What Is DEION?
DEION (Dividing Estimator for Impact Of Non-QBs) is Held Out's rating for every player who isn't a quarterback: receivers, tight ends, backs, pass rushers, interior linemen, linebackers, and defensive backs.
A box score tells you a receiver caught 60 balls. It doesn't tell you what those catches did for his offense. It tells you a linebacker made 110 tackles, and not how many of them stopped a drive. DEION puts one grade on each player, so two receivers with 60 catches each can be compared on what their catches did.
Non-quarterbacks are much harder to rate than quarterbacks. A quarterback touches nearly every offensive play. A corner might be targeted four times a game, and an interior lineman's best plays mostly leave no trace in the box score. DEION is my best try at the problem, and it says which positions it can rate and which it can't (see Which Positions Are Shown).
The Scale
Grades are on the same scale as BRADY and RICE. A 75 is the median starter at that player's position that season. Each position has its own pinned anchors, so a 90 at edge rusher and a 90 at tight end are equally rare. The scale is a straight line, with no curve squeezing the top toward 99 and no ceiling. A season far enough above the rest of its position reads above 100.
Why DEION, Instead of ___?
1. It's Built on BRADY
DEION runs on BRADY's engine. It rates each player game by game, adjusts each stat for the opponent, and weighs his history over BRADY's ladder of memory lengths (four to thirty-two games). It also measures each player against his own position. A back's carries are judged against other backs, and a thin record is pulled toward what players at his position with similar experience usually look like. The machinery that makes BRADY stable from game to game and from season to season does the same job here.
Measuring each player against his position matters more than it sounds. When backs and receivers were pooled, the gap between the two positions read as talent. DEION trusted a back's per-carry value after 315 carries, where backs alone say it takes about 734. Part-time backs and slot corners with a few big games floated to the top.
2. It's Trained on Football, Not on Awards
The stats don't all count the same, and I don't pick the weights. Each PIV-based stat counts at face value. The rest (tackles, pressures, coverage, targets, and the other box-score stats) get separate weights for each position. Those weights are fitted to one question: when these players were on the field, how did their unit do in the next game? For an offense, the answer is its points per play beyond what its quarterback's BRADY and the opposing defense explain. For a defense, it's what the defense gave up beyond what the opposing quarterback and offense explain.
No award trains it. The All-Pro teams are a useful check on the ratings (see How Well It Does). If DEION were fitted to them, it would only be a complicated way of predicting voters.
Each weight is held to the sign that makes football sense. More volume never counts against a player. Getting targeted or missing a tackle never counts for a defender. Without those limits, the fit put blocking tight ends and fullbacks at the top of the offense, which is a good reminder of why a person has to read the output.
3. Context
A receiver's targets are judged net of how his team's other receivers do on theirs. A receiver catching passes from a great quarterback doesn't get the quarterback's credit. A running back's rushing is rated beside his team's season, so the back keeps only the part of a run that follows him from team to team. How much of a run belongs to the back is one of the least stable things in football. Every stat is also adjusted for the opponent, so a corner who spends a season on the league's best receivers gets credit for it.
What Goes In
- Receivers, tight ends, and backs: targets, catches, yards, PIV, touchdowns, and first downs, plus yards after catch and air yards per target; carries with their yards, PIV, touchdowns, first downs, and fumbles. Volume is counted per play his offense ran in the games he played. A back who comes in only to carry the ball isn't projected to a full-time role he never gets.
- Defenders: sacks, quarterback hits, pressures, stops, tackles, deep tackles, tackles for loss, interceptions, passes defended, forced fumbles, missed tackles, and, from 2018, coverage stats from Pro Football Reference's charting.
Which Positions Are Shown
A position appears on the site only if its grades predicted what its unit did, on seasons before the ones graded. If a position's grades don't help forecast its unit, I don't show them, however plausible they look.
- Receivers, tight ends, running backs, edge rushers, and interior defensive linemen are shown from 2003.
- Linebackers, cornerbacks, and safeties are shown from 2018, when the coverage charting starts. Without coverage, their grades are mostly guesswork.
- Offensive linemen and fullbacks aren't shown. Linemen leave almost nothing in the box score, and nothing I tried rated them well enough to publish.
An earlier version of the estimator set these cutoffs, so they might move when I run the tests again.
How Well It Does
The main test looks back at games already played. For each game, it takes every player's grade from before that week and weights it by the share of his unit's snaps he played. Then it asks how much of the unit's remaining error those grades explain, beyond the quarterback's BRADY. The test uses the players who took the field, backups included, so it grades DEION and not a guess about the lineup.
| Unit | 2003-19 | 2020-25 |
|---|---|---|
| Offense | 0.46% | 0.35% |
| Defense | 2.64% | 3.89% |
| Both | 3.21% | 4.36% |
The version before November 2026 read slightly higher on this test: 3.49 percent and 4.65 percent for both units. The November changes measured each player against his position and counted volume per team play. They were neutral within noise, and they fixed rankings that were wrong on sight. A backup back in 2026 read 2nd in the league on two full games, and he reads 19th now.
Those percentages are small. The team rating and the quarterback already explain most of what a unit does next week. What's left is the part individual players can claim, and it's a small slice of a noisy game.
The second check is against the All-Pro teams. Voters aren't a perfect measure, but a rating that never put All-Pros near the top of their positions would be reading the wrong things. When DEION disagrees with the voters, ask which of the two is measuring the player.
Where It Looks Wrong
A few results still look off to me, and I'd rather name them than hide them:
- Randy Moss's best season reads 16th among receivers' seasons, and it's 2003, not his record-setting 2007. DEION can't yet credit a receiver for being the reason his quarterback throws elsewhere.
- The box score barely records coverage before 2018. That gap is why linebackers, corners, and safeties aren't shown before then.
Records and Honors
The Historical tab ranks each position's best seasons by the grade each player closed the season with. It ranks careers by career value. For each season with at least six full games of snaps, career value takes the grade points above the median starter, scales them by how much of the season he played, and adds them up. A season in progress doesn't count toward a peak until it's over.
The honors shown are AP first-team and second-team All-Pro, Offensive and Defensive Player of the Year, Offensive and Defensive Rookie of the Year, and Comeback Player of the Year. They're matched by player ID, never by name. No honor changes a grade.
What I Tried and Dropped
- Blending the raw box score into the play-value grade. It made the defense worse at every weight I tried.
- Volume on a log scale, a square-root scale, or per play. All three came out within noise of a plain count, so the plain count stayed.
- Updating each player's grade from his unit's result after every game. It looked like a big win on the offense test (+2.3 percent), and the gain turned out to be a flaw in how I was testing it. Deleted.
Notes
- DEION doesn't feed any forecast. RICE has its own estimates of who's playing and what a missing player costs (see Who Will Play). Changing DEION never moves a RICE number, and RICE never moves a DEION grade.
- Each season is graded only with weights fitted on earlier seasons. A season with too few earlier seasons to fit its weights isn't shown.
- Every grade is checked for leaks: rating past seasons with and without the future in the data has to give the same answer.