MLB Projection Accuracy
Every morning we archive the projections we published. Every night the box scores grade them. This page is that ledger: no cherry-picking, no retro edits, updated daily.
Scoring window: Apr 24, 2026 to Sep 3, 2026 · DraftKings main slates · players who took the field only
Measured against winners
The test that matters: when a lineup actually cashed at the top of a GPP, were its players on our board that morning?
Average number of players from a top GPP finish that sat in our top-30 board (by projection or value) the same morning, across 148 top-finishing lineups we could grade against our archive.
of those lineups had 6 or more of their 10 players on our board. A winning roster is mostly findable before lock; the optimizer's job is assembling it.
| Slate size | Winning scores studied | Median winner | Best |
|---|---|---|---|
| 2-4 games | 66 | 223.4 DK pts | 358.0 |
| 5-7 games | 44 | 308.8 DK pts | 395.8 |
| 8-10 games | 6 | 390.3 DK pts | 396.7 |
What first place has actually required, from the corpus of real DraftKings GPP results behind the optimizer's Historical Pattern Score.
Week by week
Recent scoring weeks. Baseball is loud; the sample sizes are the point.
| Week of | Slate days | Players scored | Avg error (DK pts) |
|---|---|---|---|
| Aug 31, 2026 | 4 | 551 | 7.61 |
| Aug 24, 2026 | 7 | 1,220 | 7.33 |
| Aug 17, 2026 | 6 | 951 | 6.89 |
| Aug 10, 2026 | 7 | 1,086 | 7.11 |
| Aug 3, 2026 | 6 | 886 | 7.39 |
| Jul 27, 2026 | 7 | 1,102 | 6.91 |
| Jul 20, 2026 | 7 | 1,231 | 7.08 |
| Jul 13, 2026 | 2 | 382 | 7.38 |
| Jul 6, 2026 | 6 | 820 | 7.52 |
| Jun 29, 2026 | 5 | 800 | 7.23 |
How this is scored
The methodology, including the parts that do not flatter us.
- What gets graded: the projections we published to the DraftKings main slate each morning, archived at publish time. The grade is each player's actual DraftKings score that night.
- Late scratches are excluded, not zeroed. A player who never took the field is a lineup-news problem, not a projection error. 4,740 player-slate entries were excluded on that basis and are counted here so you can see the size of it.
- Our projections run optimistic. On average, actuals come in about 4.3 DK points below projection. They are a ranking tool: the correlation and the value edge above are the honest measures of whether the ORDER is right, and the order is what builds lineups.
- The winners corpus is the set of top-finishing lineups from real DraftKings GPPs, the same data behind the optimizer's Historical Pattern Score. Coverage is only computed where a lineup's exact slate exists in our archive, matched by date and slate size.
- Nothing is regenerated after the fact. The archive is written before lock; this page cannot be improved retroactively, only by projecting better tomorrow.