2026 changes to ELWAY and QBERT
ELWAY was an ambitious model when we launched it last year.
asserted
we → launch → it
It already incorporated a number of layers: rolling team ratings dating back to the dawn of the NFL, gradually adding in advanced stats as they became available; quarterbacks in the form of QBERT; off-season roster movement; and a wide array of adjustments for factors such as injuries and home-field advantage.
asserted
they → incorporate → injuries
But ELWAY was a bit like a fancy new house that had never really been lived in.
asserted
that → live → house
It got some of the big headlines right — including becoming an early fan of the Super Bowl Champion Seattle Seahawks.
asserted
some → get → Champion
But because ELWAY had less tenure than other Silver Bulletin models, some parts felt less polished or slightly unfinished.
asserted
parts → have → models
So we put a lot of work into ELWAY and QBERT this year.
asserted
we → put → ELWAY
Most of this doesn’t change its architectural blueprint, but ELWAY’s original intentions are now implemented more effectively, we hope.
asserted
we → change → blueprint
AI tools were quite helpful for performing a thorough audit of its code and uncovering a few inefficiencies and even some bugs.
asserted
tools → perform → inefficiencies
Most of these are on the order of minor nuisances — the toilet in the upstairs bathroom doesn’t always flush — but collectively, we expect this cleanup to considerably improve ELWAY’s performance.1
Apart from this debugging, most of our work this off-season was concentrated in three areas:
A refit of QBERT;
Switching from efficiency per play to efficiency per drive as ELWAY’s main layer for assessing team quality;
And much more detailed handling of injuries.
asserted
most → flush → injuries
If this text sounds familiar, most of it is duplicated from the QBERT landing page as of when we launched it this season.
asserted
we → sound → it
We are quite confident that drops should be attributed to the WR, not the quarterback.
asserted
drops → attribute → WR
Not only does QBERT now give a QB credit for a completion on a drop, but we also estimate the additional yards and touchdowns (and even the avoided interceptions2) that would have resulted from the caught pass.
asserted
that → give → pass
QBs who had fewer drops than expected, like Drake Maye last year, suffer a slight penalty.
asserted
who → have → penalty
Yards gained on pass interference calls deserve credit.
asserted
Yards → gain → credit
Last year’s model ignored penalties.
asserted
model → ignore → penalties
But having collected penalty data, we now simply treat defensive pass interference (DPI) calls as completed passes, giving a QB credit for the completion, the first down, and the air yards associated with the “catch”.
asserted
we → collect → catch
This benefits QBs like Dak Prescott, who had 268 yards essentially erased from his stat line last year on DPIs, by far the highest in the league.
asserted
yards → benefit → league
Furthermore, intentional grounding calls are now treated as incompletions worth minus-10 yards.
asserted
calls → treat → incompletions
Other QB-related penalties are still ignored because we’ve found they don’t provide much signal.
asserted
they → relate → signal
QBERT likes rushing QBs more than the consensus, but not all QB rushes are created equal.
asserted
rushes → like → consensus
We now give far less credit to TDs from sneak plays versus other rushing TDs and also apply a penalty term to sneaks to ensure that they don’t have positive expected value relative to other QB plays.
asserted
they → give → plays
Over the past three years, two QBs have benefited disproportionately from sneaks: Jalen Hurts (30 sneak TDs) and Josh Allen (21).
asserted
QBs → benefit → sneaks
Isolating sneaks does result in slightly more credit to non-sneak rushing TDs, however.
asserted
Isolating → isolate → TDs
Pocket passers were being undervalued.
asserted
passers → undervalue → ?
Last year, the coefficients used in QBERT were fitted on data from 2016-2024.
asserted
coefficients → use → 2016
We’ve now collected advanced-stat data going back to 2006 and refitted QBERT over a larger sample.
asserted
We → collect → sample
This matters because 2016-2024 was a highly favorable era for mobile QBs like Lamar Jackson, Allen and Mahomes, whereas the late 2000s and early ‘10s were more favorable to pocket passers (Brady, Brees, Manning).
asserted
2000s → matter → passers
This makes QBERT more robust across eras and also brings it more in line with the conventional wisdom, helping modern pocket passers like Stafford.
asserted
QBERT → make → Stafford
Touchdowns now matter less, and the process that produces TDs matters more.
asserted
that → matter → TDs
Last year’s version of QBERT was essentially pegged to ESPN QBR.
asserted
version → peg → QBR
QBR is a smart system, much better than traditional NFL passer ratings, and still serves as an important scaffold for our system.
asserted
QBR → serve → system
Our reverse-engineering of QBR suggests that it places an extremely high value on touchdowns, equivalent to something like 45 yards per passing TD.
uncertain
it → suggest → TD
Other systems like ANY/A assign less than half as much credit to TDs.
asserted
systems → assign → TDs
While there is some persistent skill in red zone play, TDs are both noisy and opportunity-dependent, and TD rates don’t “travel” all that well when a QB changes teams.
asserted
QB → be → teams
Long story short: the reward given to TDs has been reduced by 50 percent, with credit instead flowing into categories like first downs that better predict touchdowns in the long run.
asserted
that → give → run
The 2025 version of QBERT accounted for the game script, giving more credit to quarterbacks who establish early leads.
asserted
who → account → leads
The rationale was that offenses are less efficient with a lead late in the game because it’s rational for them to play conservatively and run out the clock.
asserted
them → ’ → clock
However, upon review, essentially all of the difference stems from replacing quarterback plays with handoffs to running backs.
asserted
all → stem → backs
Thus, we’ve removed the game-script variables.
asserted
we → remove → variables
Mostly, they were giving QBs extra credit for being on good teams, undermining QBERT’s goal of separating QB performance from team performance.
asserted
they → give → performance
…and 274 more, not listed.
The Aporia
