Conventional wisdom as defined by Wikipedia (yeah, I know) is “the body of ideas or explanations generally accepted as true by the public or by experts in a field. Such ideas or explanations, though widely held, are unexamined. Unqualified societal discourse preserves the status quo.”

I’m not a huge fan of conventional wisdom. There are three words or phrases in the above quote that irk me:

  • generally accepted
  • unexamined
  • status quo

While I haven’t donned a lab coat and safety glasses for five years (man, has it really been that long?) I still think like a scientist and prefer to rely on research and drawing my own conclusions as opposed to generally accepting unexamined ideas or explanations that lead me to keeping the status quo. That said, I don’t always find fault with conventional wisdom, but it never hurts to corroborate. Not to mention, things change and often conclusions supported by research are in need of reevaluation.

The methodology will focus on reliability and variance using correlation. Let’s first start with a review on correlation.

We’re going to compare the player’s salary versus his points scored in a given week. The idea being the salary is a proxy for predicting how the player will perform so we’re quantifying the ability to set prices that will reflect the impending outcome. Correlation is measured from -1 to 1. Complete randomness yields zero correlation. Using our parameters, 1 means the highest salaried player performed the best, the lowest priced player scored the fewest points and everyone else was scaled proportionally in between. A correlation of -1 says the lowest priced player did the best, the highest did the worst and the rest were scaled in between.

A coefficient of 1 signifies complete reliability. Don’t confuse this with quality; it just means everyone did exactly as they were supposed to do across all levels.

A coefficient of 0 says predicting performance is a complete crapshoot. That is, the variance is high.

DFS game theory says reliability is best for cash games while variance should be embraced when chasing a GPP. The other take is if it’s unpredictable, don’t pay for it in cash games.

Correlations will be determined for the first four weeks of the season using salaries and points scored from three sites: FanDuel, Draft Kings and Draft Day. Each site sets its own prices and uses its own scoring system so with ample weeks in the book, the ultimate strategy may differ according to site. Four weeks may not be sufficient to draw those sorts of conclusions.

To that end, today’s discussion will investigate three rules of thumb that purvey the DFS football landscape.

  1. Don’t spend on a kicker.
  2. Don’t spend on a defense.
  3. Use good quarterbacks for cash games, lesser ones for tourneys.

1. Don't Spend on a Kicker

Here's the data for kickers, with the note that Draft Kings does not use the position

  Week 1 Week 2 Week 3 Week 4
FanDuel -0.02 0.18 0.08 0.10
Draft Day   0.14 0.11 0.05

Sorry, I was unable to track down Week 1 kicker prices for Draft Day.

Chalk one up for conventional wisdom as there certainly appears to be a great deal of variance with respect to performance versus prices and this assumes there is some tweaking from week to week in an effort to adjust to team and player performance. If the outcome is so unpredictable, why pay extra to chase better numbers? Keep in mind correlation does not measure quality, but this does suggest it's a fool's errand to chase quality since there's limited (almost no) guarantee of getting it.

This is actionable in both GPP and cash game formats. However, with this advice being so widespread, a sneaky way to get a player with a low percent owned into your lineup is to take a kicker one price point above the lowest, assuming you have the excess room under the cap.

2. Don't Spend on a Defense

Below are the correlations from the first four weeks for defense.

  Week 1 Week 2 Week 3 Week 4
Draft Kings -0.08 0.04 -0.05 0.13
FanDuel -0.01 0.07 -0.28 0.36
Draft Day -0.14 0.09 -0.34 0.27

Now it gets a little interesting. Through three weeks, the ability to predict defense points was nill. In fact, in Week 3, the poorer defense performed better than they superior units as evidenced by a negative correlation. Then in Week 4, things changed. It could be with three weeks of data, handicapping was facilitated or it could have been dumb luck. For what it's worth, I've been doing this sort of study for traditional fantasy football using weekly projected points in lieu of salaries and this in fact resembles what usually happens -- defense starts random and develops some predictability as the season progresses. Though, the extent of the Week 4 correlation is higher than expected. Perhaps it has to do with six teams being on bye and a few games being predicted spot on so the reduced sample amplified the impact of the accurate projections.

Even with the elevated predictability, it's still nowhere near staunch enough to confidently spend a lot of budget on a defense. But it may be enough to take a chance on one above the bottom-feeders assuming you like the match-up. Continuing this exercise going forward will elucidate if accuracy predicting defense performance increases as the season wears on which in turn may alter game theory with respect to choosing one in future weeks.

3. Use Good Quarterbacks for Cash Games, Lesser Ones for Tourneys

And now the quarterback data. It was a bit more difficult determining the sample as several starting signal-callers got hurt. I opted to include the starters but not the replacements as well as omitting those that saw time during garbage time.

