Is it Time to Scrap Zero RB?

Zero RB appears to be in trouble, and its creator saw it coming. Last August in an article about the 2016 Apex Expert’s League, Shawn Siegele wrote the following about the impending decline of Zero RB as the top fantasy drafting strategy:

We’re probably at peak Zero RB. One of the young studs at RB will likely have a great season and lead some of his owners to victory. Enthusiasm will die down, and many Zero RB drafters who do not experience success will return to their roots.

2016 provided just such an event, with both David Johnson and Zeke Elliott leading many of their owners to championships. The enthusiasm for Zero RB, even among some of its more ardent supporters, is waning. There is a crisis of confidence.

Standing here now, in the ashes of a failed Zero RB season, is it still reasonable to think Zero RB is a dominant strategy? Or did 2016 provide us with evidence that should force us to update our prior beliefs? Should we go back to attacking the draft based on value?

To answer these questions I had to take a hard look at the underlying assumptions of Zero RB. 1

Premise #1: We are not very good at projecting football players

While it may sound obvious to anyone who has played fantasy football for a while, this is actually pretty controversial. When you make this assumption concrete — when we becomes you — most of us, myself included, exempt ourselves from the rule. “Yeah, it’s true those plebes over there are hot garbage at projecting players, but I’m actually pretty okay.”2 Sadly you are a part of we in all likelihood, as tough as that is to hear. I’m sorry.

Still, I’ve made a claim without evidence. How do I know we suck?

If we look at ADP from 2009 to 20163 and compare it to end of season ranks, you can see clearly just how bad we are at this stuff. Among all positions, our collective ability to forecast an individual player is off by an average of a round and a half (18.1). 4

If we break the errors down by position we get the following:

It’s probably to be expected that when people are asked to forecast a larger number of players that their accuracy will suffer. It is also the case that, as Anthony Amico has shown, we are better at predicting volume than we are efficiency. So players with low volume, and thus high ADPs, will be harder to accurately project.

Top 50 RBs and WRs are equally difficult to predict, which is probably not much of surprise since injury rates between the two positions are relatively close in this cohort.

It’s also interesting to see QBs at the top of the list. RotoDoc has shown that 2016 was an extremely challenging year in QB forecasting. But perhaps it’s not all that surprising given only 32 QBs can start at any one time. When your pool of relevant QBs is smaller, you’d expect the mean absolute error to be smaller.

Zero RB is concerned with the high leverage rounds though. So let’s look at the top 24 ADP players by position. These are the guys taken in the first 5 rounds from 2009-2016.

Screen Shot 2017-01-11 at 4.10.12 PM

As expected, when we are forecasting half as many players, our collective accuracy improves. We are still about a round off on players on average, which is terrible, but our ability to forecast WRs has outpaced our ability to forecast RBs.

Unfortunately, my choice of error metric means we are underestimating the actual error here. First round draft picks are insulated from mean absolute error because under-estimation is truncated. You aren’t punished as much when a guy drafted at RB4 actually ends up being the RB1 by a country mile. So the error is likely underestimated for those players taken in the first round.

Finally averages can only tell us so much. They are point estimates after all. What we really want is a look at the distribution. And that brings us to premise number 2.

Premise #2: Wide Receivers are easier to forecast than running backs

The following is the distribution that drove Zero RB coming into the season.

Screen Shot 2017-01-09 at 5.30.17 PM

Tall spikes on the left are better. They indicate that there were fewer large errors in the forecasts for that position. We can clearly see that WR forecasts have a lower error more frequently than RB forecasts from 2009-2015.

Here is another view of the same data as a probability density curve.

Screen Shot 2017-01-11 at 8.43.18 AM

Again, it is clear that the odds of choosing a WR that meets his forecasted ADP is greater than that of a RB.

Now let’s add in 2016 and see how it changes the curves. Since the sample includes six years, adding a seventh represents a 15 percent increase in data.

