Those factors are about the things that define a team in a given season — not talent on paper, but how they’re actually playing together. Are they consistent? Are they clicking at the right moments? Do they hold it together when the game is on the line? A roster full of names means nothing if the timing isn’t there. V5 finally connects those questions to real answers.