Projecting Running Backs using Age Adjusted Correlation Curves and Historical Comps
tl;dr Fantasy players who draft based on efficiency will be consistent league losers. A player’s per game production in year Y correlates to Y+1 production to varying degrees depending on the player’s age. During a player’s prime years, which are longer than many might expect, the correlation is exceptionally strong. Combining these production coefficients with historical comps can provide for better pre-season projections for NFL Running Backs.
There are a couple universal truths in fantasy football, and they are directly related.
The first is that if you can accurately project volume (and its sexy and mysterious friend opportunity), you will probably win your league. Nothing correlates to fantasy points like volume of touches.
The second is that efficiency metrics are complete garbage when the goal is to try and predict what will happen in the future. Yards per carry (YPC) is famously bad, but all per-touch based metrics, including advanced metrics like Net Expected Points (NEP) and Expected Points Added (EPA), suffer from the same fundamental problem: they are noisy, descriptive only of what happened in the past, and not at all predictive.
So to recap, because per touch efficiency is worthless for projecting future performance, volume dominates.
An interesting question to ask is: why. Why is YPC such a predictive dumpster fire?
When I began studying which stats were predictive of y+1 performance I was surprised to see that efficiency metrics had such low correlations. At first I just thought football players were inherently unpredictable.
I don’t know of anyone who has been able to really explain why efficiency metrics are so noisy and vary so widely year to year. It’s a bit of a football mystery. I don’t propose to unravel the mystery here, but I do have some hunches.
My personal opinion is that efficiency metrics are analogous to BABIP in baseball. I should note that this isn’t a highly original opinion. Still, I haven’t seen a full explanation for why BABIP and football efficiency might be related. So here’s my attempt.
For anyone unfamiliar, BABIP stands for Batting Average on Balls In Play. It’s a metric that was created by Voros McCracken, who famously concluded that once a batter makes contact with a ball that lands in the field of play, the pitcher has no discernible affect on the outcome. In other words, pitchers are only responsible for strikeouts, walks and home runs, and hits on balls put in play tend to all regress inexorably toward a league average. If a pitcher has a high BABIP, its likely due to short term bad luck, and the probability is high that in the future his luck will even out.
Now if BABIP in baseball is truly analogous to YPC and other efficiency metrics in football, we would have to conclude that the running back has no influence over the outcome of any particular play. This is a very strong claim.
Nearly all fans agree that football is the ultimate team sport, and that assigning blame and parceling out credit for any successful play is extremely tricky. Saying that the offensive line contributed to a running back’s production is one thing; saying that on any given play a running back has no discernible effect on the outcome is pretty much bananas. It flies in the face of what our eyes tell us happens every Sunday.
Still, I think it is probably pretty accurate. Consider:
Prior to the running back taking a hand off, there are a string of actions that must be successfully executed before a back can even begin to bring his skills to bear.
- the QB/RB exchange must be performed correctly and on time
- the mesh point needs to occur at the correct position in the backfield
- the o-linemen/te/fb/qb (if read option) need to execute their blocking scheme or fake on time and successfully
- finally the defensive scheme needs to be fooled such that unblocked defenders are not positioned near the line of scrimmage on the play side of the field, allowing them to make a play in the backfield
But it gets even worse than all that for a running back. Even when all the above necessary conditions are met, there is a certain amount of yardage that we can expect any NFL caliber athlete to produce when a crease is available for him to run through. Football outsiders has put this number somewhere between 3-4 yards, which also happens to be right around the long run league average for yards per rushing attempt.
Finally tactical coaching decisions and gameflow can have a tremendous impact on yardage gained on certain running plays (i.e. attempting to run out the clock with a lead against a 8 man front).
Until we get better data, perhaps from Zebra, we will remain in the dark regarding running back efficiency. Until we have tools that allow us to measure and quantify the push of an offensive lineman, until we have enough player movement data to use neural networks to automatically classify run plays into outside zone/power/trap etc. like MLB did with PitchFX, until we can automatically track the position and type of defenders on the field and quantify their distance from the point of attack and its effect on the success of a running play - basically until we know much more about each and every football play - the most rational course of action is to just assume that RB talent adds very little value.
Because on a down to down basis running backs add very little to the offense by virtue of their own inherent talent, RBs derive nearly all their value from volume. Talent absolutely matters, do not misunderstand me. Talented backs will get more volume, but only rarely are talented backs more efficient over time than their peers. As a rule, a high scoring running back who puts up good fantasy points on limited touches will see his scoring drop heavily the following year unless he sees an increase in volume.
Jamaal Charles seems to be the lone exception that proves the rule. He is otherworldly and a true outlier.
In terms of fantasy football, owners who draft players based mainly on last year’s efficiency - no matter how fancily dressed in advanced metrics that efficiency might be - will lose to owners who successfully draft based on volume and opportunity. Volume utterly dominates. This is a big reason why guys who have never opened a spreadsheet can do well year after year in fantasy football. Simply following the news can win you a league.
You said there would be stats
So if per touch production isn’t the correct context to assess the worth of a Running Back for fantasy, what is?
It turns out that per game averages correlate better than raw totals, so the correct context is at the game level.
The reason for the increase we see in predictive power in game level stats vs. aggregate totals isn’t completely clear. It probably has something to do with controlling slightly for injuries. It may also be the case that over the course of a full game the marginal value added by running back talent is captured in a way that touch level metrics miss.
Whatever the case, many analysts have calculated Att/G and Yards/G year-over-year correlations for running backs. Most recently TJ Hernandez of 4 for 4 published some of his findings.
