Caleb Williams: The Year Two Leap Accuracy Didn't Build

This one is a request from one of our Discord users, so let me put my cards on the table before we start: I am not a Caleb Williams believer. Never really have been. But I'm also a data guy, and the whole point of keeping an EPA table is that it does not care what I think. So let's look at the numbers and see if they talk me out of it.

Some background. Last week I ran every quarterback season since 1999 through that table to answer one question about Drake Maye: what are the odds he does it again? (That post is here.) What stuck with me is that Maye's breakout was built on the stickiest skill in football, which is why I came out more optimistic than the raw regression math said I should be.

The Discord question was the obvious follow-up: what about the other guy from that draft class?

Caleb Williams also made a year-two leap in 2025. The Bears went from 5-12 to 11-6. His EPA/play went from 33rd in the league to 19th. On the surface it's the same story one tier down.

Except it's not the same story. It's almost the opposite one. Maye improved by getting more accurate than anybody in the history of the stat. Caleb improved while getting less accurate, and once you see where his gains came from, 2026 gets a lot more interesting than "he's trending up."

Quick disclaimer

Same as last time: this is historical base-rate work, not a Caleb Williams projection. I know nothing here about 2026 roster changes, coaching, or health. Some of the comparison groups below are small, and I'll give you the sample size every time instead of burying it. Where it's 8 guys, treat it as a hint, not a finding.

Definitions, so we are all on the same page

Repeating these from the Maye post so this one stands on its own.

• EPA - Expected Points Added. Every snap has a starting expected point value based on down, distance, field position, and clock, and an ending one. The difference is the EPA.
• EPA/play - a quarterback's average EPA across his dropbacks (attempts, sacks, scrambles) plus his designed runs. Kneels and spikes thrown out.
• CPOE - Completion Percentage Over Expected. nflverse models each throw's completion probability from depth, location, and situation. CPOE is how far past that the quarterback actually finished. It's the cleanest accuracy stat we have.
• Qualified - 200+ dropbacks in a regular season. 983 quarterback seasons since 1999.
• z-score - standard deviations above the qualified passers in that same season, so eras compare fairly.

What Caleb actually did in 2025

Counting stats: 3,942 yards, 27 touchdowns, 7 interceptions, and 24 sacks taken after 68 the year before. He added 77 carries for 388 yards and 3 scores, and finished at 18.6 fantasy points per game, QB7 on the season, up from 15.0 as a rookie.

Commentary

• The headline improvement is real. Going from 33rd to 19th, and from clearly below average to a little above it, is a real step. I'm not going to try to talk you out of it.
• But look at the completion percentage line and then the CPOE line. He completed 55.4% of his throws, which is the 16th percentile among qualified passers. And his CPOE went the wrong way, from -0.1 as a rookie to -3.5. He was throwing 3.5 percentage points worse than an average quarterback would have on the same throws.
• The sack rate is the whole ballgame. 10.6% to 4.0% is an enormous move, from one of the worst marks in football to the 8th percentile, which is elite. He took 44 fewer sacks. That alone is worth a mountain of EPA, because a sack is one of the most negative outcomes in the sport.
• His interception rate of 1.17% is the 8th percentile, meaning almost nobody threw picks less often. Hold that thought.
• He was pushing it downfield: 76th percentile in air yards, 92nd percentile in deep-throw rate, and his explosive-play rate jumped from 6.3% to 9.7%. When he connected, it went a long way.
• One thing that surprised me and cuts against his reputation: his designed runs actively hurt him. On 23 designed carries in 2025 he was worth -14.9 EPA, about -0.65 per carry. That's why his EPA per dropback (0.108) is higher than his EPA per play (0.082). Small sample and it's mostly short-yardage stuff, so don't overweight it, but the "he makes up for it with his legs" version of Caleb is not in this data. The scrambles are already counted in the dropback number.
• Team context: Chicago went 5-12 with a 310-370 point differential to 11-6 at 441-415. File this away for later: a six-win jump does not clear the 8-win threshold I used in the Maye post, and that point differential is a lot shakier than the record looks.

