Under the Hood

The Models

The dashboard's projections aren't guesses — they come from statistical models fit on real play-by-play and results, then tested on games they never saw. This page explains how they work, shows how well they actually perform, and gives you a playbook for using them to size up any team.

Try It: Matchup Predictor

SP+ MARGIN → WIN PROBABILITY

Pick any two teams. The model takes the gap in their SP+ ratings (a neutral-field points margin) and runs it through the project's margin model — final margins are roughly Normal with a 13.5-point standard deviation — to turn that gap into a win probability. Drag the pick around and watch where it lands on the curve.

VS · NEUTRAL FIELD
MODEL PROJECTION
WIN-PROBABILITY CURVE
Favorite's SP+ edge (points) →

In-Game Win-Probability Model

THE FLAGSHIP · LOGISTIC REGRESSION

A logistic-regression model that predicts the home team's win probability from six pregame signals — efficiency edges, recent form, win rate and schedule strength. It's the most rigorously tested model here: fit on earlier games, then scored on held-out games it never saw during training.

CALIBRATION — DOES 70% MEAN 70%?
Perfect calibration Model (held-out games)

Each dot is a bucket of games grouped by predicted win probability. The closer the dots sit to the diagonal, the more honest the probabilities — when the model says 70%, those teams really do win about 70% of the time.

HOW MUCH BETTER THAN A GUESS? — BRIER SCORE (LOWER IS BETTER)

What Drives a Prediction

COEFFICIENTS · R ↔ PYTHON PARITY

Every input's pull on the outcome, as an odds ratio — how much the home team's winning odds multiply for a one-unit edge in that stat (above 1 helps, below 1 hurts). The same model is fit independently in R and Python; the build fails unless the two agree, so both estimates are shown side by side.

Reproducibility gate. That's the largest gap between the R and Python coefficient estimates across every term. Two independent implementations landing on the same numbers is how the pipeline proves a result is real and not an artifact of one toolchain.

Preseason Priors — the 2026 Forecast

PREDICTING BEFORE THE SEASON

The 2026 Projected Wins on the dashboard come from a separate model that has to predict games before any current-season form exists — using only prior-year strength, returning production and recruiting. Naturally harder than the in-game model, but it still clears its naive baseline by a clear margin.

The Supporting Cast

EIGHT MORE MODELS BEHIND THE STATS

The efficiency and talent numbers you see are themselves outputs of a modeling library — each built twice (R and Python) and parity-checked. The headline figure on each card is a real result from the data.

A Playbook for Sizing Up a Team

HOW TO USE ALL OF THIS