Skip to main content
ABC News
The NFL Got Weird All Of A Sudden

If you thought you had a handle on the 2019 NFL season through its first three weeks, well … Week 4 happened. The Bucs dropped 55 on the defending NFC champion Rams; the Panthers, Raiders, Browns and Titans all scored road upsets; the Lions even came close to beating Patrick Mahomes and the mighty Chiefs. Just as the pack came roaring past the quarter-pole of the Super Bowl race, the horses kicked up a bunch of dirt — and now we can’t see the track anymore.1

The first three weeks of the season had, in fact, been comparatively normal. The Brier score between actual game results and pregame win probabilities (according to our Elo model, which assigns each team a power rating and adjusts for quarterback injuries) was a modest 0.209. That ranked 14th-lowest — lower means smaller errors — among seasons since the 1970 AFL merger in terms of accuracy through Week 3. Sure, there were some unexpected breakouts (hello, Bills and Lions!) and disappointments (Eagles, Browns, etc.), and starting quarterbacks were dropping like flies. But even through all the QB chaos, the NFL hadn’t started a season this predictably in years:

This NFL season was starting out predictably …

Most predictable first three weeks of an NFL season according to the Brier score between actual game results and those predicted by the FiveThirtyEight Elo model’s pregame odds, 1970-2019

Season Games Correct Picks % Correct Brier Score
1973 39 28.0 71.8% 0.171
1998 43 33.0 76.7 0.172
1977 42 31.0 73.8 0.173
1996 43 33.0 76.7 0.176
1976 42 30.5 72.6 0.182
1995 45 33.0 73.3 0.186
1979 42 30.0 71.4 0.187
1975 39 30.0 76.9 0.190
1993 38 28.0 73.7 0.199
1992 41 30.0 73.2 0.201
2013 48 38.0 79.2 0.201
1986 42 30.0 71.4 0.202
1984 42 26.0 61.9 0.207
2019 48 31.5 65.6 0.209

The Brier score is the average squared difference between a probabilistic forecast and the actual outcome. A game pick counts as “correct” if the pregame favorite won; ties are considered half-correct.

Source: ESPN,

But this past weekend changed all that. Week 4’s Brier score of 0.310 made it — coincidentally enough — the 14th-least predictable individual week of any NFL season since 1970:

… before it all went off the rails

Most unpredictable weeks of an NFL season according to the Brier score between actual game results and those predicted by the FiveThirtyEight Elo model’s pregame odds, 1970-2019

Season Week Games Correct Picks % Correct Brier Score
2015 10 14 5 35.7% 0.360
1978 8 14 4 28.6 0.359
2001 6 13 4 30.8 0.354
1991 13 14 7 50.0 0.343
1995 8 13 4 30.8 0.338
1983 1 14 4 28.6 0.328
1990 12 14 6 42.9 0.321
1994 17 14 4 28.6 0.319
2002 2 16 4 25.0 0.319
2017 6 14 6 42.9 0.316
1992 9 12 6 50.0 0.312
1981 8 14 7 50.0 0.311
1980 1 14 6 42.9 0.311
2019 4 15 7 46.7 0.310

The Brier score is the average squared difference between a probabilistic forecast and the actual outcome. A game pick counts as “correct” if the pregame favorite won; ties are considered half-correct.

Source: ESPN,

Here’s a rundown of the games that contributed most to the off-target predictions — and what information we might legitimately glean from them:

  • Tampa Bay over L.A. Rams: The biggest miss for the model was easily the Bucs’ aforementioned 55-40 victory over the Rams on Sunday afternoon, in which we gave them only a 17 percent chance to win. It’s easy to read too much into such an aberrant result, in both directions. For L.A., the defensive numbers are eye-catchingly bad: In addition to the 55 points allowed, their -19.0 defensive expected points added (EPA) represented the sixth-worst game by any defense so far this year. But on the season, L.A.’s defense has still been roughly as good (11th in the league by EPA) as it was last year (13th). The larger takeaway from the game might be that the Bucs’ offense is for real. The team now ranks third in offensive EPA per game after adjusting for opponent quality and second (granted, far behind the Chiefs) in adjusted EPA specifically via the pass.
  • Carolina over Houston: Elo gave Kyle Allen and the Panthers just a 22 percent chance of going into Houston and knocking off the Texans, but instead they never trailed en route to the 16-10 win. Both of these teams are difficult to figure out: Carolina has been much better (+9.3 total adjusted EPA per game) in the two games Allen started than in the two games started by the sixth-highest-paid quarterback in the game, Cam Newton (-2.8 EPA per game). And the Panthers defense gave a lot of trouble to Deshaun Watston, forcing him into his second terrible start in his past three games. Watson now ranks just 10th in our Elo QB ratings after ranking sixth going into the season, and the Texans (with a 29 percent chance of winning the AFC South) are just about the most tenuous division favorites possible.
  • Oakland over Indianapolis: Our model had the Colts as heavy home favorites (73 percent) over the Raiders — not unreasonable, since Indy’s Jacoby Brissett played well in a Week 3 win over the Falcons while Oakland’s defense looked horrible against the Vikings. But lo and behold, the Raider D showed up, QB Derek Carr played well enough to win, and Oakland earned the 31-24 upset. So what did we learn about both teams? Maybe just that they’re deeply mediocre, along with many other squads this year. Including Indy and Oakland, 13 teams currently have a record of 2-2, which is tied with the 2010, 2014 and 2017 seasons for the most 2-2 teams through a Week 4 since 1970.
  • Cleveland over Baltimore: After divergent starts to the season — Lamar Jackson and the Ravens looked surprisingly great, Baker Mayfield and the Browns looked disappointingly meh — it wasn’t shocking to see Baltimore installed as 72 percent favorites over Cleveland. What was shocking, however, was how well the Browns played on both sides of the ball in their 40-25 win. The Ravens offense was held below their usual EPA output, while Mayfield broke out of his slump as the Browns offense had its best game of the season by far. Mayfield was streaky last season as well, so he could be turning a corner, but the bigger concern is probably with Baltimore’s defense. A year removed from ranking fourth-best in EPA, it ranks sixth-worst this season after suffering the fourth-worst defensive EPA game of the season (-20.3 EPA) thus far.

And that doesn’t even get into the other weirdness of the week, which included Detroit’s near-upset of Kansas City (a result that ended up helping Week 4’s Brier score because K.C. had a 67 percent chance of winning). All told, Week 4 proved once again that we never know as much about the NFL as we think we do — and it probably won’t be the last time we have to relearn that lesson before the season is over.

Looking Ahead: Week 5

Best matchup: No. 3 L.A. Rams at No. 7 Seattle (-1.5), 8:20 p.m. ET Thursday
Matchup quality: 89th percentile2
Matchup evenness: 83rd percentile

We’ve written before that Thursday Night Football is often used as a dumping ground for the NFL’s ugliest, most unwanted matchups. But that hasn’t really been true so far this season, as the TNF slate has also contained a surprising number of marquee games, at least on paper.3 And in no week is that more true than Week 5, when the Rams and Seahawks find themselves on a weeknight collision course. Los Angeles is fresh off that stunning loss to the Bucs (see above), while Seattle is just a few weeks removed from a perplexing home loss to the Saints and backup QB Teddy Bridgewater. Still, these are two of the best teams in football according to Elo, and our ratings give the Seahawks a slim 55 percent edge to come out on top.

See also: Green Bay at Dallas (87th/53rd); Cleveland at San Francisco (67th/50th)

Biggest playoff implications: Rams at Seahawks
Potential shift in playoff odds: 30.1 total percentage points

For the first time so far this season, the best/most evenly matched game of the week is also the most important game of the week, in terms of its effect on team playoff probability: Rams versus Seahawks. Both teams are 3-1 at the moment, but Seattle has faced only one NFC West opponent (Arizona, last week), and L.A. hasn’t played any until now. With three teams — the Rams, Seahawks and San Francisco 49ers — bunched between a 25 and 40 percent chance of winning the division, each of these games matters a lot. The Rams currently have better playoff odds than Seattle (68 percent to 49 percent), but that number would change by +/-14.5 percentage points depending on the result of this game. The stakes are even higher for the Seahawks, whose playoff chances would rise to 63 percent with a win or fall to 32 percent with a loss. Although teams with a 3-2 record aren’t doomed, historically speaking, the NFC West is currently tracking for three teams with a 10-6 record or better, so the path to the playoffs for both Seattle and Los Angeles is complicated by the strength of their division.