Where the gains actually came from, and whether they stick

This is the part I actually care about. In the Maye post I ran year-over-year correlations on every qualified quarterback pair from 2006 to 2025 to sort skill from noise. Here's that same table with Caleb's 2025 percentiles dropped in:

Commentary

• Sort his 2025 season into two piles and it splits almost perfectly. The good stuff that sticks: the sack rate and the downfield aggression. The good stuff that doesn't: a 1.17% interception rate, which is the least repeatable number in football at r = 0.21. The bad stuff that sticks: accuracy.
• That's an unusual profile, and it's the mirror image of Maye. Maye's gains sat almost entirely in the sticky column. Caleb's are split, with his single worst trait sitting in the sticky column.
• So 2026 looks like this: keep most of the sack improvement, give back the interception luck, stay a below-average accuracy quarterback. Those pull against each other, which is why his projection below barely budges.

Does a sack-rate collapse like that hold?

Sack rate carries an r of 0.478, so it's about as repeatable as CPOE. But Caleb's move was extreme, so I looked for the extreme cases specifically: quarterbacks who cut their sack rate by 4 or more percentage points from one qualified season to the next, then checked the year after that.

There have been 8 since 2006. As a group: 9.1% in year one, 4.1% in year two, 6.3% in year three. Four of the eight held most of the gain, meaning they were still at least 2 points below where they started.

So the pattern is keep about half. Eight quarterbacks is a thin sample and I wouldn't bet a paycheck on it, but "some giveback, not a full reversal" matches both the small sample and the r = 0.48 correlation, and when two different cuts of the data agree I'm more willing to believe it.

The accuracy problem, and the Josh Allen question

This is what decides his ceiling. A -3.5 CPOE is not a rounding error, and CPOE sticks. So the question is whether quarterbacks with bad accuracy ever fix it.

Since 2006, 93 different quarterbacks have posted a qualified season with a CPOE of -3 or worse. Of those:

• 51% (47 of 93) later posted a season with a CPOE of 0 or better. So roughly half do climb back to average accuracy at some point.
• 19% (18 of 93) ever posted an elite season, a z of 1.0 or better.

19% is the number I keep chewing on. It's not zero and it's not a coin flip. Here's everybody who pulled it off, because the names matter:

Lamar Jackson, Josh Allen, Jared Goff, Nick Foles, Andy Dalton, Carson Wentz, Ryan Tannehill, Michael Vick, Matthew Stafford, Alex Smith, Brett Favre, Colin Kaepernick, Deshaun Watson, Kirk Cousins, Cam Newton, Donovan McNabb, Baker Mayfield, Ryan Fitzpatrick.

Commentary

• Two MVPs on that list and a pile of long-term starters. Bad early accuracy does not disqualify you.
• But 18 out of 93 also means the other 75 didn't get there, and you've never heard of most of them for exactly that reason.
• The Josh Allen arc is the one Bears fans are going to bring up, and honestly it's the right comp, so here it is:

• Allen went from a -6.9 CPOE as a rookie to +6.6 in year three, a 13.5 point swing, and never looked back. It's the best argument out there that accuracy is coachable.
• It's also a 1-in-5 outcome, per the base rate above. Caleb is starting year three from roughly where Allen started his. That's not the same as saying he gets Allen's year three.
• One thing I didn't expect: Caleb is behind Allen's year-two accuracy, -3.5 against -1.6, while being ahead on EPA, 0.082 against 0.030.

The comp set: year-two quarterbacks with bad accuracy

Here's the group Caleb actually belongs to. Every quarterback since 2006 who put up a negative CPOE in his second season, and what he did in year three. I've pulled the top of the list by EPA:

Across the full group of 26 who had a qualified year three:

• Mean z went from -0.34 to -0.39. As a class, flat.
• 42% improved. 27% reached a z of 0.5 or better.
• Mean CPOE moved from -2.6 to -1.1, so accuracy does drift back toward average, just not all the way.

