The website is unlocked and in off-season mode. The 2023 season is displayed, and you can explore all the site features freely!

Week-to-week Predictability: "The 0.3 Target"

While making all my kicker model improvements, I have had my eye on this particular accuracy number. A correlation coefficient of 0.3 is sometimes considered an unofficial "cut-off" for what is reliably predictable. You'll see why below.

There is no fantasy position with "good" predictability. A simple scatterplot shows how frequently busts occur for a more predictable position. Despite the already-high level of randomness in fantasy, kicker rankings have historically been the least accurate. I have measured other ranking sources for some years-- Most sources yield a projection accuracy of 0.15, and the best ones reach 0.20-ish, on average. Well below 0.30. (In comparison, RB1 and QB reach predictability in the >0.35 range. The other poor-predictability outlier is WR1, floating around 0.25.) My ambition with improving kicker accuracy is to make it even more fantasy-worthy than WR1. And ideally above 0.30.


The good news is I think I've nearly cracked it; but the bad news is people won't always notice it. There are 3 reasons that accuracy improvements are not obvious

  1. fantasy randomness plagues all positions, so 0.30 is still "bad" (this is gambling folks! That's why I have a kicker wall of shame). 

  2. You can still end up picking my duds, even during an otherwise good-accuracy week. 

  3. Ranking accuracies will fluctuate, week-to-week.


This 3rd point is what I want to show you today. Example: Although I had a good 0.45 accuracy in 2021 week 5, my week 4 accuracy was almost 0 (therefore "random"-- 4 of my 8 picks were bad). So let's ask "how often does that happen?"

Here is your answer, in the form of a smoothed histogram:

My model (applied historically), yellow, yields a correlation accuracy that is negative-or-near-zero in less than 10% of weeks. Meanwhile, the blue curve represents "other sources" (with an optimistic average of 0.23 I can explain the details in the comments). As you can see, a less-accurate model will be "around zero or negative" in up to 20% of the weeks.

And there is your partial explanation of why a 0.3 average correlation is desirable: we hope to ensure that only 5% of weeks will have worse-than-random correlation. An accuracy of 0.20 is not enough to guarantee that. But 0.30 can apparently come close.