NFL Week 5: Jacksonville Leads. Baltimore and San Francisco Buck the Market

Jacksonville is my strongest Week 5 winner pick at 70.4%. Baltimore and San Francisco are the two picks that oppose Wednesday evening’s saved market favorites. Most of this board is much closer to even.

The Week 4 recap finished with 11 correct winners for the original forecast and 10 for the research comparison. The research version had lower log loss than the original final forecast, but slightly worse Brier score; the original raw blend beat both on those probability measures. That mixed result stays beside the new picks.

Week 5 has one verified issued forecast from the rebuilt pipeline. It was generated Thursday, October 8, at 3:17 a.m. Central and separately frozen unchanged at 3:39:58 a.m. The two Wednesday snapshots below are price observations, not two prediction versions.

One board, five models, no confidence sharpening

Elo, Logistic regression, XGBoost, Bayesian and Random Forest all contribute. The source history includes all 64 completed 2026 games through Week 4, and the fitted weights are now close to one-fifth each: 20.62%, 19.70%, 20.48%, 19.54% and 19.66%, respectively.

The attempted calibration failed its held-out quality checks. The forecast therefore uses T = 1, preserving the raw probabilities exactly. There is no separate sharpened column to add. Passing the integrity checks establishes which forecasts were saved; it does not establish calibrated probabilities.

The historical evaluation rebuilt 40 weekly folds covering 608 games. The expanding-weight comparison had 592 eligible predictions, with 64.4% accuracy and 0.6397 log loss. Equal weights had marginally lower log loss, 0.6395. This reconstructed evaluation does not demonstrate a weighting advantage or add games to the record of forecasts actually published.

The complete Week 5 board

The market column is BetMGM’s paired moneyline probability for the same team selected by the model, after normalizing away the quoted margin. Prices were captured Wednesday, October 7, at 8:59 p.m. Central. They are an observed comparison, not live or closing prices, and that normalization is not a known true probability.

Handicaps below always belong to the listed home team. The total is the bookmaker’s points threshold, not the model’s predicted score.

Jacksonville is the designated home team in London. LAR means the Rams, recorded as LA in the source; LAC means the Chargers. Carolina and Kansas City are on bye. The NFL’s Week 5 guide confirms the fifteen-game slate.

Only Dallas, Detroit and Minnesota have unanimous five-model winner picks. Twelve ensemble favorites are below 60%. Counting agreement and measuring confidence remain different exercises.

Dallas opens Thursday; Jacksonville leads Sunday

Tampa Bay at Dallas starts tonight at 7:15 p.m. Central. The Buccaneers are 0–4 and will start rookie Jalon Daniels again with Baker Mayfield out because of a thumb injury. Dallas is 2–2. The Bucs’ preview supplies that quarterback and game context.

Every model takes Dallas, but the estimates range from 50.1% in XGBoost to 68.6% in Logistic. The blend’s 59.8% is far below the saved market’s 79.2%, calculated from Dallas -475 and Tampa Bay +360. Agreement on the favorite does not mean agreement on how heavily to favor it. The personnel news matters to the matchup; I have not applied an injury adjustment to the frozen probabilities.

Philadelphia meets Jacksonville in London on Sunday at 8:30 a.m. Central. Jacksonville enters 3–1 after two straight wins; Philadelphia is 2–2. The Jaguars’ Wednesday preview describes the start of their consecutive London games.

Jacksonville’s 70.4% leads the board without unanimity. Elo narrowly favors Philadelphia at 50.4%. The other four give Jacksonville 67.0% to 82.1%, with Random Forest highest. Their stronger estimates outweigh Elo’s small opposing lean. The London home designation identifies the forecast row; it does not demonstrate a separately validated travel adjustment.

Two picks against the saved market favorite

Baltimore at Atlanta, Sunday at 7:20 p.m. Central: the model gives Baltimore 54.3%, against 39.1% in the normalized BetMGM pair—a 15.2-point gap, calculated before rounding. Four models favor Baltimore; XGBoost prefers Atlanta at 51.9%.

The quarterback situation belongs next to that number. Lamar Jackson missed Wednesday practice with an ankle injury, after missing the second half against Tennessee. The Ravens’ October 7 update did not establish a final Sunday designation. The saved Baltimore lean should not be read as a confirmed-starter forecast or a measured injury-price advantage.

San Francisco at Seattle, Sunday at 3:25 p.m. Central: the 4–0 49ers visit the 3–1 Seahawks, per Seattle’s game center. San Francisco gets 53.3% from the blend versus 40.3% from the saved market pair. Elo gives Seattle 62.4%; all four other models favor San Francisco, led by Random Forest at 67.1%.

These are modest model leans against stronger market opinions. The gaps identify games to scrutinize, not established betting edges.

Predicted points and spread picks can disagree

The point projections come from the XGBoost margin and total models, with team scores derived from those two outputs. Spread probabilities combine three trained models. Totals combine three trained estimates with an explicit pull toward 50%. These remain uncalibrated outputs, separate from the five-model winner blend.

Jacksonville remains the favorite at -7.5, and the point model’s 8.7-point win would cover that spread. The separate spread blend gives Philadelphia a 51.9% chance to cover +7.5. Philadelphia can cover while still losing by seven points or fewer; this is a pick against the spread, not a pick for Philadelphia to win.

Baltimore’s winner blend favors the Ravens even though its point model projects Atlanta by about one. Those disagreements stay visible; I am not choosing whichever output tells the tidiest story.

Every displayed probability matches its saved threshold. In Wednesday’s 8:34 a.m. to 8:59 p.m. BetMGM comparison, Green Bay moved +3 to +2, Miami +7.5 to +7, and Arizona +5.5 to +5. Giants–Washington fell 43.5 to 42.5, the largest total move at that book. These are same-book observations, not claims about opening lines or every bookmaker.

Monday closes with almost no separation

Buffalo visits the Rams Monday, October 12, at 7:15 p.m. Central, confirmed in the Rams’ preview. The 50.1% Rams winner pick is effectively a coin flip. Bayesian gives Los Angeles 62.0%, while Random Forest gives Buffalo 62.5%; those opposing estimates help explain the narrow blend.

After the games, I will grade all fifteen saved winner forecasts and their components, alongside the spread and totals predictions at their exact lines. Jacksonville’s stronger lean, the Baltimore and San Francisco disagreements, and Monday’s near-even call all stay in the record whichever way they finish.

Source note: Forecasts, issuance verification and Wednesday market observations come from prediction-repository revision b6040c0. Companion editorial evidence retains full-precision values, model and threshold attribution, capture times and source hashes. Football context was checked Thursday morning before kickoff. This article uses the saved forecast without refitting models, updating prices or changing grading records.