Wednesday, December 12, 2012

Turnover Index - Week 15

Here are the Turnover Index numbers for Week 15 of the NFL season.  The purpose of these weekly posts is to find betting opportunities that exploit the market's tendency to overvalue a team's season to date defensive turnovers.  See here and here for more background.

Week 14 Performance

Results were split for week 14, going 3-3 against the spread.

Turnover Index through Week 10 (Against the Spread): 14-10-1

Overall, that is 58% against the spread, still consistent with the results from 1998-2011.  But recent weeks' picks have not performed as well as the early season picks.

Week 15 Picks

Once again, there are six betting opportunities this week.  As a reminder, the criterion we use is any game in which there is at least a 10 defensive turnover differential between the two teams.  We bet on the team with the lower amount of turnovers.  See the table at the bottom of the post for the details.

  • Cincinnati @ Philadelphia - Pick: Philadelphia
  • NY Giants @ Atlanta - Pick: Atlanta
  • Detroit @ Arizona - Pick: Detroit
  • Green Bay @ Chicago - Pick: Green Bay
  • San Francisco @ New New England - Pick: San Francisco
  • Indianapolis @ Houston - Pick: Indianapolis

Here is the complete table of matchups:


  1. > Vegas appears to give the Seahawks a 3.5 point
    > home advantage. They're number 1. And the
    > Giants, Saints, and Bengals are at the bottom,
    > closer to 2 points.

    That's a way smaller range than what my
    regression produced. IIRC The G-men are at around
    -1 (yes, that's negative) at one end, and the
    Vikings are around +6.

    What I did wasn't exactly rocket science, so this probably speaks to the bookies balancing the
    action, rather than finding true odds.

    I was going to ask how you generated the numbers,
    but it occurs to me that I could just do a
    regression on the spreads myself.

    Do you know how the NFL generic points favored
    that you show on the ticker are derived?

  2. Running the spread regression against the score
    regression suggests taking New England and Oakland
    on the road, and that San Francisco and New Orleans
    might not be as good as I thought.

    This information could be used to drive a
    somewhat more sophisticated betting strategy...

    Team Spread HFA Score HFA HFA Correction
    NYG: 2.52 -1.43 -3.95
    CAR: 2.93 -0.88 -3.81
    PHI: 2.49 0.26 -2.24
    NE: 2.64 0.47 -2.17
    CIN: 1.82 -0.34 -2.15
    OAK: 3.11 1.00 -2.11
    NO: 2.12 0.33 -1.79
    MIA: 2.97 1.36 -1.61
    WAS: 1.99 0.85 -1.14
    CHI: 2.94 1.90 -1.04
    NYJ: 2.97 2.03 -0.94
    TEN: 2.30 1.81 -0.49
    CLE: 2.48 2.02 -0.46
    ATL: 3.28 2.91 -0.37
    TB: 2.43 2.23 -0.20
    DEN: 3.12 3.05 -0.07
    GB: 2.61 2.57 -0.04
    BUF: 2.61 2.61 0.01
    PIT: 2.07 2.27 0.20
    HOU: 2.53 2.79 0.26
    DET: 1.96 2.58 0.62
    SD: 3.11 3.85 0.74
    IND: 2.49 3.32 0.83
    JAC: 2.82 3.88 1.07
    KC: 2.10 3.65 1.55
    ARI: 2.26 3.91 1.66
    DAL: 3.37 5.17 1.80
    SF: 2.65 4.59 1.95
    STL: 1.77 4.25 2.47
    BAL: 2.95 6.25 3.30
    SEA: 3.03 6.34 3.31
    MIN: 3.01 6.53 3.53

  3. I looked against win/loss rate against the spread, and, while that list is somewhat predictive, I noticed some oddities. From 2000-2011, Saint Louis 78-115-7 against the spread with a significantly losing record both at home, and on the road. Dallas, while near the end of the list, does average against the spread, and a couple of teams not on the list have strong or weak records against the spread. Pittsburgh is 60-47-2.

