The first week is over for the model, so it’s time to see how we did. Recall that I expect Luigi (the less complex model) to do better at picking winners and the over/under, but Mario should do better at picking against the spread. At least that’s how things have gone the past two years, and I think 500 games is a pretty good sample size. There are two ways to evaluate how the models did. One is simple accuracy, and the other is how much money would be made if they were used for betting. So, in order of the presentation from the predictions post, here are the results:
Over/Under: Mario 8-8, Luigi 9-7. Both models were wrong about Dallas-Chicago, which was the only game to go off at -115 on the over, meaning -105 on the under. So if you had bet $100 per game you’d lose $100 on that one game and go 8-7 or 9-6 on all the -110 games, where the pay odds are the same regardless of which side you took (over or under). Thus you’d lose $800 (or $700) on the losses and win $727.27 (or $818.18) on the wins. So Mario would have cost you about $70 and Luigi would have won you about $120.
Moneyline: I should have kept track of the odds for the away team, but I didn’t, so I can only evaluate the betting outcomes for games where the models picked the home team. In terms of just picking a winner, the ROC for Luigi said that the best criterion is to take the home team if they have a probability of 43.2% or better. Using that criterion for both models, Mario went 8-8 and Luigi went 6-10. Mario took Cleveland, Dallas, Carolina, Denver, and the Jets as home bets (also Cincinnati, Green Bay, and Washington, but there were no odds listed for those games). Denver and the Jets came through at -180 and 135, meaning that $100 on each game would win $55.56 and $135. The losses on the other three games would cost $300, for a net of a loss of about $110. Similarly, Luigi would have taken Cleveland and Detroit and lost both. However, the models can’t be fully evaluated because I don’t know how they would have done when they picked away teams, which would have had better pay odds in general. So next week will be better, both in terms of having all the information and having more games in the sample.
Spread: Mario went 8-7-1 (meaning one push), as did Luigi. Sadly, that wouldn’t be enough to make money with my picks, although you’d roughly break even (depending on exactly how much you bet and if you adjusted your bets using the Kelly criterion or the like). If you managed to get every game at -110, you would come out just ahead, but that doesn’t happen very often. Instead you might pick Detroit to cover at -120 (winning you $80 instead of $91), or Miami at -115 (winning you $87). So this week was close but no cigar.
Bill Simmons: Bill’s lines were close enough to Bodog’s that the models had the same predictions. The results turn out to be the same, so both went 8-7-1. Bill also went 8-7-1. Including week 1 (using a different model), I’m a game ahead. I expect my advantage to grow over time.
Overall, the models had a weak start to the season. They were treading water at best, but at least they weren’t catastrophically wrong. I’m going to be optimistic (and try not to think about the gambler’s fallacy) and say that the models are due for a better time next week.