  Week 1 Week 2 Week 3 Week 4
Draft Kings 0.39 0.35 0.39 0.32
FanDuel 0.35 0.39 0.42 0.29
Draft Day 0.34 0.41 0.40 0.29

If nothing else, quarterback performance isn't a complete crapshoot, though there is still considerable variance. There's an additional consideration with this position that wasn't relevant to kickers and defense and that's the price of the quarterbacks, especially the superior ones, encompass a much larger portion of your budget than kickers and defenses and more importantly, score a lot more raw points. The offshoot is you don't want to make a mistake with your quarterback since you're less likely to make up for it with the rest of the roster. Missing at quarterback means being short a ton of points.

Due to the increased importance of the position, a second study will be undertaken to lend more data. What we'll do is break the inventory into thirds, then measure how well each third did relative to expectations. The idea here is the above table suggests there is some measure of predictability. Even more useful is if one tier of players displayed more predictability than the others, especially the top or bottom groups. Conventionally, the top tier is the target for cash games while the bottom one populates most GPP rosters.

Let's start with the top third, noting the players that comprise each group differ between sites.

TOP THIRD Week 1 Week 2 Week 3 Week 4
Draft Kings -0.02 0.41 0.09 0.29
FanDuel 0.07 0.33 -0.05 0.57
Draft Day -0.04 0.38 -0.18 0.47

Oy vey. One or the other, please? So in two weeks, the tier was a crapshoot but in the other, many of the better quarterbacks performed as such. With a split vote, you really can't say if using a top tier quarterback is the most efficient use of your assets. We need to conduct this analysis for a few more weeks to see which are real and which are outliers.

MIDDLE THIRD Week 1 Week 2 Week 3 Week 4
Draft Kings -0.24 0.22 0.43 -0.21
FanDuel 0.01 0.15 -0.03 -0.68
Draft Day 0.00 -0.34 0.02 -0.13

This tier is interesting since it isn't a target, at least on paper. However there is something very intriguing. If true, it's DFS gold (though I'm nowhere near ready to say it is so). Check out the negative correlations. These imply the quarterbacks at the lower end of the tier played better than those at the top, at least relative to expectations. This could be a latent area to target for tournaments. It must be noted Week 4 sticks out and if it's an anomoly, this all is useless - but it's definitely worthy of further study. In a week with two teams on bye, we're looking at QB16 - QB20 as GPP candidates.

LOWER THIRD Week 1 Week 2 Week 3 Week 4
Draft Kings 0.31 -0.19 -0.07 0.09
FanDuel 0.16 0.32 0.31 0.44
Draft Day 0.38 0.46 -0.01 0.20
This subset is a bit like the elite as some weeks appear to be predictable while others, no so much. Though on FanDuel, the worst quarterbacks have some measure of projectability each week. In other words, they were supposed to play poorly and they did. This isn't what we want for a tourney; we want the chance of an exceptional game. The message is targeting the rock-bottom signal-callers might not be optimal for GPPs.
To be honest, there isn't nearly enough quarterback data to decree any of these observations as money-makers. Factor in the machinations double on sites using two quarterbacks and we definitely need a larger sample. Though, it does suggest not to omit the middle tier as it may be optimal for tourney play.
The white elephant is each and every one of you is thinking none of this matters. You go strictly by your own projections versus price for all positions. That's great; so do I. But the only way it doesn't matter is if you feel you're a better prognosticator than those doing it for the various sites. And you very well may be. The point is each week, the sites adjust their salaries based on performance, albeit some more than others. They know what they're doing so completely ignoring the results isn't wise.
The other missing link is the reliability and variance of running backs, wide receivers and tight ends. This will be the topic of a future discussion. As a spoiler, in traditional fantasy football, the correlation for each spot is greater than that of quarterback. Something I suspect will be enlightening is how the research on the other positions can be used to aid in choosing a flex. Intuitively, you want reliability in cash games and explosiveness (or variance) in tournaments. It will be interesting if this research points to one position to the flex for cash and another for GPP.
With the disclaimer that we're still in the early stages of the season and ensuing weeks will avail more data, here's a list of take-home messages I am going to apply to my own game-play. Your mileage may vary.
  • In both cash and GPP, I'm using the kickers just above the lowest priced option
  • In both cash and GPP, I'll look to save money on defense but I won't constrict the inventory to the very bottom squads
  • In cash, I'll use my own estimations but will look to the upper third unless there is someone I really like elsewhere.
  • In GPP, I'll jump up my concentration from the bottom third to just above that in the middle tier.
Before I call it a day, I'd like to thank Dave Hall for hooking me up with a link to his web site where he archived the data used in this research.
And finally,  I'm more than happy to put my money (a little of it anyway) where my mouth is. If you want to take me on head to head on any of the sites, shoot me an e-mail (todd dot zola at gmail dot com), post in the comments or hit me up on Twitter @ToddZola. We'll play for more than a buck but it won't be for triple digits,
Thanks for your indulgence and good luck this Sunday. Unless you're playing me.