My board over-rates receiving tight ends. Touchdown regression isn't the reason.

What the consensus says

Two takes share a bloodstream every offseason. The first is the perennial breakout-tight-end tier: this is finally the year, and here are the names. Footballguys runs a literal "Catch-A-Rising-Star" tier; RotoBaller hunts "sneaky TE1 candidates"; FantasyPros publishes its breakout-tight-end list like clockwork. The second is the cooler-headed counter, touchdown regression: CBS's projection model names Dallas Goedert (11 scores in 2025 against a 3.4-per-year career rate) and Harold Fannin (seven TDs, 3.8 expected) as guys whose end-zone luck is about to correct. Both takes are really one bet: a tight end's big receiving year tells you something durable, and the touchdown spike is the part most likely to be a lie.

I have a reason to care beyond the discourse. My own engine's error report, the one that grades old rankings against what actually happened over the next three seasons, flagged the receiving-tight-end archetype as a place SignalTuned over-rates players. So I went looking for the TD mirage in my own board.

The claim, in plain English

The honest version is a split. Take the receiving tight ends, the ones commanding at least a 12 percent target share, the archetype the engine treats as WR-like and slow to decline. Inside that group, flag the ones whose year carried an elite touchdown rate, top quartile, scores per catch. Then ask whether those TD-spike seasons fall apart the next year relative to receiving tight ends doing the same volume without the spike.

def predicate(row) -> bool:
    # in-cohort = elite Y0 receiving-TD rate among receiving TEs (top quartile)
    return row["td_rate"] >= TD_RATE_CUTOFF   # receiving_tds / receptions, cutoff 0.104

The control is the whole game. TD-spike tight ends score more, of course they do, touchdowns are points. The useful question is not "do they come back to earth," it is "do they come back to earth faster than a tight end with the same target share who never spiked." Hold the role fixed and the only thing left moving is the touchdown luck.

How I beat on it

scripts/h58_receiving_te_td_mirage.py builds Y-to-Y+1 pairs from player_stats_season.csv (REG, TE, 2015 through 2024 Y0 with Y+1 reaching 2025, at least six games and 40 targets so the receiving role is real), keeps only the receiving tight ends at a 12 percent target share or higher, and bands them by that target share. n = 183 paired seasons, 46 of them top-quartile TD-rate spikes. Scored under half-PPR with the tight end reception premium baked in, 1.0 per catch. Full framework: the Session 97 two-gate spec.

What the data actually said

First swing, the takesmiths look dead right. The TD-spike tight ends averaged 11.4 points a game in their spike year and 9.94 the next, giving back 1.46 points a week. The non-spike receiving tight ends averaged 10.44 and held at 10.31, surrendering a rounding error. So the touchdown regression is real, and it is large. Sell the breakout, the wrench is already out.

Because aggregates lie, and this one is lying about the part that matters. The spike cohort regressed off its own inflated line, yes, but the question for your roster is where it landed, and it landed right on top of its peers. Hold target share fixed:

Every band a wash. The alpha tier, the tight ends actually worth a roster spot, splits by a quarter of a point. The deltas flip sign as you walk down the table, minus here, plus there, which is the signature of noise, not signal. Aggregate, controlling for role, the spike cohort sits 0.37 points a week below its non-spike peers. No band clears the 1.5-point bar with ten players on each side. Gate 1 fails. The headline is that touchdown regression for receiving tight ends is real off the player's own line and worth almost nothing against the field, because target share, the thing the engine already reads, was carrying the production the whole time. The touchdown was the costume.

What the engine already figured out

The engine would not have been fooled by this particular trap. SignalTuned encodes no forward touchdown-rate bump for tight ends; the only production-regression machinery it owns is running-back-only. It ranks tight ends off recency-weighted volume, which is exactly the variable that just neutralized the spike. So on the narrow question, the takesmiths and the engine were holding hands: the TD spike is not a durable edge, and neither of them bet on it.

Which leaves the uncomfortable part. My error report still says the engine over-rates receiving tight ends, and this test just ruled out the obvious within-position reason. At the same target share these tight ends do not bust relative to other tight ends. So the over-rating is not a tight-end-projection error, it is a cross-positional one: the engine is slotting receiving tight ends above the running backs and receivers they should sit behind once you score everyone on the same realized-value scale. That is the third straight archetype my own audit flagged, after pocket-passer quarterbacks and the big-fast backs, where the "error" turned out to live in board placement, not in the position's own curve. A pattern that big is now its own to-do.

What to do about it

For your roster: do not pay the breakout tax on a tight end coming off a touchdown-stuffed year, the scores will regress. But do not fade him relative to other tight ends either, because his target share, not his end-zone luck, is what sets next year's floor. Same rule, both directions: price the volume, ignore the touchdowns. The breakout-TE tier and the TD-regression tier are arguing about a variable that washes out; the one that does not wash out is whether the targets are real and sticky.

Engine-side, this one ships no change. Inventing a tight-end touchdown penalty to match the discourse would make the engine worse, because the penalty does nothing once volume is in the model, which it already is. The real work order, the cross-positional placement of receiving tight ends versus the running back and receiver tiers, is filed to the warm loop where a board-level question belongs, not patched with an archetype knob that would only move guys past their 32nd birthday. Negative result, filed and published. The trap that fooled the experts on both sides: they were debating the touchdown column when the target column already had the answer.

← Back to all posts