A pick'em tool is only worth its probabilities. So before trusting ours, we checked them against real results, including the parts that failed.
For every prop we took the sportsbooks' closing prices, then asked the model for the chance of going over lines away from the book's number, up to 40% higher or lower, because that's where pick'em lines sit. Each week was priced using only what earlier weeks had taught the model, exactly as it runs live, so it never saw the answer.
Then we grouped the predictions and counted what happened. If the model says 70%, those picks should hit about 70% of the time. The dots should sit on the dashed line, and they do. Brier score 0.2059 (lower is better; guessing the average scores 0.2492).
| Model said | Actually happened | Lines |
|---|---|---|
| 4.5% | 9.0% | 1,830 |
| 14.7% | 16.8% | 1,782 |
| 25.9% | 28.2% | 3,200 |
| 35.2% | 37.3% | 5,390 |
| 43.8% | 45.6% | 3,094 |
| 55.5% | 54.1% | 3,642 |
| 64.6% | 62.7% | 5,036 |
| 74.6% | 73.9% | 2,502 |
| 83.7% | 87.5% | 1,539 |
| 92.9% | 88.4% | 327 |
Most pick'em lines match the sportsbook's number. The app pays the same multiplier on both sides, while the books charge more for the likelier side. So we priced each prop at the book's own closing line and took the side the model preferred, exactly what the edge board does.
| Model's chance | Picks | Model expected | Won | 95% range |
|---|---|---|---|---|
| 56% or more | 2,532 | 60.1% | 58.5% | 56.6-60.4% |
| 58% or more | 1,646 | 61.8% | 59.2% | 56.8-61.6% |
| 60% or more | 965 | 63.8% | 58.7% | 55.6-61.8% |
Raw model, before calibration. Break-evens: 54.2% per pick on a 5- or 6-pick flex, 57.7% on a 2-pick.
The model runs hot at the book's line. When it said 60%, those picks won 58.5%. So every probability you see is scaled back by an honesty factor of 0.88, refitted daily from these blind results. After that, picks clearing our 56% bar:
| Calibrated chance | Picks | Expected | Won | 95% range |
|---|---|---|---|---|
| 56% or more | 2,152 | 59.5% | 58.4% | 56.3-60.5% |
| 58% or more | 1,266 | 61.3% | 59.1% | 56.4-61.8% |
Honest version: after calibration, few lines clear the bar, and the sample is too small to call them profitable yet. We publish them because they're our best read, and we grade every one in public. The card's probabilities are where the value is proven: they match what happens.
Injury exits, benchings, games without a target. Textbook models give that almost no chance. Ours learned it from the results.
That's the data, not a rule: the zero-risk and the way props are priced make "less" win more often at the book's line. We'll keep reporting it either way.
Every graded game becomes training data. The model and this page update automatically, the numbers above will move as the season goes on.
Textbook distributions. Our first version used standard statistical shapes. It was overconfident: when it said 85%, those picks hit 71%. We replaced it with the learned model above.
"Sharp book vs soft book" line gaps. When Pinnacle and FanDuel closed on different numbers, pricing from Pinnacle won 53.4% of 335 disagreements, no edge. We don't use it.
A record is only worth something if the picks were really made before the games. Since Nov 14, 2026, every pick we publish also goes out on X @SlipSharpPicks the moment it's published. X stamps every post with its time, and nobody, us included, can backdate one. So far that's 48 picks posted before their games.
On the record, each of those picks links straight to its post, so you can check the time against the game start yourself. If a pick is dropped before its game, we reply under the original post saying so, and it never counts. Picks from before the X account existed carry the publish time from our own database.
Live picks are graded in public on the record.