How accurate are our predictions?
Every forecast we serve is the model's best guess before a show happens. Once the real scores post, we grade ourselves. This is the honest scorecard for the 2026 season — World Class + Open Class only.
Accuracy over the season
Average error per show date. Dashed lines mark model upgrades.
Forecasts sharpen near showtime
Average error grouped by how many days before the show the forecast was made.
Which way do we miss?
Distribution of signed error. On average we under-predict by 0.68 pts.
Accuracy by corps
Best forecast to worst. ▼ = we tend to under-score them, ▲ = over-score. Min 3 shows.
Show by show
Most recent first. Tap a show to see the full forecast-vs-actual diff.
By model version
How each served model era graded out. Newer models cover only the latest shows.
Methodology
“What we said going in” is the last forecast saved strictly before each event's start — exactly what was served the night before the show. Because the model never sees the show it's forecasting, these numbers are leakage-safe.
Predictions are joined to the official posted scores per corps. Scope is World Class and Open Class only — the models don't cover all-age, so those shows are excluded. “Error” is predicted total minus actual total; MAE is the average of its absolute value. “Winner called” is how often the predicted first-place corps actually won. The lead-time chart uses every saved run, not just the final one.
The model has changed over the season (final2 → v10.5 → v11 field-pace ensemble); each show is graded against whichever model actually served it. These are AI-generated forecasts — an estimate for fun, not a guarantee.