Best honest estimate ≈ 47% win rate (range 45–49%) vs the 50% breakeven against a true $5 double-up field. The machinery works — the field's projections are simply as good as ours. Total cost of learning this: $0.
| Season | Slates | vs raw line | Mean margin | |
|---|---|---|---|---|
| 2022 | 18 | 55.6% | +0.8 | in-sample |
| 2023 | 17 | 47.1% | −1.2 | |
| 2024 | 18 | 61.1% | +0.5 | carried the edge |
| 2025 | 18 | 38.9% | −9.9 | collapse |
| Prop-implied projections | beats LineStar's own projections OOS (MAE 5.38 vs 5.42) from free market lines |
| Cash-line reconstruction | 124 real double-ups; line = field median +3.3 ±0.7; twin contests agree within 1 pt |
| Levitan calibration | 0.910 correlation with published real results; field bias measured, not guessed |
| Cheap target-spike screener | ≤$4K + prev wk 7 tgt/5 rec → +4.1 pts over band, half the bust rate re-validate OOS live |
| Target value by depth | 54,314 targets: screens 1.41 pts/tgt → 20+ air yds 2.09 ±3.4; EV flat 0-15 yds |
| Variance model | TD-share = #1 widener (±5.9 vs ±7.3 at equal means); catch-share = #1 tightener |
| Data archive | DK salaries/pts/slates/ownership 2014-25 · props 2019-25 · pbp w/ EPA+wind · 4.9M-word pod corpus |
The September paper season described here was never built: no lineup logger, no 2026 salary or ownership mirror, nothing in the pipeline touched after 2026-08-07. Week 1 has since been played and DraftKings purges contest standings after four days, so that week's field data is gone.
Three findings from this work outlived it and moved to Terms, where they get used: the cheap target-spike screener (now the breakout flag on the box scores), target value by depth, and the variance model.
Reopen only with a new mechanism — a measured soft lobby, or a projection edge the field does not already have. The optimizer was never the problem.