I’ve done some work in the off-season to improve that quality of week 1 to week 3 predictions. In these weeks there are not enough connections between the teams for an unbiased model. Thus some sort of bias needs to be introduced to be able to compare all teams. The best bias discovered so far is to use last year’s power scores.
Here is the data so far this year:
Week Games Correct 0 24 0.667 1 176 0.705 2 173 0.769 3 173 0.809 4 175 0.817 Total 721 0.771
Data shows a nice gradual improvement in performance. I expect 0.817 to be a relatively high number; probably some week in the future will have a correct ratio less than 0.817.