In a big win for followers of the picks here, last week the Redskins managed to hold the Buccaneers to only three points, despite giving up 501 total yards. Winning the turnover margin 4-0 and the Bucs missing two field goals certainly helped. The Redskins have been winning close games all season backed by winning the turnover battle, with a very solid +11 turnover margin in 9 games.
A well-known idea in football analytics is that turnovers are often luck-based events, with a low degree of repeatability. While turnover margin is extremely correlated with winning in any one game, teams with a strong turnover margin in past games are unlikely to be able to carry that forward, and are likely to suffer a decline in performance. This leads to a possible angle of fading teams in-season who have generated strong turnover margins up to that point in time.
To start out, let's take a brief look at how much of an impact turnovers actually have on the game in which they occurred. We do this using a regression model, linking yards per play and turnover margin in any one game to the final score:
For each one more turnover the home team commits over the away team, we can expect the final score to favor the away team by between 4.1 and 4.4 points. Not of all of this is likely to be causation as teams who fall behind tend to throw more interceptions, but most of it probably is; fumbles lost does not correlate with whether a team is ahead or behind at any point in the game and their impact is very similar. Another item of interest, that doesn't really have anything to do with turnovers, is that even after accounting for turnovers and yards per play, the home team still has an advantage of 1.22 points. This is mostly due to penalty margin, which is not included in the above yards per play statistic.
We now move on to the angle. To start, let's take a look at whether past turnover margin predicts future success. We use the same simple model we used in the previous post, using past scoring margin and quarterback yards per attempt, only this time with past turnover margin per game added. To ensure our past turnover margin statistic is reasonable, we use season-to-date turnover margin, plus an additional five games of the mean value of 0 turnover margin. So, for example, if a team has a +6 turnover margin in five games, their turnover margin for the purposes of this model would be (6 + 0)/ (5 + 5) = 0.6 turnovers per game. Using this statistic we arrive at the following model:
We see that past turnover margin has no statistically significant impact on the model. This is actually notable, because it is not all that negative. We know that past turnovers are worth about 4.2 points, so if past turnover margin is complete luck, we'd expect a negative coefficient above, as we are including that 4.2 point impact in our past scoring margin statistic. But it turns out that there is some repeatable skill in forcing and preventing turnovers, and the level of value they have is roughly equal to past scoring margin. Running a simple regression of the difference in past team turnover margin to the score in the following game, we find that a 1-turnover difference in past games leads to a predicted difference of 3.4 points per game. While we have regressed past turnover margin to the mean, even un-doing this regression we would still arrive at a significant fraction of the actual value from past games of 4.2 points per game.
This would suggest that it is unlikely that turnovers are over-valued. Public bettors, who tend to base their plays on past scores, will not be making any big mistakes by ignoring past turnover margin, since it does not make much of a difference. Still, it is worth looking at whether the market over or under values past turnovers. We do this by looking at one last model, predicting final scoring margin based on a combination of the market point spread and the difference in past turnover margin of the two teams, and arrive at the following:
We see that after including the market's point spread, past turnover margin adds absolutely no predictive value at all. In other words, the market is weighing past turnover margin correctly. If bettors were improperly valuing past turnover margin, we'd see it here.
While this would suggest turnover margin is somewhat hopeless as an angle we can still take a look at how a strategy would have performed. Looking at the difference in team turnover margin in games since 2009 after making our regression to the mean, a little less than 16% of games have featured a difference (home-away or away-home) in team turnover margin of more than 0.5 turnovers per game. These represent the "strongest" picks in our sample. Betting this angle (fading the team with the better past turnover margin) would have lost 6.1 units over 422 plays.
While an appealing idea from an analytical perspective, the turnover margin angle just doesn't work, as the market prices this factor well, and the fact that there is some degree of repeatable skill in forcing turnovers means that the public does not price it that badly either.