Friday, February 3, 2012

NFL Turnover Differential and the Point Spread


In the comments section of the Advanced NFL Stats post on the conference round post-season projections, an anonymous poster claimed the following:
For a very long period, winning NFL bettors made money simply by betting *against* the team with the better turnover difference. That's how drastically the public has overestimated the predictive value of turnovers. 
NFL turnovers are far more random than most people realize, and only correlate weakly with prior game performance.  Teams that have achieved strong season-to-date records by "winning the turnover battle" are more likely to regress to the mean.  Vice versa for teams with bad records due to poor turnover differentials.


I was curious to see if this market inefficiency actually existed and at what level.  A quick google search on the topic yielded conflicting results.  According to one site, teams with the unfavorable season-to-date turnover differential covered the spread more than 75%(!) of the time.  I would be skeptical of any statement of the form "When [simple rule], the spread is covered [x%] of the time", where [x] is anything north of 60%.  Especially if the rule can be applied to virtually every game.  A different site appears to claim just the opposite: you should try to pick teams that protect the ball well, since teams that win the turnover battle tend to cover the spread (then again, this site also claims to tell you how to pick NFL games "accuratley").



The Data


Using data from Killer Sports, I was able to look at performance against the spread for all NFL games going back to 1998 (excluding games for the first three weeks of the season).  I stratified the games according to the teams' relative turnover differential.  The ">=1 " column shows performance against the spread for all teams whose season-to-date turnover differential was at least 1 turnover worse than their opponent's.  The ">=5" column shows performance for teams whose turnover differential was at least 5 turnovers worse, and so on.


The data shows a weak correlation between covering the spread and unfavorable turnover differential.  If you had bet on every team with an unfavorable differential, you would have won 51.4% of the time.  Unfortunately, you need to win 52.5% of the time in order to break even.  The correlation does get stronger if you restrict to teams with a wider gap.  Greater than 5, and you're up to 52.0%.  You're in to positive returns for greater than 10 (54.4%) and 55.8% for greater than 15.  Somewhat encouraging.  Here's the data:

'Unfavorable' Turnover Differential
>=1 >=5 >=10 >=15
Season Games %Cvr Games %Cvr Games %Cvr Games %Cvr
1998 171 55.3% 108 53.7% 47 57.4% 13 65.4%
1999 179 54.7% 117 57.7% 50 64.0% 26 59.6%
2000 179 46.9% 115 46.5% 63 40.5% 34 42.6%
2001 178 59.8% 115 61.7% 60 64.2% 23 65.2%
2002 188 55.3% 108 53.7% 53 53.8% 24 52.1%
2003 185 51.1% 120 52.5% 58 64.7% 21 78.6%
2004 184 46.2% 120 46.2% 76 47.4% 31 51.6%
2005 188 48.1% 122 43.0% 76 45.4% 35 44.3%
2006 182 53.3% 106 53.3% 47 64.9% 19 60.5%
2007 179 46.4% 111 49.1% 54 47.2% 23 54.3%
2008 182 51.9% 107 53.7% 40 58.8% 15 73.3%
2009 182 52.5% 110 55.5% 55 60.0% 15 60.0%
2010 188 50.3% 120 52.1% 50 50.0% 16 43.8%
2011 187 48.7% 129 50.8% 58 54.3% 25 54.0%
Total 2552 51.4% 1608 52.0% 787 54.4% 320 55.8%
1998-2004 1264 52.7% 803 53.1% 407 55.3% 172 57.3%
2005-2011 1288 50.2% 805 50.9% 380 53.6% 148 54.1%

Defensive Turnover Differential


I also looked at performance when focusing just on offensive or defensive turnovers. You can quickly go insane searching for hidden patterns in point spread performance data (e.g. "Road dogs coming off a 8+ point loss at home against a non-divisional opponent are 24-15-2 ATS when the over under is at least 50 points", and so on).  For that reason, I was hesitant to start slicing and dicing the data.  You cut the data enough times, and you're bound to find a correlation just due to random chance.  


But I had a rationale.  Based on data I've looked at, offensive turnovers are more consistent than defensive turnovers (but I've seen conflicting thoughts on this).  If defensive turnovers were more random than offensive ones, you should be able to find better betting opportunities just looking at defensive turnover differential.


The results show an even stronger correlation than total turnovers.  For games with at least a 10 defensive turnover differential between the two teams, the unfavorable team covered 58.7% of the time.  In addition, such a strategy would have paid off for each season, save for two.


