Turnovers are one of the most telling statistics in football but turnovers are also highly deceptive. An average team with a turnover margin of +10 would likely make the playoffs with a 10-6 record while an average team with a -10 turnover margin over their 16 game schedule would likely finish with a 6-10 mark. That statement is based on the regression equation that predicts wins as a function of turnover margin and which has a slope of 0.21 (i.e. every turnover is worth 0.21 wins). Some argue that turnovers are a function of skill, which is true to a degree, but the correlation to turnover margin based on the previous season’s turnover margin is very low (an R-squared of 0.01) and the slope of the regression equation is just 0.11. That means that only 11% of a team’s turnover margin one season is explained by the previous season’s turnover margin. So, a team that is +20 in turnover margin one season would be projected to be just +2.2 in turnover margin the next season.
There are exceptions, of course, as New England, with Tom Brady at quarterback, is likely to throw considerably fewer interceptions than an average team year after year after year while a team with a perennially bad quarterback is likely to throw more interceptions than average. However, for most teams, interceptions are highly variable and fumbles lost and recovered are even more variable from season to season.
What we’ve done is project interceptions thrown, defensive interceptions, offensive fumbles and defensive fumbles as a function of the statistics that have proven predictive. From those equations, we can project what a team’s turnover margin should have been in any given season. For interceptions, the most predictive statistics are passes defended, which is passes broken up plus interceptions, and completion percentage. The most significant statistic to project fumbles is times sacked on offense and sacks on defense – although the correlation is still very low. We’ve also discovered that the percentage of fumbles recovered by the defense depends on the type of fumble. For instance, fumbles after completed passes are recovered by the defense more often than fumbles on running plays because there are generally more defensive players than offensive players in the vicinity of a fumble after a completed pass. So, for fumbles we used algorithms to predict the number of each type of fumble (on a run, on a completed pass, on a sack, punt and kickoff returns, and a fumbled snap or handoff) and then used the recovery percentages for each type of fumble to project fumbles lost by the offense and fumbles recovered by the defense. While there are some statistics that correlate to fumbles the level of correlation is still very low and fumbles lost and recovered are mostly random. In fact, 65% of the projections for fumble margin are between +1.0 and -1.0 for an entire season with the extreme projections being at plus 3 and minus 3. Anything out of that range is certainly random.
Using predictive statistics to assess turnover variance is more accurate than just assuming every team will turn the ball over at the same rate. For instance, New England was projected to throw just 6.4 interceptions last season, which is the lowest projection for offensive interceptions of any team in our 17-year study. The Patriots’ 2 team interceptions in 2016 were also the lowest of any team, with the next lowest being the 5 interceptions that the Pats threw in 2010. So, while Tom Brady is still the most likely quarterback to lead the league in fewest interceptions, last year’s 2 picks were still randomly low and resulted in 24 points of offensive interception variance.
Of course, not all turnovers are created equal. A fumble at 1st and goal that is returned 99 yards for a touchdown is far more costly to the fumbling team than throwing an interception on a bomb 50 yards down the field with no return on 3rd and long, which is really synonymous with a 50-yard punt. Rather than simply comparing projected turnover margin with actual turnover margin to determine turnover luck/variance, we use expected points added, or EPA. EPA is used to assess the value of each play of an NFL game and is a far better indicator than yards or yards per play. For instance, a 10 yard run on 3rd down and 9 yards to go adds more expected points to the offensive team than a 10 yard run on 3rd down and 15 yards to go.
The average value of a turnover in the NFL is 5.3 points (5.2 per interception and 5.5 per fumble recovered) based on our EPA analysis and there is variance in points per turnover as well. For instance, Cleveland threw 14 interceptions on offense last season, which is just over 1 higher than the league average, yet those 14 interceptions only cost the Browns 48.7 points, which is considerably better than the league average of -66.0 points on interceptions. Denver, meanwhile, gained 89.2 points on their 14 defensive interceptions, which is nearly 18 points more than would be expected from that number of interceptions.
The tables below shows the 2016 offensive and defensive interceptions and fumbles for each team and the statistical projections for interceptions and fumbles, along with the variance (i.e. luck) in terms of points, which is the difference between a team’s actual turnover EPA differential (defensive turnover EPA minus offensive turnover EPA) and their projected turnover EPA differential. We assume that each team should have the same EPA per interception and fumble as the league average, which is an assumption based on the lack of correlation in a team’s EPA per turnover from one season to the next.
Because turnover margin one season has a low correlation to turnover margin the next season there is a significant correlation between win differential (wins minus previous season wins) and the previous season’s turnover margin. Thus, teams with positive turnover luck the previous season are likely to have fewer wins while teams that suffered from negative turnover variance are more likely to improve. In fact, projected win differential is equal to -0.16 times the previous season’s turnover margin or is equal to -0.024 times the previous season’s turnover luck in terms of points.
You can see from the turnover tables that the Chicago Bears were the NFL’s unluckiest team in terms of turnover variance, as turnovers cost the Bears 105 points more than projected. Based on the regression equation projecting win differential as a function of turnover luck the Bears would be expected to improve their win total by 2.5 wins based solely on having neutral luck in turnovers in 2017. The improvement could be even more than that if history is any indication. Chicago is just the 7th team to experience negative 100 points or more in turnover luck and the other 6 such teams had an average improvement of 6.3 wins with 5 of the 6 improving by at least 6 wins after their unlucky season.
The teams with the most positive turnover luck in 2016 were Minnesota (+61.9 points), Oakland (+54.4 points) and Philadelphia (+50.9 points) and the Raiders, in particular, will find it a difficult challenge to reach last season’s 12 wins despite improving their overall talent. There have been 43 previous teams in our 17-year study that won 10 or more games with turnover luck of +50 points or more and only two of those teams increased their win total with 3 matching the previous season’s win total and 38 winning fewer games. The average win differential in those 43 cases was -3.1 wins so the Raiders could be a better team and still finish a couple of wins shy of last season’s 12-4 regular season record.
[wpdatatable id=10] [wpdatatable id=11] [wpdatatable id=13]
Our series of articles on luck will continue with special teams, 3rd down variance and Red-Zone variance in the next couple of weeks.
Last season was the first season using our advanced play-by-play NFL model and the results were incredibly good. The 2016 NFL Best Bets were a very profitable 100-69 (59.2%) against the spread, including an incredible 66-26 (71.7%) on Best Bet sides. The NFL Strong Opinions were also profitable at 64-52-4 (55.2%), so the model worked very well.
NFL Best Bets and Strong Opinions
Year Category Sides Totals All Plays Win Pct.
2016 Best Bets 66-26-0 34-43-0 100-69-0 59.2%
Strong Opinions 27-30-3 36-22-1 64-52-4* 55.2%
*Includes a Super Bowl prop bet Strong Opinion win.
There was some positive variance in the NFL Best Bet results last year (won more close games than expected) and the win percentage should have been 57.4% rather than 59.2%, but it still didn’t make up for the negative variance I had in my College Best Bets, which were a profitable 48-38 (including season win totals) despite the bad luck (just 12-23 on Best Bets decided by 7 points or fewer).
You can view my Past Performance page for more details.
For details on my subscription packages you can visit the Best Bets page.