Commentary

• Caleb sits near the top of this group by EPA and near the bottom by CPOE. That's his whole profile in one line.
• The upside cases are Allen and Mayfield, and to a lesser degree Stroud bouncing back last year. The downside cases are Bortles, Sanchez, Fields, and RG3, and the common thread is that none of them ever fixed the accuracy.
• If you take one thing from this table: year three was decided almost entirely by whether the CPOE moved. Every guy who jumped also fixed his accuracy. Every guy who stalled didn't.

The 2024 draft class, two years in

Since we're here, the whole class in one table:

Commentary

• Two years in, the class ordering has completely inverted. The guy who was 27th as a rookie is now 1st. The guy who was 5th is now 28th, on only 243 dropbacks.
• Daniels is the cautionary tale I flagged in the Maye post, and it played out: a 1.07 rookie z, then -0.53 in year two.
• Nix is the quiet one. Nearly identical in both seasons, 15th then 14th. Boring and stable is a real outcome, and it's underrated.
• Caleb is third in this class by 2025 EPA and second-to-last by CPOE among the guys with real snaps.

So what are the odds?

Same regression line as the Maye post, fit on all 667 qualified year-over-year pairs:

next_z = 0.100 + 0.417 x this_z     (residual SD = 0.83)

Caleb's 2025 z of 0.15 goes in, and a projected z of +0.16 comes out, or about 0.083 EPA/play. Right about where he already is. Around that projection:

Commentary

• Bears fans are not going to like this part. The regression line doesn't project Caleb forward, it projects him flat, because a z of 0.15 is barely distinguishable from average and there's nothing in the math that knows he's 24 years old and improving.
• That's where this approach falls down, and I don't have a great fix for it. The math treats a 0.15 from a second-year quarterback exactly like a 0.15 from a tenth-year journeyman. The comp tables above are my attempt to patch around it, and they read a little better: 42% of his comp group improved and 27% got to a z of 0.5.
• A 16% shot at a top-five season is not nothing. It lands right on top of the 19% rate for bad-CPOE quarterbacks ever going elite, and I trust a number a lot more when I get to it two different ways.
• Now the part where I check myself. I told you up top I'm not a believer, and then the data came back agreeing with me, which is exactly when I should be most suspicious of my own read. So here is the strongest version of the other side: he is 24, he cut his sack rate by two thirds in a single year, he throws deep more than almost anybody, and half the bad-accuracy guys in that 93-quarterback group did climb back to average. None of that is nothing.

Take Aways

1. Caleb's leap was real but narrow. He went from clearly bad to a little above average, and almost all of it came from not taking sacks. 10.6% to 4.0% is his whole season.
2. Watch the accuracy and basically nothing else. In his comp group, year three came down to whether the CPOE moved. At -3.5 on a stat that carries an r of 0.48, that's his ceiling until he moves it.
3. Expect some sack giveback. The 8 comparable sack-rate collapses kept about half the gain. If he lands around 5-6% instead of 4%, that's a normal outcome, not a red flag.
4. The interception rate is a mirage. 1.17% at the 8th percentile, on the least sticky stat in football. That's going up, and it will cost him EPA even if he plays exactly as well.
5. In fantasy I like him more than this post sounds. He was QB7 at 18.6 points per game, and fantasy scoring rewards volume, deep shots, and rushing floor in ways EPA/play doesn't. His 92nd percentile deep rate is a fantasy asset even when it's an efficiency liability. I'd draft him as a low-end QB1 and feel fine. I just wouldn't pay for a leap.
6. Against Maye, it isn't close. Same class, same year-two leap narrative, wildly different foundations. Maye's gains sat in the stickiest column on the board. Caleb's best trait is sticky and his worst trait is too.
7. On the Bears' win total, I lean under 11. Six extra wins on a +26 point differential is a team outrunning its underlying numbers, and that usually corrects.

If Caleb comes out in 2026 completing passes at even a league-average clip over expectation, tear this post up. That's the tell, and it will show up in the CPOE column weeks before it shows up in the record. It's exactly what Josh Allen did in year three, and it's the one number I'll be watching.

Thanks for reading, and if you want to see our top player projections each week, create a free account by clicking this link.