WEEK IN REVIEW · SEPTEMBER 30, 2026
What the models got right — and what Week 3 exposed
NFL sides held up. Totals did not. College was more encouraging, but the sample still says discipline matters more than confidence.
Sports Intelligence exists to show the evidence, including the ugly parts. Week 3 was a useful example. The NFL model finished 6–4–1 against the spread and 8–6 on moneyline decisions, but totals collapsed to 3–11. Across all graded NFL decisions, that produced a 17–21–1 overall record.
The biggest signal: sides were useful, totals were not
A 60% ATS week is meaningful enough to study, but not enough to declare victory. The more important lesson is the split between markets. Spread decisions held up while totals missed repeatedly. That is a warning against treating one model output as equally mature across every market.
The season-to-date numbers reinforce that point: NFL moneyline decisions are 15–13, ATS is 15–18–1, and totals are 3–11. Totals are currently the weakest part of the system by a wide margin and should be treated as experimental until the calibration improves.
Where the model saw something the market missed
Chicago over Philadelphia was the cleanest example. The final pregame model preferred Chicago both outright and with the points. Chicago closed as a +3.5 underdog in our stored pick and won 27–7. The moneyline was +160.
Denver over the Rams was another strong result. The model carried Denver +1 and Denver moneyline at -103. Denver won 30–26. The same pattern showed up with the Giants: New York -2.5 and the Giants moneyline both won in a 12–7 game.
Buffalo -7 also covered in a 24–16 win over the Chargers. Those results matter because they were not simply “pick the favorite” outcomes — the spread model was identifying specific price/value gaps.
The misses matter more than the victory laps
The most obvious miss was Atlanta at Green Bay. Green Bay was modeled around a 69% win probability and lost 35–14. The under 43.5 also lost. That game is precisely the kind of result that should reduce confidence in early-season priors rather than be explained away after the fact.
Houston was another example. The model liked Houston both outright and -1.5; Indianapolis won 19–17. Tampa Bay moneyline also failed against Minnesota. Cincinnati failed both outright and against the spread in Pittsburgh.
The lesson is not that the model is useless. It is that early-season football is noisy, and confidence needs to be earned as current-season information replaces stale priors.
College: encouraging, still experimental
The NCAA experimental model produced 49 wins, 43 losses and 1 push across 93 graded final-pregame legs during the latest completed window. That is a positive result, but it remains an experimental track rather than an official recommendation engine.
The college model has a different problem from the NFL model: the sport itself has wider team-strength distributions, more uneven information quality, and larger roster/context swings. A winning week is useful evidence, but the real test is whether the same process remains calibrated across several weeks and across different market types.
What changes next
Three things come out of this review. First, NFL totals need work before they deserve equal billing with sides. Second, the Week 5 learning gate remains important: current-season performance should increasingly dominate preseason and prior-year assumptions. Third, college results should continue to be tracked separately as experimental until the sample is large enough to evaluate stability.
The purpose of the model is not to be right every Sunday. It is to identify where its probability estimate differs from the market, preserve the prediction before kickoff, grade it honestly afterward, and learn whether that difference was actually useful.
All records above come from frozen Sports Intelligence pregame outputs and postgame grading. Historical picks are not rewritten after results are known.