Fantasy Football 2021: Running Back Age Decline
Among all fantasy football relevant positions, the running back age decline is both the swiftest and steepest. Not often – of course Frank Gore excluded – does an aging running back hang around for years after his prime while still contributing significant fantasy production. Rather, much more often what happens is that a running back – seemingly out of the blue – regresses to the point where they are no longer recognizable when compared with their old selves.
Even over the last few years, many formerly dominant running backs were drafted at a high ADP only to have below average seasons that signal the end of an exciting career. For example, in 2018 Todd Gurley was the top overall player in dynasty startup drafts which typically indicates an expectation of high performance for many more years. However, in 2019 he finished as a high end RB2, in 2020 around 30th among running backs, and as of writing this in mid July of 2021 is not even on an NFL roster anymore.
While Gurley’s case may be more dramatic than the typical running back age decline is, it is certainly not an anomaly. In this article we want to study how likely it is that a running back will begin his decline. As before in our article about wide receiver age decline, we will study the problem from both a descriptive and predictive perspective. Not only will we present statistics about how quickly running backs have declined at various ages in the past, but we make predictions using a logistic regression model for this upcoming fantasy football season about which running backs are the safest bets based on age and past production.
Let us begin with presenting our results about which running backs will be the safest in the upcoming 2021 fantasy football season.
Top 25 2021 Fantasy Running Backs by Confidence Level
In our previous article about wide receiver confidence levels, we used raw position rank from 2020 and player age to measure confidence in top 12 and top 24 finishes. We saw familiar names like Justin Jefferson, DK Metcalf, and AJ Brown near the top. However, many online fantasy communities have Brandon Aiyuk as their 2021 breakout star while he was nowhere to be seen in our model. This is because our model considered season-long rankings instead of per-game stats. Aiyuk was 15th in fantasy points per game – right between Keenan Allen and Tyler Lockett – as a rookie. Points per game, Aiyuk was only about one point worse than Jefferson.
In this article, we’ll slightly modify what we did for wide receivers and rank them by points per game instead of raw total points. This should be a much better predictor of the future. We’ll threshold those players who played less than 50% of the games. We also allowed our logistic regression model to treat aging differently in the 22-27 range than it treats aging after 27 in order to better represent reality. Then, the running back age + production confidence predictions are in the table below.
Name | 2020 Rank | Top 12% | Top 24% |
D. Cook | 1 | 64% | 85% |
A. Kamara | 2 | 62% | 84% |
N. Chubb | 4 | 59% | 82% |
D. Henry | 3 | 56% | 81% |
J. Taylor | 6 | 56% | 81% |
J. Robinson | 7 | 54% | 80% |
A. Jones | 5 | 53% | 79% |
D. Montgomery | 8 | 53% | 79% |
J. Jacobs | 9 | 51% | 78% |
A. Gibson | 12 | 46% | 74% |
C. Carson | 10 | 45% | 74% |
M. Sanders | 13 | 44% | 73% |
E. Elliot | 14 | 43% | 72% |
K. Hunt | 15 | 41% | 71% |
R. Jones | 16 | 39% | 69% |
A. Ekeler | 18 | 36% | 67% |
D. Swift | 19 | 35% | 66% |
K. Drake | 17 | 34% | 64% |
C. E.-H. | 21 | 32% | 63% |
JK Dobbins | 22 | 30% | 61% |
Da. Johnson | 11 | 28% | 56% |
N. Hines | 26 | 25% | 56% |
M. Gordon | 20 | 25% | 55% |
S. Michel | 27 | 24% | 54% |
M. Davis | 24 | 20% | 49% |
Obviously, near the end of this list our model begins to break down. Melvin Gordon and Sony Michel should not do as well as my model thinks simply because their roles in the offense will not be the same as they were last year. Moreover, the rookies are noticeably missing because we don’t have any data on them to work with. However, this list at least gives you a way to value guys at different parts of their career.
Running Back Age Basic Statistics
As before, we want to study the running back age decline problem by looking at the proportion of player’s last top 12 and last top 24 seasons that happen at a given age. We pulled all running back data since 1990 and, for each player, figured out the age at which their last top 12 and last top 24 season occurred. Then, we computed the percentage of these seasons that happened at each age. These percentages give numbers to help us study running back age decline from a fantasy perspective by looking at how long players hang on to be fantasy relevant. First, we present the percentage of last top 12 seasons that happened at each age.
Age | % of ‘Last Top 12 Seasons’ | Age (cont.) | % of ‘Last Top 12 Seasons’ |
22 | 5.6% | 29 | 7.7% |
23 | 7.7% | 30 | 10.6% |
24 | 13.4% | 31 | 3.5% |
25 | 9.2% | 32 | 3.5% |
26 | 12% | 33 | 1.4% |
27 | 12.7% | 34 | 0% |
28 | 12.7% | 35 | 0% |
And, now, the percentage of last top 24 seasons split by age:
Age | % of ‘Last Top 24 Seasons’ | Age (cont.) | % of ‘Last Top 24 Seasons’ |
22 | 5.1% | 29 | 10.2% |
23 | 6.8% | 30 | 6.4% |
24 | 12.8% | 31 | 6.4% |
25 | 8.1% | 32 | 3.4% |
26 | 11.1% | 33 | 1.3% |
27 | 13.2% | 34 | 0.4% |
28 | 12.8% | 35 | 0.4% |
Looking at the raw data above, it seems as if more running backs have their last top 24 season at age 27 than at age 28. While this is true, we notice that more players have already dropped out before their age 28 season than before their age 27 season. Thus, that 12.8% actually represents a larger proportion of the remaining players at age 28 than the 13.2% of remaining players at age 27.
