How Newton v3 Was Built, Step by Step
Every measured gain, every idea we cut, and the checks we set before scoring anything. The gains are modest, and the method behind them is the part we are proudest of.
Newton is the intelligence layer behind Saturview. The app's rankings, metrics and projections run on their own models, and Newton is where the predictions for individual games come from. This note walks through how the current version, v3, came together: the measured result at each step and the ideas we dropped along the way.
One scope note before the numbers. Development results come from replays, where each season is predicted using only the seasons that came before it. Nothing here is a live record.
Where it started
The original rating model picked winners 72.4% of the time across the 2021 to 2023 development seasons, with a Brier score of 0.1844. (A Brier score measures how well probabilities match results. Lower is better.) That was the baseline every later change had to beat.
Adding signal one piece at a time
The first real gain came from blending two different views of team strength. Winners barely moved, at 72.3%, but the Brier score improved from 0.1844 to 0.1823. Adding this season's play efficiency lifted winners to 72.9%. Adding schedule context brought the model to 73.1% and a Brier score of 0.1815 in the development seasons.
Plenty did not make it. A pregame Elo rating, a set of twenty situational inputs (which overfit), QB continuity, form against expectation, travel, kickoff time, elevation and a more complex machine-learning approach were all tested and dropped. Game weather, rest days, short rest and pace were tested too. None of them helped, so they stayed out.
One new input did help: knowing that a starting quarterback is out. How that was tested, and why it only counts when an absence is recorded before kickoff, is in The Human Layer.
Newton v3 and the gate
The v3 plan and its benchmarks were written down and committed before any candidate was scored. The candidates were frozen before evaluation, and every scoring run was logged. There were three looks in total.
On 1,498 paired games from 2024 and 2025, v3 beat the original model on Brier score (0.1788 against 0.1813), margin error (12.40 against 12.55 points) and winners (72.0% against 71.7%). Each season was trained only on the seasons before it. Those two seasons are reused history, so this is a replay, not a blind test.
The gains are modest, about 0.3 points of winners and 0.16 points of margin, and they show up in aggregate rather than week to week. Newton's 80% ranges covered 81.4% of final results, inside the 77% to 83% target. Projected season win totals also got a little closer to the real ones: off by 1.23 wins against 1.25 in 2024, and 1.29 against 1.33 in 2025.
Two checks sit apart from the replay. A diagnostic on 2026 Weeks 1 to 7 showed 75.5% winners against 73.0% for the original model. That is neither a blind test nor a live record, because the production coefficients were trained on 2026 games through October 11. The first truly blind check is a set of 99 FBS forecasts, frozen on October 16 before the first kickoff. They have not been scored yet.
Rankings: a study, then a rule
Team rankings used to follow an older family of strength measures. When the board was ordered by the Newton rating instead, it beat the old order out of sample across 3,376 forecasts from 2024 and 2025: Brier score 0.1855 against 0.1921, and margin error 12.28 against 12.55 points. The Top 25 also moved less from week to week, 2.05 places against 2.39 in 2024 and 1.99 against 2.20 in 2025.
Then came a rule choice. On October 18 we decided that team ranks use this season only. That costs some accuracy, we measured how much, and we kept the rule. The full accounting is in our rankings head-to-head. Newton's picks did not change.
Projecting the committee
Newton's projected committee ranking matched 10.25 of the committee's final top 12 on average. Each of the 12 seasons from 2014 to 2025 was tested with that season held out. The old editorial formula managed 9.75 and the rating alone managed 9.42. With the this-season rank model plugged in, the projection held up on held-out 2022 to 2025, matching 11 of the top 12 on average. This is always labeled as Newton's projection, never as the committee's own ranking.
Where v3 stands
Against an outside ranking model, built from a published method and scored on the same games, Newton came out statistically tied with a slight lean its way. Against the consensus closing line, Newton is close but not ahead: the line is still more accurate on winners and on margin. Not betting advice. Both comparisons are in our rankings head-to-head, and what "75%" should mean is in The Human Layer.
The pace
The project's history runs from September 8 to October 18, about six weeks and 293 commits. The v3 work, the rankings studies and the head-to-head all landed in the final week, from October 12 to 18. The credit belongs to the gains that held up, the ideas that were cut when they did not, and the habit of writing the test down before running it.
Newton’s picks lock at kickoff and are graded in the open, wins and losses both. See Newton’s record and Newton and the betting line for where the model stands today.