TJ looked at all running backs who played on the same team in consecutive years, with over 100 carries in each year. The correlations when the analysis is performed at this level are pretty good: around .65 for Att/G and Yards/G.
Much stronger correlations can be found if you adjust for age and eliminate the carries threshold. Below is a table showing the year over year correlations for a number of fantasy relevant variables, based on age, for all running backs who played in at least one NFL game in consecutive seasons.
You can view and download the dataset upon which this is based here. It is pretty obvious that YPC is worthless. Still, where TJ and others have found year over correlations of .11 for YPC, breaking players out by age adds some slight predictive power to the stat.
Other items of interest to note:
- Att/G and it’s twin Yards/G slowly rise as the Running Back cohorts age, and then level off…for a very long time, and at a very high correlation coefficient
- running backs do not drop off a cliff when they turn 30, but very few stay in the league past that age
- the ones that do stay in the league will continue to be consistent until age 33 (at which point sample sizes get too small to predict anything with confidence)
- TD/G adjusted for age is actually relatively predictable
- Receptions and rec yards per game are both decently projectable from the previous year’s stats
The biggest issue with studies of this type is survivorship bias. But in the context of fantasy football, I think this is a feature not a bug. If it is known a player is going to get opportunity and volume, the fact that he is entering his year 31 season shouldn’t scare you in the slightest. Merely being in the league at 31 and still being projected for volume means that, absent injury, a player is very likely to perform to the level of his previous year.
It’s also important to note that high correlations to a previous year don’t necessarily mean that the next season will be good. It just means it will be similar to the previous season. But as you can see from the averages graphed below, it is a fairly gentle ride into the sunset for aging NFL RBs.
Rich Hribar, among others, has done great research on age curves for RBs. One of the most famous visualizations on player aging is the Top 36 RB seasons by age. The shape of that curve is familiar to almost all fantasy players at this point. It looks like this:
But if we look inside this graph, we can see support for the findings I describe above.
Good running backs come into the league seemingly in clusters. As the good backs age, the peak that sits centered around age 26 in the Hribar graph moves with them. It’s like a wave, building and crashing on the rocks of injury and old age. The most important feature of these graphs though is the number of top36 seasons produced by runners older than 26. In some years the linear trend is neutral or even positively - instead of negatively - correlated to age! It’s an unintuitive result given the shape of the aggregate graph.
In many ways, and through no fault of the author, the Hribar graph might be responsible for more bad decision making by fantasy football owners than any other analysis ever conducted. It has scared an entire generation of dynasty owners away from proven veteran talent at the RB position simply because a player happens to be older than 26 and therefore “past their peak.”
Putting it all together
So we have some very strong year over year correlations for running back stats that are fantasy relevant. How can we turn these into a usable projection?
In truth, when correlations are .8 (or higher), what you choose to “regress” the remaining 20% of a particular player’s performance toward will have a fairly limited impact on the final projection.
Still, if we want to be as accurate as possible, we can use a modified version of a technique first put forward by Jonathan Bales in Fantasy Football for Smart People. In it he described a process where you take a weighted average of a player’s previous year performance and mix it with the average performance of the top X number of players at his position.
Here is an example for Matt Ryan from Bale’s book:
So let’s project a player from last season. We’ll take Matt Ryan, who threw for 4,177 yards, 29 touchdowns and 12 interceptions, and ran for 84 yards and two scores in 2011.
Let’s pretend we were assessing him in a one-quarterback league. To project Matty Ice’s passing yards, we add 50 percent of 4,177 (2,088.5) and 50 percent of the league average in a one-quarterback league (1,805) for a total of about 3,894 yards…Thus, our initial 2012 projection for Matt Ryan would have been 3,894 passing yards, 24 touchdowns, 9.3 interceptions, 94 rushing yards, and two rushing touchdowns.
Bales, Jonathan (2014-06-01). Fantasy Football for Smart People: How to Dominate Your Draft (p. 68)
What Bales did was in essence run a regression, only he was able to specify exactly which player population to regress on to.
What we will do is rather than taking an average of the previous year’s performers, instead we’ll take an average of 25 of the N+1 seasons for players who we identify are objectively similar to the player we are trying to project.
Fantasy Douche at Rotoviz pioneered the idea of Similarity scores based on age, weight, height, and per game production metrics. I borrowed this idea to create a highly relevant population mean to regress to, and also as a check on the projection itself. A completed projection for Todd Gurley is shown below.
The scoring is custom for one of my dynasty leagues, so pay more attention to the per game averages rather than the projected points.
Todd Gurley’s N+1 comps are on average a full 2 years older than he will be for his second season. So it makes sense that he would have a better projection than that group (80.7 yards/G vs. 71 yards/G). He is a phenom.
Next up is Matt Forte who will be entering his age 31 season next year. So long as we believe Forte will have a similar amount of touches and opportunity as his previous year, it makes sense that his yards per game would be quite close to 2015. The 25 comp average gives Forte a projection a full 17 total yards/G lower, so that should probably temper our expectations somewhat.
A cool feature that falls out of using similarity scores is that we can quantify how many players miss the following year. I called it a bust rate, which is good enough as far as names go, but it’s really more of a injured/washed up/demoted measurement. For Forte, the bust rate is 8%, which isn’t a terribly scary number.
The major drawback of any projection system based mainly on the previous year’s stats is that they do not account for opportunity. And since opportunity drives volume, and volume wins championships, we have to dig in and do some team level projections and then make some educated guesses about playing time. I’ll cover that in a future post.