Accuracy Report
Every prediction is graded against what actually happened, always using only the results that were available before that match, so nothing is scored with hindsight. Chronological walk-forward: each match is predicted using only matches completed before it (warm-up of 6 qual matches per event). No final rankings or later results leak into a prediction.
Experimental, AI-assisted analytics
FRCinsights uses AI-assisted experimental analytics. Predictions, rankings, ratings, and insights are not guaranteed to be accurate and should not be treated as official results or professional advice. Data may be incomplete, delayed, miscategorized, or incorrect. FRCinsights is a new project under active development, and accuracy will improve over time.
Model apex-0.4.0 · 16913 predictions across 206 events · 2026-03-03 → 2026-05-02
Confidence calibration
When APEX says 70%, does the favorite win ~70% of the time? Closer bars = better calibrated.
| Stated confidence | Actual win rate | Sample |
|---|---|---|
| 50–60% | 59% | 4572 |
| 60–70% | 72% | 3923 |
| 70–80% | 83% | 3351 |
| 80–90% | 90% | 2738 |
| 90–100% | 97% | 2329 |
Baseline comparison
APEX vs naive predictors on the same matches.
Best-predicted events
Worst-predicted events
Where the model was most wrong
High-confidence predictions that missed.
Model limitations
Small samples early
Week 1 ratings rest on very few matches and swing the most. Confidence scores reflect this.
Cold-start teams
Teams with no current-season matches use heavily-regressed provisional estimates with wide uncertainty.
Rare events
Mechanical failures, no-shows, and unusual fouls are hard to predict and inflate score error.
Regression to the mean
APEX deliberately discounts single lucky/unlucky matches, so it can lag a genuine breakout for a match or two.
FRCinsights does not claim its predictions are guaranteed. Accuracy improves as more data arrives and the model is refined.