See also: Tampa Bay at New Orleans (26.9); Buffalo at Tennessee (25.7)

Best QB duel: No. 3 Dak Prescott (DAL) vs. No. 6 Aaron Rodgers (GB)

See also: No. 2 Matt Ryan (ATL) vs. No. 10 Deshaun Watson (HOU); No. 1 Patrick Mahomes II (KC) vs. No. 18 Jacoby Brissett (IND)

FiveThirtyEight vs. the Readers

As a weekly tradition here at FiveThirtyEight, we look at how our Elo model did against everybody who made picks in our forecasting game. (If you entered, you can find yourself on our leaderboard here. I am currently in 1,127th place!) These are the games in which Elo made its best — and worst — predictions against the field last week:

Elo’s dumbest (and smartest) picks of Week 4

Average difference between points won by readers and by Elo in Week 4 matchups in FiveThirtyEight’s NFL prediction game

DEN 60% JAX 51% JAX 26, DEN 24 +9.9
HOU 78 HOU 71 CAR 16, HOU 10 +7.9
KC 67 KC 77 KC 34, DET 30 +3.3
IND 73 IND 70 OAK 31, IND 24 +2.6
NE 65 NE 73 NE 16, BUF 10 +2.4
ATL 62 ATL 59 TEN 24, ATL 10 +1.9
LAC 79 LAC 87 LAC 30, MIA 10 +1.1
NYG 56 NYG 58 NYG 24, WSH 3 +0.0
BAL 72 BAL 71 CLE 40, BAL 25 -0.6
SEA 71 SEA 71 SEA 27, ARI 10 -2.1
LAR 83 LAR 84 TB 55, LAR 40 -2.9
PIT 58 PIT 57 PIT 27, CIN 3 -3.1
GB 61 GB 63 PHI 34, GB 27 -4.4
DAL 53 DAL 56 NO 12, DAL 10 -5.3
CHI 59 CHI 52 CHI 16, MIN 6 -8.0

Home teams are in bold.

The scoring system is nonlinear, so readers’ average points don’t necessarily match the number of points that would be given to the average reader prediction.

After snagging a win over Elo in Week 3, how did the average FiveThirtyEight reader do in the chaos of Week 4? The algorithm picked up some points on the field with its confidence in the Bears, Saints and Steelers, but the readers bought into Minshew Mania and correctly picked Jacksonville to win at Denver, to go with some savvy sliding of the probability sliders in the Carolina-Houston, K.C.-Detroit, Oakland-Indy and New England-Buffalo matchups. That was enough to net the readers their second victory of the season — already more wins than they had all of last season! — by an average margin of 2.7 points.

Congrats to Caleb Nides, who led all readers in Week 4 with 119.5 points, and to Jon Radermacher, whose total of 371.4 points leads the full-season contest. Thanks to everyone who played — and if you haven’t, be sure to get in on the action! You can make picks now and try your luck against Elo, even if you missed Week 4.

Check out our latest NFL predictions.


  1. Yep, that’s what a tortured horse-racing metaphor looks like.

  2. In terms of the harmonic mean of both teams’ QB-adjusted Elo ratings, relative to that figure for all regular-season NFL games this year.

  3. Of the five Thursday night matchups so far (if you include the season-opening special), two have been really good pairings — Packers-Eagles and Seahawks-Rams. One — Packers-Bears — looked solid on paper beforehand, even if it proved to be unwatchable trash once the game actually got underway. That leaves only two matchups — Carolina-Tampa Bay in Week 2 and Tennessee-Jacksonville in Week 3 — that fit the usual Thursday night template of pure mediocrity.

Neil Paine was the acting sports editor at FiveThirtyEight.