Project TD Tuning — Version 4

Three versions in. Three times the model stopped in the opening weeks of 1999. That’s not a coincidence. The first few weeks of 1999 are also the first few weeks of our entire dataset. There is no prior history to draw on. No record of how these teams performed. No trends. No recent results. The model was being asked to predict games in a vacuum, and it showed — every version collapsed before it ever got going. We kept adjusting things: thresholds, stopping rules, a couple of factors that were pointing the wrong direction. And some of that helped. But you can’t tune your way out of having no data. The fix for V4 is straightforward. Instead of starting at the very beginning, we let the model watch a full season first. It gets to see how 1999 actually played out before we ask it to perform. When the grades start, it has something to work with. Same model. Same approach. We just stopped asking it to predict before it had seen anything.