Mat Irby's Apex Score: Tight Ends

If you’ve been sticking with me lately, you’ve seen me slowly roll out a metric I’m calling Apex Score. The idea behind Apex Score is to boil indicators of future upside down to a single number. This, alone, may be foolish.

I’ve gotten really deep into the weeds on this process, even when I threatened not to. If you’d like to go back and check out the nearly 20,000 words I’ve already written on how it works and whom Apex Score likes or doesn’t like vs. ADP, check out my last three articles:

• Introducing Apex Score: Quantifying Upside and Beating Redraft ADP at Running Back
• Mat Irby’s Apex Score: Wide Receiver
• Mat Irby’s Apex Score: Quarterback

HOW APEX SCORE WORKS AT TIGHT END VS. OTHER POSITIONS

Apex Score is difficult enough with WR and RB, where there are more data points; I created a special hell for myself by doing this with the onesie positions.

At QB, Apex Score was particularly difficult because specific players with buoyant profiles entered the dataset four or five times. Finding the balance between a rush-heavy model that favored Marcus Mariota and Justin Fields and a more balanced model that suppressed Josh Allen was almost impossible. However, I eventually found a combination of stats that not only correlated well with Apex N+1 and had strong YoY correlation, but also left room for a pure pocket passer to enter the conversation (of course, we proved it sufficiently by the fact that Joe Burrow, Matthew Stafford, and Jared Goff all came in below ADP, so these QBs will probably never be favored by the model, and I think… that’s okay).

But TE presents new challenges. Obviously, we have certain players like Travis Kelce and Mark Andrews who creep into the top end multiple times, but what makes them special is a bit more opaque than what is going on at QB. At QB, we can really eyeball it and say the best players combine a high-floor rushing platform with high-end passing. Full stop. That makes testing the QB model easy, with just a sense of oughtness.

At TE, we’re trying to process a crime scene. There are all sorts of materials in evidence, and we have to get out the tweezers to figure out which things are important and which aren’t. I actually have less faith in TE Apex Score than any of the others; not to encourage you to give up and see what J.J. wrote this week, but it’s just harder to look at the finished product and know: this is right.

At QB, even after all the statistical wrangling, there was ultimately a football answer underneath it that made intuitive sense: the best fantasy QBs tend to combine rushing with high-end passing. At TE, I’m not convinced there’s an equally simple underlying truth.

For now, though, I’ve got a model that correlates best to Apex N+1 and PPR N+1. Its YoY stability is above average, and its Apex Capture (how frequently Apex TEs have hit this one metric) is well above average.

THE TRAVIS KELCE PROBLEM

To ensure specific candidates weren’t dragging us along, I tried the model by essentially removing one TE at a time and retesting. When specific parts of Apex Score seemed overly dependent on specific players, I tried different statistics that told the same part of the fantasy story – volume, target-earning, efficiency, etc. – then tried again. I didn’t want to build something that only worked because it was Travis Kelce’s player-specific move.

And Kelce did matter. Removing him dropped Apex N+1 from .410 to .339, with especially large effects on three of the model’s four inputs. That led me to a version replacing YAC/G with YACO/G, which was less dependent on any single player and improved both YoY and PPR N+1 while barely hurting Apex N+1 (.410 to .405).

So I put the two versions through another battery of tests. The original held up better when removing individual seasons, limiting players to one transition, and looking specifically for new apex TEs. In thousands of player-level resamples, the YACO version beat it in Apex N+1 only 34% of the time.

So YAC/G stayed. I’m not looking for the combination that produces the prettiest number. I want one where the signal survives when I try to break it.

TIGHT END APEX SCORE IS KIND OF A BUZZKILL

If I’m being honest, part of what I don’t like about the model is that it didn’t give me the answers I wanted.