Screen Shot 2017-01-11 at 4.57.01 PM

Notable changes include a slight shortening of the WR peak, and the addition of a bump at the extreme right tail of the WR distribution. There is also an increase in the probability of RB errors falling in the 40-50 range, or roughly 4 rounds worth of error.

All in all however, I think it is very fair to say there has been little change to the distribution.

Zero RB DGAF about value

We talk a lot about value in fantasy football. I think it is often misguided, especially in the context of structural drafting.5

When you combine the two premises laid out above, a specific set of principals for drafting fantasy players follows:

  • We know we are going to miss on our forecasts, sometimes by many rounds. This can be thought of as irreducible error, or variance. To reduce variance of our fantasy results, statistics teaches us that we must add more trials, or increase our sample size. We need to take an excess of some position so we can reduce our variance relative to our competitors.
  • To load up at one position, we must forego another. There are only two viable choices since most leagues only allow us to start one TE and one QB: RBs or WRs. Top 24 ADP WR are easier to predict than RBs. Even better, forecasting errors at the WR position cluster around low error values. So WR are the obvious candidates.
  • A final piece of the puzzle revolves around the reason for WRs being easier to forecast: the top 24 WRs by ADP are far less likely to miss four or more games due to injury than the general population of WRs. Meanwhile the top 24 RBs by ADP are far more likely to miss four or more games due to injury than the general population of RBs. Because of this, RBs are replaceable on the waiver wire in most years, further making the strategy of over-drafting WRs viable.

Nowhere in the above thesis is there any discussion of value. Zero RB is a structural strategy. Value is incidental. It’s a separate discussion.

This is an incredibly important point, and a misunderstanding of this key idea drives most of the confusion about the strategy.

Zero RB takes no position on player value! At some point we must choose WRs, this is true, but which ones you take are far less important than the fact that you take a lot of them. Furthermore, you should take WRs when the forecasting error is known to be lowest relative to RBs: in the high leverage rounds.

It’s the recognition that we suck at forecasting that drives us to try and reduce our personal variance relative to our competitors. The only way to do this is through volume because we can’t get a sustainable edge by beating the projections market. We had to choose one position or the other to load up on, and we chose WR, but not because Zero RB says that WRs are a better value. Instead we chose WR because they are more predictable relative to RBs.

Finally, it is this constraint — limiting the pool of players to those with lower variance and choosing a lot of them — that gives us our edge. It is also why Upside Down Drafting or Do The Opposite are not at all equivalent to Zero RB. Value is a sideshow.

Conclusion

Zero RB is not dead. It should not be scrapped. It is immortal.

Hyperbole? No. What drives Zero RB is a distribution. What drives that distribution is injuries, and the injury trend is secular. RBs will nearly always have a higher serious injury rate, and an injury means zero fantasy scoring.

There is a theoretical point at which league settings could make drafting a RB incredibly tempting from an expected value standpoint, but that is incidental to the structural drafting recommended by Zero RB. All RBs will see that scoring increase, and waiver wire replacements for injured RBs will grow in value commensurately.

That is your edge. Will that edge always win you your league? No, of course not. The edge is not large enough to overcome variance every year. But the edge is stable and it is exploitable. Unless the distribution of forecasting errors changes at some point in the future, Zero RB will remain viable and will provide its practitioners with a small, but real advantage over their competitors.

  1. Most of this has been covered elsewhere by Shawn himself.  (back)
  2. I know. I feel the same way. I’m really quite convinced I’m better than average at everything I do. The only problem is I’m wrong.  (back)
  3. Top 50 ADP cut-off for all positions.  (back)
  4. Accuracy here is defined as the absolute mean error between ADP and end of season rank. In other words, if LeSean McCoy‘s positional ADP was RB 11 and he finished RB 5 that would be an absolute error of 6.  (back)
  5. Of which Zero RB is one example.  (back)