The regular season is in the bag, and before moving on to the playoff predictions (and how sad they look for the Lions), I need to wrap up the models’ performance this year. It will be a sad tale of woe, but there is no avoiding it. As I noted when making the predictions, week 17 is a tough one because the model doesn’t know who is in the playoffs and who isn’t, and thus who is resting and who is playing hard. Let’s see if it led to a bad week.
Bill Simmons could have missed every pick this week and still beat me handily. In actuality, he went .500 (8-8). So did Luigi and Yoshi 2, while Yoshi 1 went 6-10. Bill finishes up at 134-116-6. Checking back on last year’s summary, we see that Bill had a four-year record of 479-488-25. Tacking on this season, Simmons is at 613-604-31. Assuming you bet $100 on every game (which you wouldn’t) and got them all at -110 (which you couldn’t), that record would have cost you about $4600 over five years, or about $3.75 per game. If you’re betting for fun, that really isn’t too terrible. People probably spend more than $900 a year on hobbies. But if you’re looking to make money, that isn’t going to cut it; of course, you wouldn’t bet every game trying to make money.
Luigi’s 8-8 brings it to 122-128-6. The previous years’ numbers are different from Simmons’ because I don’t always include week 1, when the models don’t have data to go by. Luigi’s five year record is 612-574-32. Under the same conditions as before, you would have lost about $2200 over five years, or $1.80 per game. Nearly all of that (about $1800) is just from this season.
Yoshi 1’s 6-10 brings it to 116-134-6 and a three year total of 350-356-16. With the differing number of years it only makes sense to compare performance using money per game, which is a loss of $5.60 per game. Two of the last three years have been a little rough, but that seems like a lot.
Yoshi 2’s 8-8 brings it to 116-134-6 and a three year total of 330-376-16 and a horrendous per-game number of a loss of $10.85 per game. I’m not sure why Yoshi 2 is so bad against Bill’s lines, and so much worse than Yoshi 1 over the same time period. Let’s see how some of the other numbers pan out before jumping to conclusions.
Going against the public/SBR lines, Luigi went 8-7 for a season total of 98-93-7. The five-year record is then 476-448-27, which has a per-game value of a loss of $1.60. That isn’t good, obviously, but shows that there’s a little value in a combination of a) using different lines and b) eliminating some close-call games.
Yoshi 1 went only 6-8 for a season record of 87-104-7. Its three year record is 282-295-19 for a per-game value of a loss of $6.82. That’s worse than against Bill’s lines and is theoretically being more selective, which is bad.
Yoshi 2 went 7-7 for a season record of 88-103-7. Its three year record is 268-295-18 for a per-game value of a loss of $9.17. This is relatively in the expected direction (i.e. better than picking every game like Simmons does), but obviously still not very good.
Consensus picks went 6-7 for a season total of 63-70-6. I only had last year’s results in the summary, so the two-year total here is 129-130-7. Given that it was a bad year overall, and last year was middling at best, this isn’t too surprising.
Luigi’s SuperContest picks went 2-3 for an overall record of 38-46-1. That’s pretty bad, but the five year record is 221-187-12 for a per-game value of $3.31. These picks are the closest to what you would actually do if you were picking games (although still not quite; you would pick your most confident games but it wouldn’t always be five games each week). Yoshi 1 went 1-4 for a season record of 38-45-2, and a three-year total of 122-127-6. That isn’t good, but the per-game value of a loss of $6.30 is between the numbers for SBR and Simmons. Yoshi 1 seems to be most affected by changes in the line. Yoshi 2 went 2-3 for a season record of 41-43-1 and a three year total of 130-118-7 for a per-game value of 7 cents per game. This is much better than the numbers above, and why I didn’t want to jump to conclusions. For whatever reason, the last few years have been shaky while the 2009 and 2010 seasons went fairly well. So having the most likely games to bet on break even over the last three years is nothing to sneeze at.
To sum up, it was a rough year. The SuperContest picks had the best relative results, which is as it should be and a good sign. The Yoshis are not doing as well as Luigi, but they’ve also only been running during poor years for everyone involved, so it’s hard to say if they just aren’t working well or if they’ve been unlucky. Yoshi 1 does better than Yoshi 2 when picking lots of games, but Yoshi 2 has been better when picking its most confident games. Apparently something interesting happens when you use an entire season to fit the model (as Yoshi 2 does) as opposed to just the week of interest (e.g., use only previous week 2 games to predict a current week 2 game, as Yoshi 1 does). I’ll have to see if that trend continues.