I plan on testing this strategy out for the 2012 NFL season, using the ">=10" rule for defensive turnovers only.  I'll report back here on results.  From prior seasons, it looks like you can expect about 20-25 betting opportunities over the course of the season.

'Unfavorable' Defensive Turnover Differential
>=1 >=5 >=10 >=15
Season Games %Cvr Games %Cvr Games %Cvr Games %Cvr
1998 167 54.2% 80 58.1% 20 80.0% 5 70.0%
1999 173 52.9% 80 58.8% 21 57.1% 3 33.3%
2000 175 44.9% 92 44.0% 28 41.1% 11 40.9%
2001 174 58.3% 73 61.0% 23 69.6% 6 50.0%
2002 170 55.3% 63 54.8% 18 38.9% 5 60.0%
2003 181 51.1% 83 50.0% 29 56.9% 8 37.5%
2004 184 48.9% 91 49.5% 34 52.9% 3 33.3%
2005 188 49.7% 118 46.2% 37 58.1% 7 57.1%
2006 177 55.1% 75 50.7% 22 59.1% 4 37.5%
2007 178 51.1% 69 55.8% 20 67.5% 2 50.0%
2008 182 55.2% 72 59.0% 9 61.1% 0 0.0%
2009 169 56.8% 62 49.2% 15 60.0% 5 80.0%
2010 183 53.3% 69 51.4% 16 59.4% 0 0.0%
2011 186 52.2% 88 58.0% 20 70.0% 1 0.0%
Total 2487 52.7% 1115 52.9% 312 58.7% 60 49.2%
1998-2004 1224 52.2% 562 53.3% 173 56.1% 41 46.3%
2005-2011 1263 53.3% 553 52.5% 139 61.9% 19 55.3%

Offensive Turnover Differential


And here are the offensive turnover results.  As you can see, performance against the spread is inconsistent.

'Unfavorable' Offensive Turnover Differential
>=1 >=5 >=10 >=15
Season Games %Cvr Games %Cvr Games %Cvr Games %Cvr
1998 174 49.1% 83 50.0% 21 42.9% 3 33.3%
1999 178 55.3% 77 46.1% 25 62.0% 3 100.0%
2000 175 51.4% 71 43.0% 24 43.8% 6 58.3%
2001 178 57.6% 90 55.6% 23 54.3% 8 50.0%
2002 180 54.4% 80 55.0% 25 64.0% 4 50.0%
2003 179 53.1% 88 55.7% 19 57.9% 2 100.0%
2004 183 49.2% 89 38.8% 22 43.2% 6 58.3%
2005 186 42.5% 80 38.1% 24 43.8% 2 50.0%
2006 176 52.6% 80 56.2% 19 57.9% 6 50.0%
2007 180 46.1% 92 40.2% 22 50.0% 6 66.7%
2008 175 48.0% 82 50.0% 15 80.0% 2 100.0%
2009 177 51.7% 78 50.6% 22 40.9% 3 100.0%
2010 182 49.7% 97 54.6% 33 36.4% 11 18.2%
2011 178 46.6% 72 45.8% 17 50.0% 1 100.0%
Total 2501 50.5% 1159 48.7% 311 50.8% 63 55.6%
1998-2004 1247 52.9% 578 49.3% 159 52.8% 32 59.4%
2005-2011 1254 48.1% 581 48.0% 152 48.7% 31 51.6%


8 comments:

  1. Some of the older research on the topic...when this strategy worked quite well:

    http://www.twominutewarning.com/cbjumping.htm

    http://twominutewarning.com/v04csm.htm

    Would love to see a follow up post breaking down further the ">=5" bin to see where between 5 and 9 preformance starts to improve. Also a look at favorites and dogs would be cool.

    ReplyDelete
  2. Thanks. Interesting stuff. Not turnover-based, but similar in concept. Although I doubt there is truly a "bridge-jumper" phenomenon in football betting. That's like playing Russian Roulette with half the chambers loaded.

    Here's how the numbers break out from >=5 to >=15:

    >= %Cvr
    5 52.9%
    6 53.6%
    7 54.2%
    8 56.3%
    9 55.1%
    10 58.7%
    11 57.6%
    12 57.3%
    13 53.9%
    14 52.4%
    15 49.2%

    But I would caution against trying to pick the best cut of the data. You'll fall into the multiple endpoints trap.

