Accuracy
Every number below is measured on games that came after the model's training data ended, so the model never saw them. Sports and markets that fail our own checks are listed next to the ones that pass. Built 2026-10-02.
Fantasy-point projections
17 of 22 sports pass the release check. Each sport is scored on two things.
- Rank correlation per slate
- How well the projections order the players within one slate (Spearman; 1 is perfect, 0 is random). This is what decides lineups. NBA's 0.73 means the projected order and the actual order agree closely, but not perfectly.
- 80% range coverage
- How often the actual score lands between the floor and ceiling we publish. They're the 10th and 90th percentiles, so the target is 80%. Much higher means the ranges are too wide; much lower means they're overconfident.
| Sport | Rank corr. | 80% range | Avg. miss | Games | Check |
|---|---|---|---|---|---|
| NBA | 0.73 | 80% | 7.7 pts | 53,006 | Pass |
| WNBA | 0.72 | 79% | 6.8 pts | 49,032 | Pass |
| CFL | 0.72 | 77% | 3.4 pts | 6,456 | Pass |
| College basketball | 0.72 | 80% | 6.3 pts | 152,766 | Pass |
| Formula 1 | 0.63 | 79% | 11.9 pts | 5,849 | Pass |
| NFL | 0.62 | 82% | 4.8 pts | 46,082 | Pass |
| EuroLeague | 0.60 | 79% | 6.4 pts | 30,049 | Pass |
| College football | 0.57 | 80% | 5.3 pts | 184,254 | Pass |
| League of Legends | 0.53 | 77% | 18.8 pts | 296,574 | Failits 10th / 90th percentiles aren't calibrated |
| NBA Summer League | 0.52 | 78% | 8.1 pts | 3,116 | Pass |
| Dota 2 | 0.50 | 80% | 49.5 pts | 7,434 | Failits 10th / 90th percentiles aren't calibrated |
| Soccer | 0.42 | 80% | 2.8 pts | 151,711 | Pass |
| UFL | 0.41 | 79% | 4.5 pts | 1,845 | Faildoesn't reliably rank players better than a season average |
| NHL | 0.41 | 80% | 4.5 pts | 49,554 | Pass |
| Tennis | 0.40 | 81% | 20.7 pts | 291,768 | Pass |
| Golf | 0.33 | 79% | 20.7 pts | 34,774 | Pass |
| MLB | 0.29 | 83% | 4.8 pts | 86,954 | Pass |
| Valorant | 0.28 | 83% | 18.0 pts | 23,634 | Failits 10th / 90th percentiles aren't calibrated |
| Counter-Strike | 0.27 | 79% | 21.3 pts | 11,205 | Pass |
| NASCAR | 0.25 | 79% | 16.1 pts | 13,188 | Pass |
| Call of Duty | 0.25 | 78% | 45.8 pts | 9,011 | Failits 10th / 90th percentiles aren't calibrated |
| MMA | 0.20 | 80% | 32.5 pts | 6,875 | Pass |
“Avg. miss” is the mean absolute error in that sport's fantasy points, so it isn't comparable across sports. A sport passes when it beats both a season average and a last-10 average at ranking players (95% bootstrap interval above zero), misses the score level by less than they do, and has calibrated percentiles. Each row is that sport's most recent check, run between 2026-09-22 and 2026-09-23.
Pick'em probabilities
For each market we re-predict thousands of past games from only the games before them, with the line where a book would set it. The calibration map is fitted on the earlier games and every figure here comes from the later ones. A market passes when its predictions land within 5 points of what actually happened in every probability band with 100+ games, and it predicts better than the base over-rate. 9 of 23 checked markets pass. The others, and any market not checked yet, still show their simulated distribution but the pick'em simulator claims no edge on them.
| Market | Largest gap | Games | Check |
|---|---|---|---|
| Counter-Strike kills maps 1 2 | 1.1 pts | 445 | Pass |
| MLB hits | 1.2 pts | 1,425 | Pass |
| NHL shots on goal | 1.6 pts | 1,171 | Pass |
| NBA reb | 3.5 pts | 1,478 | Pass |
| NBA ast | 3.5 pts | 1,478 | Pass |
| WNBA reb | 3.8 pts | 1,104 | Pass |
| MLB total bases | 4.5 pts | 1,425 | Pass |
| NBA pts | 4.6 pts | 1,478 | Pass |
| NHL points | 4.7 pts | 1,171 | Pass |
| NFL rec yds | — | 223 | Failtoo few later games to judge |
| NFL receptions | — | 223 | Failtoo few later games to judge |
| NFL rush yds | — | 223 | Failtoo few later games to judge |
| NFL pass yds | — | 223 | Failtoo few later games to judge |
| Tennis games won | 2.4 pts | 946 | Fail |
| NBA fg3m | 5.5 pts | 1,478 | Fail |
| WNBA ast | 5.8 pts | 1,104 | Fail |
| WNBA fantasy points | 5.9 pts | 1,101 | Fail |
| Soccer shots | 6.2 pts | 754 | Fail |
| WNBA pts | 7.4 pts | 1,104 | Fail |
| MLB pitcher strikeouts | 9.2 pts | 607 | Fail |
| NBA pra | 9.3 pts | 1,478 | Fail |
| WNBA pra | 9.5 pts | 1,104 | Fail |
| Tennis aces | 9.9 pts | 946 | Fail |
A pick'em edge is measured against the play's break-even, not 50%: each leg of a 2-pick PrizePicks Power play has to hit 57.7% to return its stake.
What this doesn't show
- Past accuracy is no promise about the next slate. A single slate is noisy even when the model is good.
- These are the projection and probability models. The contest simulator's opponent field isn't yet checked against real contest standings, because there aren't enough contests on file.
- Simulated probabilities are for research and entertainment, not betting advice.