NFL DFS Cash-Game System

Build report · July 11, 2026 · walk-forward backtest, 71 main slates (2022–2025), adversarially audited

Dynasty trade board · Rookie draft board (Aug 1)

Verdict: NO-GO for real money on open-lobby double-ups

Best honest estimate: ≈47% win rate (range 45–49%) against a true $5 double-up cash line vs the 50% breakeven. The optimizer does real work; the field's projections are simply as good as ours. September = paper trading only.

Backtest — win rate vs the cash line (needs >50%)

vs raw recon line

50.7%

honest DU offset (+1.5)

≈47.2%

calibrated line (+3)

45.1%

line +5

39.4%

random legal lineups

9.1%

Tick mark = 50% breakeven (double-up pays 2× entry to top 43.5%; rake 13%, verified from contest purse math). Bars scaled to 65% max. Bootstrap 95% CI on the headline: [39.4%, 62.0%] — statistically a coin flip at n=71.

+41.6

optimizer lift over random legal lineups (win-rate pts)

0.910

correlation: our simulated field vs Levitan's real H2H results (60 wks)

52.6% / 50.7%

predicted P(cash) vs realized — the machine's self-estimate is honest

~2,175

slates needed to prove a 53% edge — you get 17/season

Season stability (the audit's sharpest finding)

Season	Slates	vs raw line	Mean margin (pts)
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

What got built (all validated, all free)

Component	Result
Prop-implied projections (free market lines)	OOS MAE 5.38 — beats LineStar's own projections (5.42), salary-implied (5.82), trailing avg (5.84)
Cash-line reconstruction (124 real double-ups)	line = field median +3.3 ±0.7; same-slate contests agree within 1.0 pt
Calibration vs Levitan's published results	0.910 correlation; field bias measured at +4.5 pctile (≈3–4 line pts, H2H-tough)
Cheap target-spike screener (Nick's setup)	≤$4K + prev wk 7 tgt/5 rec → +4.1 pts over salary band, half the bust rate; needs live OOS revalidation
Target value by depth (54,314 targets)	screens 1.41 pts/tgt → 20+ air yds 2.09 ±3.4; flat EV 0–15 yds, variance is the difference
Variance model	TD-share = #1 widener (5.9 vs 7.3 weekly std at equal 10-pt mean); catch-share = #1 tightener
Data archive (rescued, $0)	DK salaries/points/slates/ownership 2014–25 · prop lines 2019–25 · pbp w/ EPA+wind · 4.9M-word Ballers transcript corpus

September plan — paper only, evidence decides

Weeks 1–2: paper. Lineup recorded at lock, graded Monday vs real cash lines — start collecting DK contest standings before their 4-day purge (the ground truth that doesn't exist historically).
Measure the beginner lobby. Everything above grades vs the GIANT $5 — the toughest cash pool. Sub-$5 lobbies (experienced players barred) are plausibly softer; two paper weeks measure it directly.
Re-validate the screener out-of-sample on live 2026 weeks before it counts.
Wire in weather (wind >15 mph) + funnel-EPA layers; start the Kalshi NFL prop harvest (vig-free lines on existing infra).
Projections vs the field is the only lever with real headroom — the optimizer is already fine.