Thursday, December 27, 2012

Turnover Index - Final 2012 Results

Here are the final results for the Turnover Index for the 2012 season.  In my original post on the topic, I excluded week 17 games from the analysis, so there are no more betting opportunities this season.

Week 16 Results

Things ended on a low note, with the week 16 picks going 1-3 against the spread.

Turnover Index through Week 16 (Against the Spread): 18-16-1

That's 52.9% against the spread.  Assuming the standard sportsbook bet (lay $110 for a chance at a $100 profit), you would have realized a whopping 1% return on investment.  That's below the long term performance of 59% against the spread observed from seasons 1998-2011, but at least it's positive.

So what, if anything, went wrong?

  • Random Variation - Assuming each bet truly had a 59% chance of success, the probability of less than 19 successes out of 35 trials is 23%, or about 1 in 4.  Looking back over past seasons' results against the spread, 2000 was at 41.1%, 2002 was at 38.9%, and 2004 was at 52.9%, identical to this year's results.  So it's not as if 2012 was an unprecedented statistical anomaly.
  • Total Turnovers vs. Turnovers Per Game - The betting criterion I came up with was to bet on any team with less than 10 defensive turnovers than its opponent.  It may have been best to convert that criterion into a "per game" difference.  Bets in the latter half of the season didn't perform well.  The more games a team plays the wider the spread in aggregate turnovers, which may have led me to cast too wide a net.
  • The Betting Market has Caught On - The biggest problem with a betting approach like this is that it only works to the extent that the betting market remains inefficient.  This is also what plagues stock market trading strategies based on technical analysis.  And while I don't think the market has changed its evaluation of turnover performance, it is a possibility.

Next Season

I plan on continuing this feature next NFL season with the only potential tweak being redefining the betting criterion in terms of turnovers per game, rather than accumulated season-to-date turnovers.   When faced with lackluster performance in a prediction model, there is a strong temptation to dive deeper into the numbers, searching for increasingly complicated patterns in the data that have led to higher returns (Should I ignore fumbles? Should I weight recent games more heavily? Is the pattern different for home teams vs. away teams?  Are 4th quarter turnovers less "random" than turnovers early in the game?).

And while it is possible that by digging deeper, you find a betting strategy more likely to result in consistent profits, you are just as likely (and probably more likely) to be searching for green jelly beans.

I am open to suggestions for other potential betting market efficiencies to exploit and track.  In the comments to these posts, Nate has laid out a fairly promising approach involving home field advantage that may be worth a separate investigation. Another commenter suggested looking at special teams touchdowns.

Steven Levitt (of Freakonomics fame) published a paper in 2004 which demonstrated that home underdogs tend to cover the spread more often than not.  In this 2009 post on the Freakonomics blog, Levitt acknowledged that this market inefficiency had disappeared in recent years (at the time 2007 and 2008).  In a separate post, I plan on taking a look at the 2009-2012 results to see if the pattern is gone for good, or whether 2007 and 2008 were just bad years for the home underdog strategy.

Are there other sports where the betting market and/or the general public tends to over or under value certain aspects of the game?  In basketball, 3 point field goal percentage, perhaps?


  1. Schemes with more bets per year and a more predictable betting pattern are more efficiently exploitable, and less vulnerable to variance.

    Moreover, if the intent is to profit, then having the bankroll sit idle in the off-season is less than ideal. Based on that, I'd be looking for something exploitable in the MLB, MLS or IndyCar to complement the NFL, NBA, or NHL season.

    I would expect that a for-profit betting strategy is going to be fractional bankroll, rather than fixed bet. Under a scheme like that, I would expect a record of 18-16-1 to incur a loss.

    The HFA error theory seems to break down with teams like Oakland and Pittsburgh whose performance against the spread seems to clash with the implied home field advantage error. Splitting historical performance against the spread into home and away does suggest some other teams to look into.

  2. My two cents is that a set of 35 games will tell you almost no information. I.e., if the performance had been 21-13-1, I would be no more bullish on the 2013 projected performance than I am with 18-16-1.

