Regression methods in the NFL
I don’t know how familiar you are with statistical methods, so I am going to simplify the term regression. Regression just means that over- or underperforming numbers tend to regret to the mean – or their average – over time, specifically from week to week or from season to season. By understanding regression, you can find excellent value week in week out in the NFL. The reason is that the public bettor overreacts to over- and underperformances and struggles to adjust a team’s performance to the average situation. That’s also the reaon why the schedule plays such an important role. The results are inflated lines due to public perception where the sharp bettor can take advantage. You would get some extra points that helps you winning more bets over time and increasing your ROI.
Regression numbers from season to season are such as “close wins”, “turnover differential” or the discrepancy between “estimated wins“, “pythagorean wins” and actual wins by any team (I recommend reading the linked definitions). All these numbers have shown a regression to the mean over the years. There are some exceptions to the rule, but you will find regression in the majority of the cases. Let’s start with the turnover differential:
Example: The Raiders had a -15 turnover margin during the 2014 regular season which ranked dead-last. In 2014, they improved that number to +1 which ranked 17th in the league.
Example: The Texans had a -20 turnover margin in 2013 (#32), whereas they improved to +12 (#3) in 2014. A swing of +32.
Over the last 26 years, the 50 teams with turnover ratios between +15 and +25 posted an average margin of +2.7 the following season, declining by an average of 15.8 turnovers (per ESPN). That’s astonishing, isn’t it? Guess who is set for a heavy regression in 2016? The Carolina Panthers who posted a turnover ratio of +20 last season.
Over the last 26 years, teams that posted turnover margins between -25 and -15, produced an average turnover margin of -0.1 the following season, improving by 18.6 turnovers (per ESPN). They also won 3.1 more games the next season. A team that is set for a heavy improvement? The Dallas Cowboys, who posted -22 in 2015.
Why do turnover margins improve or decline off outlining seasons? Well, it’s pretty simple and I’ve got three arguments: First, teams tend to work on things that went bad in the past. They rarely work on things that went the right way. Second, personnel and coaching changes. It’s logical that Tony Romo won’t throw as many interceptions as Matt Cassell, Kellen Moore and Brandon Weeden did. Key injuries and the player personnel do effect the turnover margin. The third reason is coincidence. Some turnovers are completely random just like fumbles or tipped interceptions. You can laser a perfect ball to your receiver, but if he isn’t capable of catching and tips it into the hands of the cornerback, it’s not the QB’s fault. That coincidence evens out in the long run. Sometimes you are on the bad side of the fumble recovery percentage, sometimes on the right one.
Another category is close wins. That means a win by one score. Generally, a high percentage of close wins cannot be repeated the following season, and vice versa. Per ESPN, eight teams won 75 percent of their one-score games in 2014, finishing a combined 35-7-1 in such contests. Those teams regressed to 34-25 in close games last season. A fumble can make the difference between a close win or a close loss and overshadow a good performance by a certain team. An over-performance in close wins is very tough to repeat the following season, only a handful of teams was able to do so in the past.
The third category is the difference between projected and actual wins. The pythagorean win calculation is based on the scoring difference over a season. Per ESPN, since 1989, 62 teams have fallen in the range of underperforming by 1.5 to 2.5 wins. Forty-nine of those teams (79 percent) won more games the following year, and the group as a whole improved by an average of 2.3 wins. But I always felt that it isn’t enough. It’s just that a few blowout wins against weak teams can inflate the pythagorean number. That’s why I highly recommend to use both, the pythagorean and the estimated win numbers from Football Outsiders. If both show a significant difference, you can assume that a team has really over- or underperformed. These teams tend to improve or decline the following season. Some examples from the past:
2014 Atlanta Falcons (4-12) underperformed by 2.5 estimated wins and 1.9 pythagorean wins. They won 8 games in 2015.
2013 Atlanta Falcons (13-3) overperformed by 3.9 estimated wins and 1.8 pythagorean wins. They won 4 games in 2014.
2013 Detroit Lions (4-12) underperformed by 3.6 estimated wins and 2.4 pythagorean wins. They won 7 games in 2014.
2010 Baltimore Ravens (9-7) underperformed by 3 estimated wins and 2.6 pythagorean wins. They won 12 games in 2011.
There are also exceptions to the regression, for example the 2013 Colts (11-5) who overperformed by 4.8 estimated wins and 3.8 pythagorean wins. They won 11 games again in 2014. The 2014 Colts are a favorable example of why you always have to put every aspect in context (personnel, schedule, dyanamics) and not rely solely on mathematical models. Andrew Luck had the advantage of playing in the worst division several years in a row. The Colts had a 5-1 head start within their division because their opponents have been so bad. From 2012-2014 the Colts went 16-2 within their division and 20-16 against all other teams.
There is a strong correlation between close wins and projected wins differential. A few lucky bounces lead to a better turnover differential, which leads to a close win, which leads to overperforming on the season. But a turnover differential doesn’t have a true correlation to close wins and projected win differential. All in all, you will find out that in almost every case, a team that has over- or underperformed by 1.9 or more games, always has had an advantage in close games.
Teams that are set for a regression in 2016
Carolina overperformed by 3.9 estimated wins and 2.6 pythagorean wins. They had a record of 7-1 in close games and a turnover differential of +20. They had the easiest schedule in 2015.
San Diego underperformed by 2.0 estimated wins and 1.9 pythagorean wins. They had a record of 3-9 in close games and a turnover differential of -4. They are going to have an easy schedule.
Baltimore underperformed by 2.5 estimated wins and 1.0 pythagorean wins. They had a record of 5-9 in close ames and a turnover differential of -14. They are going to have a much better schedule than last year – atleast on paper.
Denver overperformed by 1.3 estimated wins and 2.3 pythagoran wins. They had a record of 9-3 in close games but a turnover differential of -4.
Giants underperformed by 1.4 estimated wins and 1.5 pythagorean wins. They had a record of 3-8 in close games but a turnover differential of +8.
Tennessee underperformed by 1.4 estimated wins and 1.8 pythagorean wins. They had a record of 2-6 in close games and a turnover differential of -14.
Indianapolis overperformed by 2.5 estimated wins and 2 pythagorean wins. They had a record of 7-4 in close games and a turnover differential of -5.
Seattle underperformed by 2.5 estimated wins and 1.8 pythagorean wins. They had a record of 2-5 in close games but a turnover differential of +8. The Seahawks finished with the best total DVOA number, by a wide margin!
When also putting all the other important aspects into context, I see a regression for all of these teams, except for the Colts. The Colts are an exception to this. Their 2015 season wasn’t a normal underperformance. More details will follow in an write-up for the Colts soon.