In particular, we want to study survivorship. Survivorship refers to the percentage of players who make it to a certain point that survive past that point. The survivorship percentage at age 27 is the percentage of players who make it to age 27 before having their last good season who do not have their last good season at age 27. A bit more concretely, the survivorship percentage at age 27 is the percentage of players who have their last good season at ages 27+ divided by the percentage of players who have their last good season at ages 28+. It is 100% minus the percentage of players who make it to a certain point only to not continue past that point. It helps us measure the percentage of the remaining players who falter at a given age.
This is helpful from a fantasy perspective because it helps us evaluate whether or not a particular player is likely to continue to be successful. It is just another important piece in studying running back age decline. If we want to know how likely it is that a hypothetical 30 year old running back who just finished as the RB22 has another top 24 season, the numbers in the table below are helpful. First, the survivorship probabilities for a top 12 season:
Age | Survival % | Age (cont.) | Survival % |
22 | 95% | 29 | 65% |
23 | 93% | 30 | 66% |
24 | 85% | 31 | 48% |
25 | 89% | 32 | 43% |
26 | 83% | 33 | 25% |
27 | 76% | 34 | 33%* |
28 | 69% | 35 | 25% |
This data tells pretty much the story we would expect. Now, for the top 12 survival percentages:
Age | Survival % | Age (cont.) | Survival % |
22 | 94% | 29 | 71% |
23 | 92% | 30 | 44% |
24 | 85% | 31 | 58% |
25 | 88% | 32 | 29% |
26 | 81% | 33 | 0% |
27 | 76% | 34 | 0% |
28 | 68% | 35 | 0% |
In the following two images, we repeat the data shown in the previous two charts below as graphics so that we can visually appreciate the rapid aging of running backs after 30-31 years of age.
Now, as I’ve said before, this data is descriptive rather than predictive. If we’re considering drafting a specific player who is 28 years old, we don’t know whether or not this player is ‘still alive’. Such knowledge would require us to know if a player has at least one top 12/top 24 season remaining. If we want to make predictions we need to use observable data. So, instead of considering whether or not players have survived to a certain point, we use repeat probabilities to study running back age decline.
Repeat Probabilities For RB1/RB2 by Age
For the data in the two tables below, we studied the probability of an individual player repeating as RB1 or as an RB2. We took every fantasy football season since 1990 and looked at how often a player succeeded in repeating. For the top 24 repeats, the data is shown below. In the age 22 column, we see the number 89% which is interpretable as ‘of all the 21 year old players who finish in the top 24 running backs, 89% of them finish in the top 24 in the following year when they are 22 years old’. These numbers help us view running back age decline through the lens of drafting confidence.
Age | Repeat % | Age (cont.) | Repeat % |
22 | 89% | 29 | 45% |
23 | 57% | 30 | 46% |
24 | 65% | 31 | 48% |
25 | 60% | 32 | 23% |
26 | 72% | 33 | 22% |
27 | 60% | 34 | 20% |
28 | 55% | 35 | 25% |
And, the top 12 repeat probabilities:
Age | Repeat % | Age (cont.) | Repeat % |
22 | 60% | 29 | 31% |
23 | 60% | 30 | 43% |
24 | 50% | 31 | 22% |
25 | 50% | 32 | 0% |
26 | 58% | 33 | 17% |
27 | 46% | 34 | 0% |
28 | 36% | 35 | NA |
Comparing the two tables above, we see how much more difficult it is to repeat as a top 12 running back than it is as a top 24 running back.
All the tables and probabilities above scream one thing. The running back age decline happens very swiftly right around the 30-31 year old mark. Only 22% of 31 year old players repeat to have another top 12 season. Only 48% of the 31 year old running backs repeat a top 24 season. Moreover, we see the survival probabilities for top 12 and top 24 probabilities take a nose dive down below 50% at the 30 and 31 year marks.
Finally, we want to describe the regression model we built to study the running back age decline problem.
Logistic Regression for Running Back Age Decline
Logistic regression is a statistical technique used to predict probabilities of events based on continuous predictive parameters. For us, we want to predict the probability of a particular running back finishing in the top 12 or top 24 based on their previous production and their age. This section will be a bit mathematically dense, but if you are only interested in the results I direct you back to the first table in this article.
We made a slight adjustment relative to what we did for receivers. In our prior article studying wide receivers, the inputs to the model were only production and age. Because the exponent in the logistic model is linear in the model parameters, doing things this way implicitly assumes that the change in running back production incurred by growing from 24 to 25 years old is the same as the change incurred by aging from 29 to 30 years old. Obviously, this is not the case. In the simplest possible terms, the running back age decline is not linear in the player’s age.
How did we fix this? We added an auxiliary variable to our model. This new variable is given by max(0, age-27). Adding this variable into our model allows the relationship between age and production to be modelled by a linear spline with a single knot rather than simply being linear. We could extend this technique to more complex linear splines using the ReLU function while not sacrificing continuity of the model. Of course, extensions to arbitrarily complex linear splines will lead your model to suffer from the bias-variance tradeoff. Doing things the way I did here is an excellent example of adapting standard machine learning models by imparting domain-level knowledge to create more accurate models.
Concretely, this allows us to incorporate different age-based rates of change for different points in a player’s career. All this goes to say that our new model allows for more general conclusions. In particular, by studying the relative magnitudes of the coefficients computed by the model, we can compute a tradeoff between age and production. In particular, for every year up to age 27, the best predictor of next year’s performance is the previous year’s performance. However, for every year after age 27, one year is roughly equivalent to 4 spots of prior year performance. That means that a 28 year old number one overall RB and a 24 year old number five overall RB have about the same prospects for the next year. The main conclusions from studying the running back age decline problem is this: one year equals about 4 spots of performance.