    That said, I'm curious how the window of 8-12 (TO diff.) performed this year? Those endpoints aren't terribly optimized, as the historical perf. seems to tail off rather smoothly on both ends.

    The first thing I might take a look at is to see how perf. has varied by seasonal quarter -- that is, games 1-4, 5-8, 9-12, 13-16.

    Other stats that might be interesting to research:

    Abnormal third down conversion rates achieved or allowed (normalized for distance and team's overall ability).

    Abnormally high or low aggregate turnover impact. That is, measure the impact in terms of expected points gained/lost for each turnover, and look at ATS perf. of the most favorably/adversely impacted teams year to date. Both gross and on a per-turnover basis.

    Abnormally high/low variance in ypp (compare to perf. of "normal" variance teams).

    Abnormal rate of fumbles per QB hit on both def. and off.

    Pass interference penalty yardage differentials.

  3. One other random idea for the turnover effect: try eliminating teams with truly elite offenses from the study. Say, top 3 in the league each year. Theory would be that those few teams will cause so many adverse situations for the other teams offense (long passing) that the elite-offense team will have a def. TO rate that is high and fairly sustainable (not due to luck). Thus, they may not be great teams to bet against. Obvi, you'd have to measure the top 3 at the point in time to avoid any look ahead bias.

  4. NBA/CBB:

    Look at avg. time to first shot taken in offensive possessions. Wondering if teams that are very quick to shoot will go over the total more often than the market expects (theory is that rate of shooting is more predictable than success of shooting).


    Look at defensive strand rates (runners left on). First calc. the expected strand rate for a given team's K/BB ratio, and look for abnormal deviation at ~mid-season. Teams with high strand rates should perform more poorly than expected.

    1. On your MLB suggestion, is the rationale that a high/low strand rate is "luck" and not a repeatable skill?

    2. Yes, it is mostly random variance (sequencing of hits/outs) for a given pitcher quality. I don't know how much of this is already factored into market odds properly.

  5. Statistics suggest the following:
    Baltimore, Pittsburgh, San Diego and Seattle are strong at home.
    Carolina, Green Bay, New England, New Orleans, Philadelphia, Miami, and the New York Giants are strong on the road.
    Cincinnati, Denver, Miami, the New York Giants, Oakland, Saint Louis, and Washington are weak at home.
    Arizona, Minnesota, Tampa Bay, Saint Louis, Seattle and San Francisco are weak on the road.

    Pick teams that are strong, except against other strong teams.
    Pick against teams that are weak, except against other teams that are weak.

    Since the Rams are considered weak at home and on the road, I want to call this the 'Beat the Lambs' method.

    1. Applying that method to 2012 gives a below-parity result. I'm thinking I should study the feedback mechanisms in the spread a little more...

    2. There's some interesting stuff on home field advantage over at football perspective that you might find interesting.

    3. Interesting reading. Football Outsiders (linked in the article) also mentioned that people are wrong about New Orleans' home field advantage.

      The theory about cold weather and travel is nice, but Oakland which is a bit warmer, but right next to San Francisco with similar travel has a much smaller home field advantage. Similarly, the Giants and Jets play out of the same stadium had have very different HFAs.

      The article didn't mention that teams with artificial surface home stadia (like Seattle) also have bigger home field advantage.

  6. There might be some advantage opportunities on the over/under:
    STL, DAL, WAS, MIA, SF, ATL, CLE, JAC, PIT, TEN, KC, or TB at CLE, WAS, BUF, IND, CIN, or STL should be under.
    NO, SD, DEN, NE, CAR, BUF, NYJ, DET, or NYG at ARI, GB, DET, NYG, PHI, DEN, NO, or DAL should be over.

    That would have gone 19-12-0 this season so far. Though going 0-3 in week 2 on the first 3 bets would be pretty rough.

