A Newer RB Model
Or: Predictin' Ain't Easy
Predicting running backs is, in many ways, a paradox. The secret sauce is easy to figure out: league success is heavily contingent on a back’s college production, with receiving prowess being a particularly strong signal. Maybe you can add a bit of athletic testing in for good measure, but the overall picture is pretty clear by now.
Yet over the course of building my new model, I learned that predicting running backs is anything but a closed case. It’s easy enough to get a good model, and at least in fantasy terms, beating draft capital isn’t too hard either. But making a truly great model is a different matter entirely, with so much room for marginal improvement it’s easy to fall down the rabbit hole.
Another hurdle is juggling the burden of explainability while chasing better performance. Again, it’s not too hard to achieve good accuracy if you’re willing to throw a bunch of esoteric metrics into a complicated machine learning algorithm. Yet explaining why your model behaves the way it does is another challenge entirely, one arguably more difficult than just boosting metrics.
Designing the model
So how, exactly, did I manage to distill all these complex insights—ones learned from years’ worth of training data—into a relatively simple, five-feature model? Well, I cheated a little, combining players’ overall and route-running grades to get a composite PFF grade. (Strangely enough, simply adding the two together provided the best signal to the model; however inelegant it seems, it’s still effective.)
As you can see in the chart above, our top feature is still draft capital (specifically log-transformed pick number, which outperforms raw pick). Our composite PFF grade is also pretty important, and is the first feature I’ve seen to even come close to draft capital (it even surpassed draft capital in 2021 importance).
Much credit, really, is owed to PFF data for this model. Their elusive rating provides information that doesn’t cleanly emerge from simple rushing rate or volume stats. Ditto for targeted QBR, which finds itself tied with elusive rating in terms of importance.
Lastly, we have height-adjusted speed score (HaSS). It’s lucky that this fell out organically while searching for the best feature set, because I always like to have some athleticism numbers fed into the model. Even if it’s the least important of our main features, it’s proven extremely useful in some of the more difficult classes (e.g., 2021).
Model performance
As was the case for our receiver model, we’re going to be using the Spearman correlation coefficient to evaluate our model’s out-of-sample performance. Essentially, for each holdout year, we first train a model on all the data available to that point. For example, if we’re holding out the 2023 class, we feed it all the three-year outcomes from 2016 through 2022.
While our final model ultimately ends up using all the data available to us—i.e., the 2016 through 2023 draft classes—this gives us a decent idea of how our model will perform on data it hasn’t seen before. If, by using the same model specifications and features our final model does, we can achieve good performance on held-out historical draft classes, there’s a pretty good chance our model will perform well in the future, too.
The chart above demonstrates how our model specifications perform on recent draft classes. Already, the value-add of our setup seems pretty apparent, with our model significantly outperforming a draft-capital-only approach in the 2021 draft.
It’s a great example, frankly, of why I chose Spearman coefficient as an evaluation metric over r². While my RB model also does pretty solid on r² (.57, weighted across all hold-out classes), its performance on rounds two through six speaks volumes. By ignoring rounds one and seven—which are layups to rank—we see how our model does on the more difficult mid-round prospects.
A 2025 refresher
One piece of feedback I received on my receiver model piece was that some people wanted to see how it predicted past classes. The issue with this, of course, is that since my final model was trained on classes through 2023, it already knows how those players did.
What’s the solution, then? Luckily, since I only trained on data through 2023 (my model predicts for a three-year horizon), my model was totally blind to the 2024 and 2025 classes. As a result, they’re about as close to “pure” unseen data as you can hope for.
Overall, the new model isn’t that different from last year’s version when it comes to the 2025 class. The model rightly fêted OSU product Quinshon Judkins, while also correctly identifying how RJ Harvey’s receiving ability would boost his dynasty floor. Notable, too, is its favoring of Bhayshul Tuten over Kaleb Johnson, despite the latter going well before the former in the NFL Draft.
The mid-round guys are more of a mixed bag, with few definitive successes or failures. Woody Marks, for example, was flagged as “low-upside,” but my model also correctly did the same for Cowboys flameout Jaydon Blue. Perhaps the savviest call by my model was correctly recognizing “Bill” Croskey-Merritt as a noteworthy talent. Though his upside was marked as ‘Low’, he only missed the ‘Medium’ upside cutoff by nine points, and his prospect score roughly matches the likes of earlier picks Devin Neal and Ollie Gordon.
This demonstrates what I think the most important aspect of a predictive model is, at least in fantasy football terms. It isn’t just about ranking players correctly, or getting a higher R²; it’s about giving managers a competitive edge of some sort, a way to arbitrage. Being correctly bearish on Kaleb Johnson is great and all, but not everyone has top-10 picks to burn. But everybody could’ve bid on Bill, and that’s the kind of advantage we’re seeking here.
