Pittsburgh at New Orleans

Pittsburgh is favored by 3 over New Orleans at the Stade de France in Paris, a 9:30 AM ET kickoff, and the model makes it Pittsburgh by 7.0 — a 4.0-point disagreement in favor of laying the points. Where that projection lands matters as much as its size. Three is the most common final margin in football, 14.7% of the 2,054 settled games in the model's training data, and the projection clears it by four points rather than sitting on it.

Neither team is among the league's best, so the case rests on where they differ. Pittsburgh is 17th in opponent-adjusted net EPA at -0.022 and New Orleans is 22nd at -0.088. The offenses are the largest gap: Pittsburgh's ranks 18th at -0.021 and New Orleans's ranks 25th at -0.111, with success rates of 42.4% and 40.5%. The defenses are close, Pittsburgh's 17th at +0.001 and New Orleans's 14th at -0.023, the one unit where New Orleans rates better. Two more things separate them. Pittsburgh's turnover margin of +0.333 ranks 6th, while New Orleans's -0.529 ranks 29th. And Pittsburgh's special teams rank 12th at +0.145, against New Orleans's 31st at -0.030.

Pittsburgh is 4-2, with Aaron Rodgers taking 250 of its 252 dropbacks. Its losses came 20-3 at New England, rated 2nd, and 27-24 at Cleveland; its wins came 20-13 over Atlanta, 30-27 over Cincinnati, 34-22 over Indianapolis and 26-19 at Tampa Bay last week. New Orleans is 2-4 and has not won at home: it is 0-3 at the Superdome, outscored 103-64 by Las Vegas, Atlanta and Minnesota. Its two wins came on the road, 24-17 at Baltimore, rated 6th, and 34-27 at the Giants last week, so the Saints have played their best football away from home.

This is the only path from an injury to the projection, so it is worth showing rather than describing. The model reads the expected starter’s rating:

When a starter is ruled out, the depth chart promotes his backup, so the play probability above reads 100% and the post-model quarterback adjustment does not fire. The absence shows up in the rating instead, which falls to the backup’s — and that rating is one of the twenty features the model selects. A play probability below 100% with no designation filed is the default weighting applied to any unfiled starter, not a report that he is hurt.

Every input is perturbed by the standard error of a full 17-game season average and the game re-projected, four hundred times. The question each trial answers is whether the model still favors PIT -3 after the perturbation.

It does in 100.0% of trials at that realistic noise level, and 96.0% when the noise is doubled.

The other side of this game is priced too — NO +3 — and the model puts it at 47.3% against 51.9% implied by its own price. It is the same projection seen from the other end, not independent evidence — preferring this side and declining that one are one decision, not two.

The side held in 100% of trials and 96.0% under doubled noise. Put plainly: every input to the model was nudged at random by the standard error of a season's average, 400 times over, and Pittsburgh stayed ahead of 3 every time. At twice that noise, 16 trials in 400 flipped. The site works in the same direction. The model has no input for a neutral field and reads New Orleans as the home team, as the schedule lists it; across the 2,054 games in its training data, home teams won by 1.6 points on average. To whatever extent the projection carries that edge for New Orleans, a game in Paris takes it away.

PIT opened at -1.5 and sits at -3: the market has moved 1.5 points toward PIT. 46% of tickets and 46% of the money are on this side. That is reverse line movement: the line moved toward PIT while the majority of tickets were on the other side, which usually means the move came from bettors the public splits do not count.

The money splits match the tickets exactly, so the move did not come from a few large bets on Pittsburgh either. The move also changes the bet itself, because it put the line on the key number. At -1.5, a three-point Pittsburgh win covered. At -3, that same win is a push, and Pittsburgh has to win by four or more to cover.

Public betting splits from Action Network, captured 2026-10-24, across 20,795 tracked bets. They describe the market; they are not an input to the model.

No injury moves this projection directly: the model's only injury input is the quarterback adjustment, and it does not fire, because Rodgers carries no designation and Tyler Shough, who has taken every New Orleans dropback this season, practiced in full with nothing filed. The rest is context, and it leans Pittsburgh. New Orleans is the more depleted side at 1.211 against Pittsburgh's 0.757, and its largest absence is in the backfield: Travis Etienne is on reserve at 46.8% of the running-back snaps. Kelvin Banks is out with an ankle injury at 15.9% of the tackle snaps. Cesar Ruiz is questionable at 18.0% of the guard snaps after missing practice, so two offensive linemen are out or in doubt for an offense that already ranks 25th. Nathan Shepherd is questionable at 28.1% of the interior line snaps after full practices, and Kaden Elliss practiced on a limited basis with nothing filed. Pittsburgh is without Derrick Harmon, on reserve at 17.2% of the interior line snaps, and Roman Wilson, on reserve at 12.0% of the receiver snaps. Sebastian Joseph-Day is out at 13.1% of the defensive end snaps. Rico Dowdle is questionable at 34.3% of the running-back snaps after a limited week. Cameron Heyward missed practice at 30.6% of the interior line snaps but carries no designation on the final report.

Final report, filed Friday. A player ruled out or listed as doubtful counts as not playing; a questionable non-quarterback plays about 62% of the time, a questionable quarterback about 35%. Reserve lists are folded in separately because players on IR never appear on a weekly report.

Depletion is the share of a normal week’s snap workload expected to be missing: each player’s share of his position group’s snaps in prior games, multiplied by his probability of not playing. 1.00 is roughly one full starter’s workload. It is context, not a model input.

The case against starts with the shape. Across 1,391 walk-forward games, laying points at a 3-7 point disagreement has covered 44.1% of the time, below the 52.38% break-even, so a four-point gap is no evidence on its own. Neither team rates in the top half of the league, which makes this a margin estimate between two middling sides, and the line now sits on 3, where a field-goal win pushes. New Orleans also owns the best win in this matchup, at Baltimore, rated 6th; Pittsburgh's best came over Cincinnati, rated 14th, and the Steelers lost 20-3 to the one top-five team they have played. Heyward missed practice without a designation, and Dowdle is questionable.

Play-by-play from every game since 2019 — roughly 345,000 plays — plus snap counts, depth charts and multi-book closing lines. An ensemble of gradient-boosted trees, a regularized linear model and a random forest predicts the point margin directly; that continuous prediction is converted into a probability against any line using the model’s own out-of-sample error distribution.

Twenty features, ranked fresh each week inside the training window and never using the games being predicted. Validation is strictly walk-forward: every season is predicted using only the seasons before it. Betting odds are never permitted as model inputs — the model is not allowed to see the market it is being asked to beat.

See every Week 4 game on the edge board →

Every game played this season, with the margin. These are results, not ratings: the ranking above is opponent-adjusted expected points added per play, which is why a team’s record and its rank can disagree.