2 Running Backs Who Are Due for Fantasy Football Regression in 2017
Sometimes, a player's performance on the field does not match up with his fantasy production.
I created some models last year to help evaluate fantasy production as it relates to numberFire's Net Expected Points (NEP). On every play, there's an expected point value an NFL team has for the drive based on yard line, down, and distance. What happens on that play can change the expected point value on said drive. What NEP does is aggregate the value gained or lost on every play into a single, net number.
You can read more about NEP in our glossary.
These models helped identify players whose on-field performance varied drastically from their fantasy production and predicted that they would regress closer to their mean in 2016. So, we're back here to do it again with the same exercise for the upcoming season.
2017 Model Changes
The running back model has probably seen the most changes of all the positions from last season.
Last year's model looked at the last five years of data, while the 2017 edition looks all the way back to 2000. The touch threshold last year was set at 140 carries, but that was changed to 100 opportunities this year, defined as a carry or target. These two changes brought the player sample from 168 to 938, a massive swing.
This is also the only model that includes two different inputs: Rushing Successes per game and Reception NEP. A Rushing Success is any rush that yields positive NEP, or an expected points gain greater than zero.
These two components were combined to create RB Score, which was defined as Rushing Successes per Game times a coefficient of nine plus Reception NEP. This year, the equation for RB Score was found by using a multiple regressions between Rushing Successes per Game, Reception NEP, and fantasy points per game. Here were the results.
| Component | Coefficient |
|---|---|
| Y-Intercept | 1.201395267 |
| Rush Successes/Game | 1.540353709 |
| Reception NEP | 0.200909451 |
This means that the true best-fit equation between the inputs and fantasy points per game is (roughly) Rush Successes per game times 1.54 plus Reception NEP times 0.2 plus 1.2. The coefficient of determination for this equation is .8482, better than last year's .8221.
The following table shows the history of running backs in the sample based on different Fantasy Point Per Game over Expectation (FPPGoE) thresholds. The first column gives the threshold, the second gives the number of running backs above that threshold, and the third gives the number of running backs in that sample who were worse the following season.
Performance is being checked in this way because we expect backs who outproduced their on-field play to decrease the following season. Also included is the percentage of running backs in each bucket who regressed and the average change in scoring the following year for the entire group.
| FPPGoE Threshold | RBs Above | RBs Who Regressed | Percentage Regressed | Avg +/- | Opportunities | Opportunities Year N+1 |
|---|---|---|---|---|---|---|
| 3 | 46 | 37 | 80% | -2.623 | 273 | 286 |
| 2.5 | 66 | 50 | 76% | -2.007 | 268 | 282 |
| 2 | 84 | 64 | 76% | -1.986 | 268 | 270 |
| 1.5 | 122 | 90 | 74% | -1.975 | 270 | 271 |
| 1 | 171 | 123 | 72% | -1.856 | 270 | 261 |
| 0.5 | 228 | 154 | 68% | -1.480 | 261 | 256 |
You'll notice the number of opportunities the running backs received were included here. Since volume is so critical to running back production, I wanted to make sure that any change in scoring was not brought on by additional noise.
The biggest change in scoring comes at the 3.0 FPPGoE threshold. At that point, 80 percent of running backs regress, and at a significant rate of -2.623 points per game. This will be the threshold we use to identify our negative regression candidates.
We can also check for significant thresholds for positive regression candidates, and we did the same exercise for this situation in the table below.
| FPPGoE Threshold | RBs Below | RBs Who Regressed | Percentage Regressed | Avg +/- | Opportunities | Opportunities Year N+1 |
|---|---|---|---|---|---|---|
| -3 | 25 | 16 | 64% | 0.415 | 303 | 264 |
| -2.5 | 40 | 23 | 58% | 0.021 | 281 | 244 |
| -2 | 82 | 42 | 51% | -0.120 | 276 | 241 |
| -1.5 | 124 | 60 | 48% | 0.007 | 262 | 233 |
| -1 | 178 | 93 | 52% | 0.187 | 253 | 234 |
| -0.5 | 237 | 124 | 52% | 0.305 | 248 | 234 |
This is where things get interesting, and it brings up some intriguing discussion points. You'll notice there is no significant FPPGoE threshold -- the highest regression percentage is 64 percent, but it comes with an average plus/minus of just 0.415.
The reason for this appears to be volume. While these running backs produced below their play on the field, it led to reduced opportunity the next season (Year N+1). The average number of opportunities across samples in Year N was 270, while that number dropped to 242 in Year N+1. As a result, even if these running backs performed better, that was mitigated by a loss of touches, so fantasy scoring remained largely unaffected.
This contrasts well with our negative regression candidates, whose touches remained largely unchanged after outperforming their play on the field. Negative regression candidates averaged 268 opportunities in Year N, and 271 in Year N+1.
I find this to be especially interesting because it shows that teams are probably a little too obsessed with counting stats, even when scouting their own players, rather than looking at the true value the running backs brought to their team. For instance, the player lowest in FPPGoE for 2016 was Spencer Ware at -3.53. We would expect him to bounce back...except the Kansas City Chiefs traded up in the draft to select Kareem Hunt, a move that suggests the organization is displeased with Ware's production.
The odds of improving or decreasing in scoring for a random running back was slanted towards decline. Of all running backs in the sample, 57 percent saw a decrease in scoring, while 43 percent improved. That puts the model above random chance in predictive power by 40 percent for negative regression candidates.
While there has been a lot of interesting information dissected so far, it has led us to just two regression candidates at running back for 2017.
Theo Riddick, Detroit Lions
FPPGoE: 5.58
Theo Riddick has been a tremendous producer in the receiving game for the Detroit Lions for two straight years. In 2016, he was asked to do a lot more on the ground after Ameer Abdullah got hurt (before getting injured himself).
While he finished as the 8th-highest scoring running back in fantasy points per game, his expected production was closest to Giovani Bernard, who finished as the 28th-highest scorer. With Abdullah back, Riddick should fall off in scoring somewhat naturally, and he could crater if his production regresses back to its mean. Riddick's FPPGoE was the second highest since 2000.
LeSean McCoy, Buffalo Bills
FPPGoE: 3.95
Lesean McCoy was exceptional in 2016, finishing as the fourth-highest scoring running back in fantasy points per game. However, he vastly outproduced his expected performance. McCoy's expected production was incredibly close to that of Mark Ingram, who finished as the 10th-highest scoring running back.
That doesn't sound like a big difference until considering that McCoy averaged 4.7 more points per game than Ingram. Shady's status should not be in question -- he just shouldn't be considered an elite fantasy running back.
