Overall model score isn't the most important thing IMO. Someones model might win 65% of the time but if the 50% confidence projections and the 80% confidence projections all win 65% of the time then it isn't really as helpful. Personally I focus on finding that differentiation and making sure the 80% projections actually win roughly 80% of the time, 70% wins 70% of the time, etc.
Spread and totals models are the toughest to build. You're basically putting the model you built up against the model that the country's largest sportsbooks spent tens of millions of dollars (if not more) building. An additional fraction of a percentage point in your direction equates to real money in the long run.
As @THC_IPA said, it also helps to see how your model does at certain confidence intervals. If you hone in a few games each night that look particularly interesting, you could end up doing even better than your model score because unlike the books who put up a line for every game, you just need to bet the ones you like.
So what's the next step here?
The top 9 NFL ATS prediction models are all within 1% of each other. What is the next step to find a more significant percentage of picking winners?
Win your way.
Your AI model belongs to you.
Spread and totals models are the toughest to build. You're basically putting the model you built up against the model that the country's largest sportsbooks spent tens of millions of dollars (if not more) building. An additional fraction of a percentage point in your direction equates to real money in the long run.
As @THC_IPA said, it also helps to see how your model does at certain confidence intervals. If you hone in a few games each night that look particularly interesting, you could end up doing even better than your model score because unlike the books who put up a line for every game, you just need to bet the ones you like.
Hope this help!