Hello oh great and patient fans. I haven’t been posting these last few days because a) I’m working on a couple changes to the weighting of certain formulaic pieces in preparation for the 2020 season based on learning from this year and correctly some mathematical logic that skewed some teams up or down a tad more than they deserved, b) I’m working on looking back over all of the games played (not including non-division I opponents) and seeing how the current predictions would be against the actual outcomes (I’ll talk about that more in a minute), and c) I’ve been debating how to blog the predictions and outcomes for the bowl games. Let’s talk about that first.
I have all of the games (except the unknown rounds of both playoffs) coded into the GCS and, therefore, have pct to wins for each game. In a normal week, the number of non-Saturday games is so relatively small that it’s not really worth resetting the predictions for the next day (it’s also possible that I’m just too lazy to do it properly). Bowl games are different. Some days have 1 game, some have a bunch. Each game changes the probabilities, potentially even the predicted victor, in some cases by a substantial margin. Since the bowl season is simultaneously the most and least important games of the season, depending on the teams, I guess, I decided that there will be predictions on a day by day basis. Tomorrow, I’ll post the 12/20 games (both of them). Then each evening before any games, I’ll post those. So… more posts, but all of them on a few games. Should be fun.
Over the course of the season, or at least the last 2/3 of it, I’ve kept track of the predictions and shown the model is sound (the GCR picks the “correct” number of winners based on the percentages to win). But, I’ve been thinking about how good it is in arrears: meaning, now that the season is almost over, how did the GCR predict every game in hindsight. Predictions, like the rest of the GCR, adjust each week as games are played. What could have been a 74.5% chance at the time could have shifted, based on results of future games up or down a lot – again, even shifting the “predicted” winner. But the formula never changes, so I thought it would be interesting to see how good the model gets with the tons of data it has now. The other fun thing I’m doing is using the 2020 model against the 2019 season to see if there is a difference in the predictions – kind of a calibration. I’ll publish information on the outcome of all of this (it’s rather time consuming to capture) as well as the “real” predictions at the end of the bowl season.
One more thing about the changes in the 2020 GCR: I’ll continue to add and adjust. The basic formulas are the same, but I will try to explain all of the changes. I do not share the formula, not because of copyright or trust or anything like that, but because it’s really large (a 12.5 MB excel file) and has a lot of moving parts. It’s always about improvement. We are either actively getting better or we are passively getting worse.
Happy holidays to everyone whatever you choose to celebrate and we will start this party up tomorrow night.
Thanks, G
