---
title: "Touchdowns are mostly rented, not owned"
slug: touchdown-regression
dek: "A walk-forward test of the fantasy maxim — and a more useful way to apply it."
access: free
email_subject: "The 3.0 touchdowns a hot season gave back"
email_preview: "Our first study puts a number on touchdown regression."
candidate_generated_at: 2026-08-31T01:15:20.866111+00:00
published_at: null
data_through: 2025
---

# Touchdowns are mostly rented, not owned

Fantasy analysis says “touchdown regression” so often that it can become a substitute for analysis. We wanted a falsifiable version: **does an opportunity-based expected-touchdown baseline predict next season better than a player's raw touchdown total?**

Across **649 consecutive WR/TE season pairs**, it did. Carrying forward actual touchdowns produced **2.65 TD** of mean absolute error. Carrying forward expected touchdowns cut that to **2.34 TD** — an **11.7% reduction**.

![Forecast error comparison](figures/forecast-error.png)

## The result in one picture

Players in the highest overperformance quintile scored **8.3 touchdowns** in year one and **5.3** the next year: a decline of **3.0**. The biggest underperformers moved the other way, from **2.9** to **4.4**.

![Regression quintiles](figures/regression-quintiles.png)

That pattern is the useful core of regression. It is not “high touchdown players are bad.” It is “the part of touchdown production unexplained by repeatable opportunity should receive less weight.”

## What we modeled

We aggregated nflverse regular-season receiving data for WRs and TEs with at least 40 targets from 2014–2025. For every test season from 2018 onward, the model was trained only on earlier seasons. Five same-season opportunity signals — targets, receptions, yards, air yards, and receiving first downs — estimated a touchdown baseline. Ridge regularization reduced instability from correlated volume measures.

The next-season test compared two deliberately simple forecasts against the same outcome:

1. next season equals this season's actual touchdowns; and
2. next season equals this season's expected touchdowns.

The paired MAE improvement was **0.31 touchdowns**, with a 95% paired-bootstrap interval of **0.18 to 0.45**. Those 5,000 resamples quantify uncertainty inside this one evaluation sample; they are not independent replications. Both forecasts still missed by more than two touchdowns on average, so this is a modest benchmark improvement rather than a complete projection model.

## How to use it

At the population level, large touchdown-over-expected gaps can support a weak tiebreaker—not an individual projection adjustment with a known magnitude. If two players have similar projected opportunity, the evidence supports giving less weight to the one whose prior production depended more on outlier conversion. Do not mechanically fade elite players with elite volume, and do not turn a historical baseline into a current projection without team, quarterback, health, and market context.

## Where this can break

The test includes only players who returned to at least 40 targets, so it describes volume survivors. Selection can affect both tails differently: high scorers whose roles collapse disappear from evaluation, while low scorers whose roles grow are more likely to enter it, and the net direction of bias in the forecast comparison is uncertain. Season totals also hide red-zone usage and route context. The model is a benchmark, not a causal estimate of touchdown “skill.” Finally, the public source snapshot ends with 2025; this is not a current player-ranking model.

## Reproducibility and provenance

Sources: nflverse `player_stats.csv` through 2024 plus `stats_player_reg_2025.csv`, both from the project's CC BY 4.0 GitHub releases. The run manifest records each URL and SHA-256 alongside the code path, seed, model coefficients, walk-forward predictions, bootstrap outputs, reviewer findings, and publication decision. Underlying NFL data remain subject to their owners' terms.