Project TD Tuning — Version 2

Version 2 — What changed from V1 V1 stopped after one week. 47% correct, and yeah, that looks bad. But it wasn’t really the model’s fault — it was a cold start. Week 1 of 1999 is literally the first game in our dataset. The model had nothing to pull from. No prior scores, no win rates, no sense of which teams were any good. Every single prediction came out to roughly “home team by 3.” That’s not a model, that’s a coin flip with home field. Same weights, same factors in V2. We just added a 4-week buffer before the stopping rules kick in. Let the thing get its footing before we pull the plug. Seemed fair.