Skip to content


I hope there are some people taking advantage of the NBA this season. The rating systems having been on fire against the spread for almost the entire season.
For example, the system average is
64% ATS over the last 2 weeks.
57% over the last month
59% over the last 2 months
56% over the last 3 months

turnovers and the accuracy of predictions

Yesterday I showed a simple plot of the line by the actual results for the 2014 NFL season. What this shows is that even in the best predictor of the final score, the Vegas line, there is is still a very large margin of error, about +/- 24 points per game. Is there anything that can explain most of or at least some of that error? At least one thing comes to mind and that is turnovers. Turnovers are considered to be random for the most part. And in fact the turnover margin is roughly normally distributed with a mean of about 0. A regression model shows that the turnover margin explained about 28% of the error in the line last year in the NFL.
So here is another plot, this one shows the average error for each value of turnover margin.
nfl turnover margin by difference in line

From this you can see that when the turnover margin is equal then the difference between the line and the final score is almost exactly 0. This is what I measure as bias in my prediction tracker results. And each turnover is worth approximately 4 points. But in reality only 20% of the games have a turnover margin of 0. The home team wins the turnover battle 40% and the road team the other 40%.
This second plot shows the average absolute error in the line by turnover margin.
absolute error by turnover margin

From here you see that when the turnover margin was equal the mean absolute error was reduced from the dotted line(11.26) down to about 7.6.
So what does all this mean? Well, turnovers do help to explain a chunk of the error in predictions. But even after accounting for turnovers, you have only reduced your margin of error from +/-24 points a game to around +/- 17 points. Which means there is still a lot of room for improvement.

football prediction

You might not realize it but predicting a football game accurately is a very difficult thing. Take a look at a plot of the line against the actual result from the NFL games this past season. Basically, the results is going to end up being the line plus/minus about 24 points.

line vs actual outcome nfl 2014

basketball results

I’ve made an improvement to the NCAA basketball results. The tables are now sortable by any of the columns. I will probably do the same with the NBA. Eventually it might also be nice to add a link to a daily updated plot of the results. They seem to give you an idea if any system is on a hot streak.

year to date NCAA basketball predictions performance

Given time I may try to start putting up some graphics. Here are how well the various rating systems have done through the course of the season to date in college basketball. I know it is kind of messy.

update: I increased the resolution of both these and the NFL plots below. You can right click them to pull them up full screen.

winning percentage over time

Against the spread over the season to date

Mean square error across the season to date

Bias across the season to date

Plots of NFL results over time

NFL systems Straight Up.

NFL system  winning percentage over time

NFL systems Against the Spread

NFL system ATS over time

NFL systems Absolute Error

NFL system absolute error over time

NFL systems Mean Square Error

NFL system MSEover time

NFL systems average Bias

NFL system average bias over time

Best NFL Systems

I get asked quite often which is the best system? Who should I follow, etc.
My typical response is that I don’t really believe there is any best system. There is something about the landscape each season that makes each season unique. Very few systems, if any, are at the top every year. I now have 15 years worth of results and I have tried to determine a way to say who are the best. To do this I decided to judge based only on second half results. A system needed to have made at least 100 predictions in the second half to qualify. Like I mentioned before, every season is different. So I didn’t want to just calculate the best overall winning percentage. Some seasons all numbers are high and other seasons all numbers are lower. In looking for something that I could compare relative performance year to year I decided to base these results on average percentile ranking. So if you were the best every year you would have a score of 1, the worst every year would be a score of 100.
I think I will only list those that are still making predictions.

Best ratings for straight up winners
1. Kenneth Massey (avg percentile 26.6. So note on average nobody is even in the top quarter every year!)
2. Computer Adjusted line/Updated Line/opening line (just rolling these all into one)
3. Game Time Decison
4. Beck Elo
5. Sonny Moore

Keep an eye on these, they could have been on the above list if they had more seasons:, Pi-Ratings, Lou St. John
And two good ones that are no longer around: Eric Hollobaugh and JFM Power Ratings

Best against the spread
1. Nutshell Retro (avg percentile 29.6)
2. Game Time Decison
3. Statfox
4. Sagarin overall rating
5. Roger Johnson

Systems to keep an eye on: Turnover adjusted Least squares,

Best at coming closest to the actual win margin
1. Computer Adjusted line/Updated Line/opening line (average percentile 9.0)
2. Game Time Decision
3. System Average or Median
4. Kenneth Massey
5. Dokter Entropy

Keep an eye on these,
Good but no longer around: Jeff Self, Hank Trexler

The accuracy numbers, absolute error and square error, do have certain systems that are typically closer to the top. But for either picking games straight up or against the spread there really is no always good system. The truth is that there is a huge margin of error on these things which means for the most part every system is within another systems’s margin of error. If a game, like tomorrow’s Superbowl, is a pick’em, then a 95% confidence interval on what the actual outcome will end up being is something like either team will win by 24 points or less.

NFL play by play data

I have never looked at NFL play by play data and am considering starting a new project to look at it. Does anybody know the best source for this data? I found Brian Burke’s data. Is there anything else, more recent and maybe with more info? Anybody know? Leave a comment.

NFL Divisional Playoffs

  home           p(win)   p(cover)   road            p(win)   p(cover)   line   lineavg

  Green Bay     0.59256    0.43811   Dallas         0.40744    0.56189    7.0    4.2221
  Denver        0.63827    0.46731   Indianapolis   0.36173    0.53269    8.0    6.5460
  Seattle       0.73402    0.51969   Carolina       0.26598    0.48031   12.0   12.8822
  New England   0.65145    0.49735   Baltimore      0.34855    0.50265    7.5    7.3808

NFL Wildcard playoffs

If the BCS was still being used the championship game would have been Alabama vs Florida St. With Alabama most likely winning easy and continuing the SEC overhype. I think now we know that all those BCS years the champ was still the ‘mythical’ national champion.

Last week the top 3 NFL picks went 3-0 to finish the regular season 30-18, 62.5. Last season the top 3 were 29-18-1. So two very similar seasons back to back. The top game of the week was 10-6, also 62.5%.
So why are you paying for a service and still getting a losing record?
Interestingly two years ago the top 3 NFL games were only 14-28. Which is about what happened with the NCAA numbers this season.
Only one of the games is worth considering this week, Arizona at Carolina playing again with their third string qb.

 home            p(win)   p(cover)      road       p(win)   p(cover)   line    lineavg

  Carolina       0.47526    0.32776   Arizona      0.52474    0.67224    7.5   -1.11671
  Indianapolis   0.60211    0.53952   Cincinnati   0.39789    0.46048    3.0    4.80465
  Dallas         0.63076    0.47145   Detroit      0.36924    0.52855    7.5    6.21634
  Pittsburgh     0.55781    0.49184   Baltimore    0.44219    0.50816    3.0    2.63254