Lions at Bills: Three Forecasts Favor Buffalo. Two Drive the Ensemble

My pick is Buffalo. The saved final forecast gives the Bills 79.4%. The raw blend starts at 63.1%.

All three available model forecasts favor Buffalo over Detroit tonight. That sounds familiar after the Seattle and Rams previews. There is an important change underneath it: only two models contribute to this week’s ensemble.

The Lions visit the Bills on Thursday, September 17, at 7:15 p.m. Central, on Prime Video. Both teams enter 1–0, and this is Buffalo’s home opener at the new Highmark Stadium. The Bills’ official game page lists 8:15 p.m. Eastern; their viewing guide confirms the broadcast.

Three forecasts, two contributors

The available models span only 8.6 percentage points on Buffalo, from Bayesian’s 56.8% to Logistic’s 65.4%. They agree on the winner without any individual model reaching the final forecast’s 79.4%.

The distinction is Bayesian’s role. Across the full Week 2 board, its home-win probabilities have a standard deviation of 0.02184, below the pipeline’s 0.05 threshold. The existing health check excludes a model when its probabilities vary that little across the slate. Bayesian still has a saved forecast for this game; it receives zero ensemble weight.

That rule does not establish that Bayesian is wrong about Detroit or that excluding it improves this forecast. It explains which numbers actually produced the saved final probability. XGBoost and Random Forest remain unavailable in this saved board.

After the exclusions, the effective weights are 34.2% Elo and 65.8% Logistic. Applying those weights to the full-precision saved inputs reconstructs Buffalo’s raw probability at 63.15%. The same calculation reproduces all sixteen Week 2 final probabilities after temperature scaling, to rounding precision.

Where the extra confidence comes from

The retained T = 0.4 adjustment takes that 63.15% raw blend to 79.35%, adding 16.21 percentage points. It changes the probability, not the pick, and adds no new observation about either team.

This is not quite the same ensemble I wrote about in Week 1. Bayesian contributed to that blend. Tonight’s final number comes from Elo and Logistic, with Logistic carrying nearly two-thirds of their combined weight.

The calibration limitation remains. The temperature setting was fitted using a different model combination, and some of its evaluation inputs were fitted-sample predictions. That does not provide clean evidence for the confidence of this two-model lineup on unseen games. A strong result tonight would add one observation; it would not settle the setting.

Two wins on Sunday, two different stories

Detroit beat New Orleans 31–30 in overtime. The Lions had led 21–0, then needed an overtime touchdown and a stop on the Saints’ two-point attempt to finish the game. That is the sequence in the Lions’ official recap, not an explanation I can attribute to a model feature.

Buffalo won 36–31 at Houston. Trailing 31–30 late in the fourth quarter, the Bills drove from their own four-yard line and took the lead on Josh Allen’s 34-yard touchdown pass to Joshua Palmer. The Bills’ official recap records the road win and the late response.

Both teams have a short turnaround. Those opening results and the new-stadium setting make the matchup interesting. The saved forecast does not document the separate effect of the short week, stadium opening, or individual plays on tonight’s probability. I am keeping the football context separate from claims about what the models learned.

Grade the probability, whatever the score looks like

The Seattle recap corrected my earlier tendency to connect confidence to how comfortable a win should look. The Rams recap then showed the cost of sharpening a losing pick.

Both lessons belong here before kickoff. 79.4% is a probability of a Buffalo win. It is not a predicted margin or a promise of a comfortable game. A one-point Bills win and a four-touchdown Bills win receive the same score for this winner-only forecast. A close Detroit win would penalize it exactly as much as a Detroit blowout.

Here are the two possible decisive outcomes, calculated from the full-precision inputs. These are hypothetical scores, not recorded results.

Lower is better. Binary Brier score is (p - y)^2, where p is Buffalo’s probability and y is 1 for a Bills win or 0 for a Lions win. Log loss is the negative natural logarithm of the probability assigned to the winner. A tie would need an explicit grading rule; this table covers either team winning.

Sharpening earns more credit if Buffalo wins and a larger penalty if Detroit wins. The first week is a small sample, and one more game cannot tell us how often comparable 79% favorites should win. It can still be graded honestly.

The pick I am putting in the record

Bills to win: 79.4% final ensemble, 63.1% reconstructed raw blend.

Detroit retains 20.6% in the final forecast and 36.9% before sharpening. I want both numbers preserved, along with Bayesian’s excluded forecast, so the follow-up can evaluate what was actually said before the result.

There is no score, margin, or total prediction in this saved row. The market-line fields do not supply a fresh, timestamp-verified betting comparison. This is a winner forecast, with its confidence and limitations on the page.

The Week 1 recap grades the opening slate, and the Week 2 preview carries the full new board. Tonight, Buffalo is the pick. The follow-up will keep the original probabilities and grade all three available forecasts, plus both ensemble stages, against the result.

Source note: Forecasts come from the Week 2 prediction artifact for game 2026_02_DET_BUF, with successful generation logged September 15 and the CSV committed September 16. They were inspected September 17 before kickoff. The raw blend is an editorial reconstruction using the saved model probabilities, retained weight configuration, and the pipeline’s across-slate health rule; it is not a separate raw-probability field in the CSV. The final probability is preserved exactly as saved. Game logistics and Week 1 results were checked against the official team sources linked above. No model was retrained, forecast refreshed, or prediction ledger updated for this article.