Atlanta Won 35–14. All Three Models Picked Green Bay

I picked Green Bay. Atlanta won 35–14 at Lambeau Field, and the Falcons controlled most of the night after falling behind early.

The pregame preview gave the Packers 63.3% in the raw blend and 79.5% after temperature scaling. Elo, Logistic regression and Bayesian all favored Green Bay. Every one of them missed.

The preview also identified a specific weakness in the betting comparison: only the confidence adjustment pushed Green Bay above the break-even threshold for Wednesday’s saved -250 moneyline. Thursday delivered the losing outcome for that pick, with a larger probability-scoring penalty for the sharper forecast.

Atlanta took over on the ground

Green Bay led 7–0 before Atlanta scored 27 unanswered points. The Falcons led 17–7 at halftime, 24–7 after three quarters, and 27–7 early in the fourth. The win moved Atlanta to 1–2, dropped Green Bay to the same record, and ended the Packers’ run of 13 consecutive home-opening wins. The Packers’ official recap records the sequence.

The clearest measure of the mismatch was 242 rushing yards for Atlanta against 17 for Green Bay. Atlanta finished with 498 total yards and held the ball for 37:05. Each team committed one turnover, so the margin did not come with a lopsided takeaway count. Those totals are in the official gamebook.

Bijan Robinson ran for 194 yards and two touchdowns on 29 carries. Drake London caught nine passes for 194 yards. Michael Penix Jr. completed 18 of 25 passes for 256 yards, one touchdown and one interception in his return. Jordan Love threw for 312 yards and two touchdowns, but needed 53 attempts and also threw an interception. The official box score makes the contrast clear.

Green Bay still had an opening after halftime. Down 17–7, the Packers drove deep into Atlanta territory before a failed fourth-and-one pass ended the possession. Atlanta answered with a 92-yard touchdown drive, capped by Brian Robinson Jr., to make it 24–7. That sequence widened the gap when Green Bay had a chance to close it. Official gamebook

The preview’s question about Penix and sustained drives received a convincing answer on the field. His return coincided with a much more productive Atlanta offense. That does not isolate his contribution from the blocking, rushing, receiving and opposing defense, or establish exactly how a forecast should have adjusted for him.

Grade every forecast that picked the loser

These are the probabilities preserved before kickoff, not predictions regenerated after the result. The table shows how much probability each forecast left for the team that actually won.

Lower is better. With Green Bay losing, binary Brier score is p², where p is Green Bay’s pregame win probability. Log loss is −ln(1 − p), using the probability assigned to Atlanta. All calculations use the retained full-precision values.

Logistic gets the smallest penalty because it left Atlanta the most probability. It still picked the wrong winner. The final ensemble gets the largest penalty because it left Atlanta the least.

All three component models contributed this week. Their effective weights were 21.09% Elo, 40.58% Logistic and 38.33% Bayesian. XGBoost and Random Forest were unavailable. The raw and final ensemble rows are two stages of the same blend.

Sharpening raised Green Bay from 63.26% to 79.55%. On this result, it increased Brier score by 0.2326 and log loss by 0.5858. Those penalties match the Atlanta-win calculations retained with the pregame evidence.

The 21-point margin makes this a decisive football loss. It adds no separate penalty to these winner-probability scores: a one-point Atlanta win would receive the same grades. No score, margin, spread-cover or totals prediction was available in the saved forecast.

The apparent price edge ended in a loss

Wednesday’s archived Caesars -250 quote required a 71.43% Green Bay win rate to break even. Every individual model, and the raw blend, sat below that threshold. The final 79.55% forecast was the only one above it.

At that price, the raw blend implied an expected net return of −11.44% per dollar staked. The final blend implied +11.37%. The selection engine used the final forecast and flagged Green Bay with MEDIUM confidence while the system’s performance status was already YELLOW.

A hypothetical $1 stake on Green Bay would have lost $1. That is the settlement of the illustrated bet, not a claim that I placed it or a change to the paper ledger. The pregame expected return was an average implied by an assumed probability; it was never the promised payout from this game.

One loss cannot prove that a quoted price had no value beforehand. It does supply another adverse observation for the probability adjustment that created the positive estimate. The price comparison and model recommendation were both dependent on that estimate.

The running record is now 17–16

Adding this game to all 32 original forecasts covered in the Week 2 recap leaves both ensemble stages at 17 correct winners in 33 games.

This includes the complete original Week 1 and Week 2 boards plus Atlanta at Green Bay. It is a running editorial record through Thursday, not a completed Week 3 grade or a refreshed production dashboard.

The raw blend still has lower average penalties on both measures. These are the forecasts actually issued across changing model combinations, including Bayesian’s exclusion from the original Week 2 ensemble and its return in Week 3. They are not a controlled test of one unchanged blend.

What carries into the rest of Week 3

The Buffalo follow-up gave sharpening credit when it helped a winning pick. Atlanta adds another game where it magnified a miss. Both stay in the record.

A 20.5% outcome can happen. The question is whether the issued probabilities hold up across many games, including every miss. This result worsens the accumulated scores for the confidence adjustment; it does not identify the right replacement temperature or justify rewriting the forecast after the fact.

Atlanta earned a convincing win. My models selected Green Bay, and the final ensemble expressed more confidence than any of its contributors. The original preview stays as written. The remaining fifteen Week 3 games will be graded against their preserved forecasts under the same rules.

Source note: Game 2026_03_ATL_GB finished ATL 35, GB 14 on September 24, 2026. The result and game details were checked September 25 against the official sources linked above. Probabilities and the Wednesday price come from the evidence saved with the pregame preview; running averages use full-precision evidence behind the Week 2 recap. Companion editorial evidence preserves the calculations and source hashes. This article did not regenerate forecasts, change model settings, run weekly grading, or modify evaluation or paper-betting records.