    It looks like I may have gotten lucky with the choice of >=10. Based on that, I wouldn't expect long term performance at 58.7%.

    ReplyDelete
  3. My bad. I sent the wrong link from that site. Here is the correct link:

    http://www.twominutewarning.com/g13.htm

    It might worth investigating performance using net turnovers/game (after week x) rather than raw net total turnovers.

    ReplyDelete
  4. A few more articles:

    http://www.twominutewarning.com/g2003/gs8.htm

    http://www.twominutewarning.com/g2003/gs7.htm

    http://www.twominutewarning.com/g2003/gs18.htm


    Obviously, all of this data is ancient, but gives an idea of what people have looked at in the past.

    ReplyDelete
  5. Don't mean to split hairs and I should qualify first by saying - great post, very thought provoking. Few points.

    First - the break even rate for picking ATS is more like 52.36 and that is in a vaccum (i.e. in a single week). To be fair to readers whose behavior may be influenced by your advice, if in 2011-12 you picked 134.03 games right ATS you would have attained the break even rate of 52.36. However, that is assuming you had zero standard deviation - more clearly you picked at a 52.36 rate week in week out without variation. Personally, a more accurate measure of ATS performance is cumulative wealth or at what rate you can generate profits above your initial investment. By doing this you account for the natural volatility of ATS picking and the compound effect of money.

    For example, I track 69 different NFL predictive ATS algorithms,18 of which bested the 52.36 break even rate. However, only 10 ATS algorithms ended the 11-12 season with values greater than their initial investment. And some barely drew profits. This is due to the natural swings in ATS picking.

    Second. An interesting question is how turnovers affect the intrinsic value of a team as opposed to the matchup. It's telling that we often search for similar matchups to predict future performance. Where as we spend less time determining how a teams propensity to give away possessions and turn the ball over contributes to their on the field value. I understand that we are looking across a broad set of data and we hope that aberrations regress to the norm, but turnovers appear to be much more circumstantial.

    ReplyDelete
  6. "An interesting question is how turnovers affect the intrinsic value of a team as opposed to the matchup."

    What do you mean by this? Turnovers are generally assumed by casual fans to be ~100% the product of the player/team involved. What this research says is that the market is still over-crediting the player/team with respect to turnovers forced.

    Also, this post is simply pointing out an anomaly. Your points about money management and the inherent volatility of betting on binary outcomes are not relevant to the research.

    ReplyDelete
  7. New mexico, California, TCU, Ohio Condition, Wake Forest,

    Virginia Technical, Texas Tech, Xavier and The south all play game titles that will either help solidify their at-large berth/seed or make Selection

    Weekend more of an troubled

    March Madness
    March Madness Live
    March Madness Live Stream
    March Madness Bracket
    March Madness 2017
    March Madness 2017 Live
    NCAA March Madness
    2017 March Madness
    march madness schedule
    march madness schedule 2017

    ReplyDelete
  8. Idaho: Even though a Tar High heel wasted inside ACC contest semifinals, on the list of Duke relating to the ropes designed for most of the main half. Justin Fitzgibbons, who has formulated a experienced outside sport, and then a foul-free Fran Berry II are the recommendations.
    March Madness

    March Madness Live

    March Madness Live Stream

    March Madness 2017

    Certainly, we realize Recognise Number of has not made one further 3. This approach group, stacked by using proficient scorers that can also lock down attackers, are going to be her initial
    March Madness 2017 Live

    ncaa March Madness

    ncaa march madness live

    ncaa tournament

    March Madness Bracket

    ncaa final four

    When you’re still in the containment system about that 8-9 game, or are trying to find your complete viewpoint onto your Very last Four picks, we’ve got most people covered using a search engine ranking off 68 competitors inside the competition. You can now with certainty generate your decision concerning some of those mid-major darlings you’ve already been eyeing to get upset possibilities.
    March Madness

    March Madness Live

    March Madness Live Stream

    March Madness 2017

    March Madness 2017 Live

    ncaa March Madness

    ncaa march madness live

    ncaa tournament

    March Madness Bracket

    ncaa final four

    Villanova: The top overall seedling. Quick grown timbers . team’s placement in the most difficult region (the East, with matchups next to Duke in addition to SMU looming), Villanova’s three-man a particular problem core from Jalen Brunson, Josh Hart and additionally Kris Jenkins gets the likely to help reiterate like national champions.

    ReplyDelete