2026 season
Every game on NC State's schedule, predicted before the season started and frozen exactly as it was written.
This is the model's first attempt at a whole season. Twelve games, predicted before a single snap was played, written down on July 30 2026. The point isn't to be right; it's to be a fixed, checkable record of what a homemade model said in July, so that in December you can see exactly where it went wrong and what changed in response. This is a model being built in public, flaws included.
What it does reasonably well is margin: who wins, and roughly by how much. That comes from every play of every FBS team last season, adjusted for the quality of who it happened against, then regressed toward the average to account for a year of roster turnover. It is a real method, not a hunch. Version 3 added returning production, which adjusts a team by how much of last season's production it actually brings back.
Known limitation: the total
Version 1 could not predict total points, so it didn't pretend to. Add up either Version 1 row below and you get 52.9. That isn't a bug: it's the 2025 league average, used as a stand-in because two attempts at predicting totals scored worse than simply guessing that average. Version 3 replaced the constant with pace, each side's drives per game, measured from play-by-play. It is about two per cent better on total error and no more than that, so the margins are still the part worth reading.
Games 9 and 12 show what the constant cost. Under Version 1 both read 28.1 to 24.8, because the two margins land 0.06 points apart and a fixed total turns equal margins into identical scorelines. Version 3 separates them, 29.9 to 24.0 and 28.1 to 22.8, because Duke and North Carolina play at different speeds. Both numbers are still on the board, in the two columns.
Loading the frozen forecast…
Still to come on this page