  7. Something that seems conceptually strong is finding profitable split middles:

    For example, in Week 1 of 2012, the Vikings at home were favored by 3.5 over the Jaguars. The win/lose line was -$195 +$170.

    So let's say that we take the Vikings to win for $409.50, and the Jaguars to cover for $324.5. 409.5 * 295/195= 324.5 * 210/110 = 619.5 so we're really only wagering $114.5 And we win $505 if the Vikings win by 3 or less. In my database, this has happened 1 in 7 times -- way less than the 1 in 4 we'd like to see for this to be profitable - it loses around 50%.

    Baltimore hosting Cincinnati was a 7 point favorite, and -350 to win. Taking Baltimore to win for 735, and Cincinnati to cover for 495 is guaranteed to net at least 945, so it's a net wager of 285 to win 660 - we need it to happen at least once in 2.3 times to profit, but it happens one time in 3.9 - this bet loses around 30%

    Philadelphia at Cleveland is favored by 9 points. If we take 550 for Cleveland to Cover, and 945 for Philly to win, then it's a net wager of 340 to win 1155. 9 point favorites have won by 9 or less 39 of 126 times (11 of 27 on the road, 28 of 99 at home) and home ones. This means the wager has an expected return of around 5%.

    Things get a little sparse beyond that, but I expect that as the spread gets larger, the expected return improves. It looks like roughly 1 in 6 games will have spread of 9 or more.

    1. I ran this for 2012 through week 16 - I got 27 bets, winning 13.5 (on the Saint Louis/San Francisco tie) and netting 8.8 units (33%) on a flat wager strategy.

    2. I don't understand how it is "conceptually strong" to simply bet on large money line favorites and against the same teams ATS. You're looking at a very small sample of past results (e.g., something that has happened 39 times) and assuming that it will be 100% predictive.

    3. The reason it was interesting is that there's a potential for exploitable inefficiencies if the spread implied by the winning odds doesn't match the spread.

      That said, you're right. I had the payout calculation wrong, and this would be a losing bet, and while it would have won over this season, it would have lost over the last 10.

  8. To better factor in teams' individual HFAs, we'll have to evaluate whether they are affected by a team's fortunes in a particular season. It's been largely proved that the HFA is influenced by referees which in turn are influenced by the fan presence, so it's plausible that a team having an unexpectedly good season will enjoy a bigger HFA than usual (fans of the Pats could be too used to winning, hence the team's mediocre HFA).

    In addition to turnovers, special team play seems capable of inflating/deflating a team's apparent performance. You can see how many expected points every team gets from ST at (go to any team's home page to see game-by-game EP splits). I calculated teams' game-to-game correlation for ST EPs and it came to 0.0 (comparing to around 0.2 for passing EPs).

    So theoretically we could start to determine how much a team is likely to be overrated or underrated by combining their turnover numbers (perhaps converting to 4.5 points per TO) and special team EPs.

  9. Oh, P-F-R also have teams' game by game EPs off turnovers, both offense and defense.

  10. Looks like the 2013 schedule has been announced...

    Week 1:
    Ravens at Brocos (pick Broncos)
    Sun 1pm
    Falcons at Saints (pick Falcons)
    Chiefs at Jaguars (pick Chiefs)
    Bengals at Bears (pick Bengals)
    Vikings at Lions (pick Vikings)
    Dolphins at Browns (pick Over)
    Sun 4:25pm
    Packers at 49ers (pick 49ers)
    Cardinals at Rams (pick Rams)
    Sun 8:30pm
    Giants at Cowboys (pick Cowboys)
    Giants at Cowboys (pick Over)
    Monday 7pm.
    Eagles at Redskins (pick Eagles)

    1. A couple of those are wrong. It should be:
      Lions over Vikings,
      Giants over Cowboys,
      Panthers over Seahawks was completely missed.

  11. Looking forward to the first update of this season. Thanks!

    1. I plan to have a write up this weekend. I can let you know that there are no betting opportunities for week 5. In general, teams are getting matched up against opponents with similar defensive turnover rates.