The 2026 Class
Before we look at my model’s actual predictions, it’s worth addressing the bad vibes engulfing this RB class. Much of this negativity likely comes from whiplash due to how good last year’s class was, but it’s important to remember just how exceptional the 2025 class was. Most recent classes resemble this year's crop instead: an elite guy or two, then some decent second-tier dudes who, per my model, score in the 70’s and 80’s.
The real difference between this year’s class and prior ones is its atrocious depth. If you tab over on the table above, you’ll see that the next prospect after Emmett Johnson, Nicholas Singleton, scores 30 points below him. This is likely to improve somewhat after the draft (assuming some of these guys go earlier than their consensus board ranks), but it’s still a stark indication of how shallow this RB class is.
Early notables
My model clearly likes Love and Price, but they’re known commodities. The real question is, how do we find enough other quality players to salvage this less-than-stellar RB class? Luckily, my model found three more guys who, while not all necessarily first-round dynasty locks, are at least worthy picks in the early second round.
The best of these three is Washington product Jonah Coleman, whose composite PFF grade and targeted QBR are both top three in this class. The only thing holding him back is his just-OK elusiveness rating, which comes with a bit of an asterisk. Assuming a sufficient sample size, our model uses a player’s best collegiate season (in fantasy PPG terms) to pull targeted QBR and elusiveness from.
By this approach, Coleman’s best college season was 2025, largely on the back of receiving yards and touchdowns. His standout work as a pass-catcher this year obviously boosts his targeted QBR, but his elusiveness rating took a nosedive, clocking in at barely half his previous number. Even with his terrible 2025 elusive rating, though, the model still likes him a lot, and if his previously strong ratings are to be believed, the sky is the limit for Coleman.
Scoring slightly lower than Coleman are Arkansas’s Mike Washington and Nebraska’s Emmett Johnson. Near-polar opposites, Washington is largely carried by his combine testing, while Johnson gets by on his solid PFF grades and receiving ability. It’s worth noting, however, that speed score is my model’s least important feature, meaning Washington’s profile is mostly carried by his earlier expected draft position. Thus, while Washington was the better tester, Johnson is likely the better overall player.
Late standouts
While I do think the time spent on the top guys in this class was worthwhile, there are some late-round guys who still warrant attention. Eagle-eyed readers may note, for example, that Pitt RB Desmond Reid’s score far outpaces his mock-consensus rank, but because this means scouts see him as just a priority UFA, he’s still flagged as a ‘Low’ upside player. Thus, I think it’s better to use this space to address the three late-rounders our model pegs as having ‘Medium’ upside.
The first of these guys is Penn State product Nicholas Singleton, who could go earlier than expected this weekend due to his past pedigree. He’s uncannily similar, frankly, to Coleman, at least when it comes to his excellent PFF grades and targeted QBR, the latter of which almost exactly matches Coleman’s. Only his elusiveness rating is meaningfully worse, but while that is a considerable red flag, the numbers support the notion that Singleton can be a valuable post-hype sleeper.
The other two late-rounders of note—Wake Forest’s Demond Claiborne and Kentucky’s Seth McGowan—couldn’t be more different. Like similarly small speedster Desmond Reid, my previous analysis flagged Claiborne as a no-go, and with his light-in-the-pants profile (he weighed in below 190 at 5’10”) it’s easy to see why. Yet he’s got real receiving bona fides, with a borderline elite targeted QBR, suggesting a creative coach could get something out of him in the league.
McGowan, meanwhile, weighed in over 220 pounds at 6 feet flat, making his 4.40 and elite broad jump numbers staggering. Yet his miserable elusive score hints at some underlying red flags, with his just-OK 10-yard splits suggesting there might be some holes in his profile after all. Still, if I had to pick between the two, I’d definitely lean McGowan, and if he goes early enough (or at least slots into an open enough backfield), managers should definitely take notice.
Summary
Ultimately, while I think this pre-draft projection exercise has been pretty useful in defining the overall shape of this running back class, the specifics are simply going to be murky until we get actual NFL draft data. While using log-transformed draft pick makes the difference between late-round guys a lot smaller, a guy going earlier or later than projected is still going to affect his projection fairly strongly.
Nowhere is this truer than with the mid-round guys. At the end of the day, fantasy football (and real football, really) is about opportunity more than anything, and the teams and the level of draft capital a team puts into these guys can drastically impact their NFL career outlook. While I wouldn’t overreact, where these guys land in the draft is still going to matter, and to pretend otherwise would be misleading my audience.
But I’ll also say that my running back model this year has a lot more going for it than just draft capital, which isn’t something I can say of last year’s. Some of our holdout test years even showed PFF grade surpassing draft capital in predictive power, which, in my opinion, speaks to the importance of player quality. At some point, you have to go beyond mere notions of value and ask yourself one simple question: do I want this guy on my team? And if your answer is yes, you’ve got yourself a player.
Loves the article. Great work.