Apex Insider Blog

After writing The Peak Age for an NFL Running Back, I became intrigued with how age affects running backs and how predictable the decline is. Data analysis programming languages, like Python and R, are very helpful when studying historical seasons. Let’s see how successful a predictive model using running back age and PPR scoring can be.

Defining a Peak Season

To win fantasy football championships, we are all attempting to identify quality or peak running back seasons — elite players we can start each week.

Typically, the RB12 hovers around 230 PPR fantasy points, which equates to 14.4 PPR fantasy points per game. For the purpose of this study, I included all running back seasons since 2000 to score at least 230 PPR fantasy points or at least 200 PPR fantasy points and 14.4 fantasy points per game.

To determine successes or failures, I charted whether the player maintained the same pace (14.4 fantasy points per game) in the following season. If the player maintained the same pace, he was charted as a success. If he failed to score 14.4 fantasy points per game, he was charted as a failure.

The study includes 234 individual running back seasons.

Note: A fair number of seasons were removed because of inconclusive results. For example, Jamaal Charles’ 2011 season was not included because in the following season, he was injured in Week 2. If results for the following season were not conclusive, the season was removed.

The table below contains all of the seasons that were included to create the models. The ‘0’ or ‘1’ under N + 1 indicates whether the player succeeded in the following season.

Preface

When it comes to machine learning, you must build and teach a model with a training set. Then, to analyze the accuracy of the model, some observations must be set aside to test the model.

In Kernel SVM I used 75 percent of the observations to train the model, meaning 175 of the 234 seasons were used as a training set. The remaining 25 percent, or 59 observations, were used to test the model.

Based on the type of data being analyzed, the two best algorithms appear to be Kernel SVM and Naïve Bayes. We’ll start with Kernel SVM and see which algorithm is more accurate.

Kernel SVM Model

Let’s first understand the chart. The Y-axis refers to PPR points per game while the X-axis refers to Age. The green points indicate players who succeeded in the following season, the red points indicate those who did not. The top left is young, high-scoring players. The bottom left refers to low-scoring, young players. The top right refers to old, high-scoring players. The bottom right refers to old, low-scoring players.

Note that there are far fewer points on the right side of the graph — 84.6 percent of the seasons took place before age 29. Once you include the age-29 seasons, it accounts for 214 of the 234 seasons — 91.4 percent. It is hard to just make this list at 28 years old or older, never mind succeeding in the following season.

The conclusion we can draw so far is the model doesn’t like the older, low-scoring seasons. The drop-off begins for low scoring seasons between age 27 and 28 seasons. At age 29, a running back play looks likely to fall off. This is the training model, which the machine learns from. Now so we will see how the model fares in predicting the remaining 59 observations.

Let’s check the results of the Kernel SVM Test.

Out of 59 total observations, the Kernel SVM model predicted 45 correctly, which is an accuracy rate of 76.2 percent. While there are some incorrect predictions on the boundary, the model does a decent job of predicting the drop-off.

Naïve Bayes Model

For the Naïve Bayes model, there are also 59 observations to test the model or 25 percent of the seasons. 175 observations were used to train the model.

The Naïve Bayes algorithm returns a similar result – as the curve appears to follow a very similar line. Let’s check the results of the Naïve Bayes model test.

Out of 59 total observations, the Naïve Bayes model predicted 47 correctly which is an accuracy rate of 79.6 percent. It was a little more accurate than the Kernel SVM model and on average, will predict four out of five seasons correctly.

What Does This Mean for 2020?

The two things that cause the models’ concern are old age and low scoring seasons. I wrote this article last season and the model advised avoiding David Johnson.

In 2019, 14 players scored at least 230 PPR fantasy points or at least 200 PPR fantasy points and 14.4 fantasy points per game. Each player is listed below.

Mark Ingram will play next season at 31 years old and with a relatively low-scoring RB1 season, the model lists him as a clear fade.

Playing next season at 24 years old after averaging over 29 fantasy points per game, the model couldn’t be more optimistic about McCaffrey’s odds of remaining elite next season.

The model prefers the running backs who will be 25 next season to the backs who will be 26. The soon-to-be 25-year-olds are Ezekiel Elliott, Austin Ekeler, Dalvin Cook, Leonard Fournette, Nick Chubb, and Alvin Kamara.

The soon-to-be 26-year-olds are Aaron Jones, Derrick Henry, Chris Carson, Todd Gurley, and Kenyan Drake. Jones and Henry avoided the danger zone by scoring over 19 fantasy points per game. The model is less optimistic about Carson, Gurley, and Drake, who all hovered around 15 fantasy points per game.