  4. I've just started sports betting and have found this site to be really useful. Wondering if you can help validate I'm using your numbers right. Let's use TB@NO. The GPF has NO as a 1.1pt fave, but accounting for Nate's HFA (-1.47 for NO) results in TB being favored by .5 pts. Since the line is currently TB+4, I should grab that right? Did I miss anything or forget to consider other factors? Thanks for the help guys, this is super interesting stuff.

  5. Actually, I should be using the middle HFA column, right? So the TB/NO should really be NO -1.43, right? 1.1GPF + .33HFA?

  6. I picked TB +3 in an earlier post. IANAL YMMV No Liabily. Void where prohibited...
    For what it's worth, I don't bet on sports -- If I look at a line, I have to spend a few minutes decyphering it.

    Column 2 is the 'Vegas' home field advantage for the team (in points)
    Column 3 is the 2000-2011 score home field advantage for the team (in points)
    Column 4 is the differece.

    In theory, New Orleans GPF (0.0) plus the column 1 number for New Orleans (2.12) minus the Tampa Bay GPF (-1.1) should give the Vegas spread:
    So we'd expect to see the line at Tampa Bay +3.23
    I see Tampa +3.5 posted, which matches that prediction.

    In theory we can add the home team's column 3 number and subtract the away team's column 3 number to get a 'corrected' line:

    That said, when I looked at my historical data, that kind of calculation did not look significantly predictive.

  7. Ah, got it. So did you "bet" TB +3 (and your other bets) due to the difference between your Vegas and corrected lines, or did you use a different model?

    1. Originally, I only regressed the home field advantage (by points), and then I looked at against-the-spread history for the teams with the most extreme home field advantages. It looks like the spread tends to underestimate the extremity of those teams' home/away performance.

      (For example, the New York Giants are much stronger on the road or weaker at home than the 'average NFL team', and they when 66-39 against the spread on the road, but 43-56-3 at home over 2000-2011.
      Similarly, Seattle is 39-57-3 on the road, and 56-43-3 at home.)

      This led to the following betting strategy:
      Take the Vikings, Seahawks, Ravens, Rams and 49ers against the spread when they're at home, and their opponents when they're on the road.

      Take the Giants, Panthers, Bengals, Eagles, and Saints against the spread when they're on the road, and their opponents when they're at home.

      For normal games, when teams from one list meet teams from the other, don't bet. When teams from one list play each other, bet two units. (For neutral field games, it's reversed.)

      Much like Micheal with the turnover index, I'm looking to see how it performs through this season.

    2. Interesting. Do you happen to have the entire list of H/A ATS for all teams dating back to 2000?


      Though he's not that consistent with the stadium names.

  8. Interesting discussion. You may want to give ridge regression a try here. It's a fairly simple way to avoid overfitting. For example, I truly doubt that the Giants have a negative home field advantage or that the Vikings are worth an extra 6 points at home.

    Ridge regression "squeezes" your regression coefficients back together and it is controlled by a single tuning parameter (usually called lambda). You can use cross validation to find the best value of that tuning parameter (for example, train your model on the 2000-2010 seasons and then see which value of lambda best predicts the 2011 results, and then repeat for each season).

    After you've got your properly tuned model, you can then turn it loose on the 2012 season.

    1. Let's say that a team with home field advantage X and one with home field advantage Y play once at each home stadium, then we'd expect the net points to swing by (X+Y) since one team loses X points, and the other gains Y. So:
      Home/Away Swing=(HFA_X+HFA_Y)
      This rearranges to:

      Now, the average NFL team has a home field advantage of 2.6 points or so, so we can approximate the home field advantage of a team by:

      The Giants net +1.4 points at home, and +0.8 on the road. That's a swing of +0.6 points. Therefore their approximate HFA is 0.6-2.6=-2.

      The Packers net +7.5 points at home, and +2.0 on the road. That's a swing of +5.5 points. Therefore their approximate HFA is 5.5-2.6=2.9.

      The Vikings net +4.2 points at home and -4.7 on the road. That's a swing of +8.9 points. Therefore their approximate HFA is 8.9-2.6=6.

  9. Got some of that extra-tasty mean reversion tonight 3-7... (Well, the Pats are showing signs of life, but 21 points is a long way to go.) I'll be reading up on ridge regression.