Last week there were more differences between the model's predictions and Vegas lines than I could remember. I happened to be going to Vegas, so I had plenty of options for games to bet, and bet I did. There is more detail below for interested parties, but spoiler alert: the results were disappointing. It could be game level variance, it could be that the model is missing key info on some of the teams I bet, it could be karma (remember my warnings about Vegas overconfidence!). Whatever it was, I had a looooong day on Saturday.
This week though, the model is in lockstep with Vegas. The model's correlation with Vegas lines this week is 0.985. It was 0.961 last week (week 3). Those two numbers may seem close but look at how far each is from 1 (perfect correlation). Week 3 is over twice as far away. Suffice it to say my model and Vegas disagree on very few games this week (UW vs. Cal being one of them; go dawgs!).
On one hand that makes me feel good. Vegas lines, especially those with a lot of action on them, are well-calibrated, and aligning closely with them is good. On the other hand, lines shouldn't be taken as canon and one clear goal in building a college football model is to beat or competes with, not match, professional odds-makers.
Week 4 looks like a standard awesome college football week, propped mostly by the Pacific 12 Conference which figures to just deliver. USC/ASU, UCLA/UA, Utah/UO and Cal/UW? How can you go wrong?
On one hand, I was down betting college football. On the other hand I more than made up for it with poker, so that's something. College football results were disappointing though. If I had my druthers I would have crushed college football and lost at poker, I know I'm a winning poker player; the jury is still out on my model's predictive value compared to Vegas lines.
I bet 9 games, and won 3 of them, losing roughly 1/3 of the money I bet. I would categorize my results into the following categories:
- Won in OT and shouldn't have been a sweat
- Syracuse, Kansas State, and UTEP were all favored by Vegas and heavily favored by the model, they were each taken to overtime by their inferior opponents where they eeked out wins
- Lost in close or in OT as an underdog
- Colorado State (OT), BYU (by 1) and Texas (by 1) all lost super close games as small to large underdogs. So on one hand it's validation of the model that I bet them as underdogs and they played good games, but the other hand they don't pay out for validation of the model, I asked.
- Couged it
- Cougs (couldn't beat Wyoming by more than 17), Auburn (LOL), and Texas State (lost as a favorite at home)
Notes on the data presentation
- A team is shaded green according to their chance to win (darker = better chance)
- I'm experimenting with new names for the Watchability index. I really like the idea behind that number, but the name has never sat very well. I'm currently testing out Expected Game Quality, which I like for its simplicity. Don't hesitate to share comments on the new name or suggestions for other names
- Expected Game Quality (formerly Watchability) is a combined measure of how good the teams are and how likely the game is to be close. Put another way: it's an estimate of how likely the game is to be a close, well-contested game
- This post has more detail on the math behind Game Quality
Post a Comment