BlackArbs LLC · NFL model research

Four seasons · FanDuel and DraftKings

This Model Beat FanDuel AND DraftKings on NFL Totals. Then the “Best” Bets Lost Money.

Unfiltered LightGBM returned +4.13% at FanDuel and +7.17% at DraftKings. When both models picked the same side, the book-local results were +10.05% and +10.40%.

+4.13%Unfiltered FanDuel LightGBM.

+7.17%Unfiltered DraftKings LightGBM.

+10.05% / +10.40%Same-side agreement at each book.

+4.13%

FanDuel LightGBM

425 settled total bets produced +$1,753.
Sharpe 0.48 · geometric Martin 0.78.

+7.17%

DraftKings LightGBM

537 settled total bets produced +$3,852.
Sharpe 0.87 · geometric Martin 1.92.

Why this matters

Separate models. Separate prices. Approximate 95% intervals still crossed zero.

FanDuel about -4.96% to +13.21%. DraftKings about -0.80% to +15.15%. Promising, not proof.

Main story

Unfiltered FanDuel Tier-1 stepped profit and drawdown paths across native and derived markets

This retained FanDuel path is a book-local comparison. It is historical and post-hoc, and it does not authorize capital.

Book-local path

Unfiltered DraftKings Tier-1 stepped profit and drawdown paths across native and derived markets

This chart is the unfiltered book run, not the same-side filter. The blue native-total line finished near +$4,000. XGBoost’s orange line stayed negative.

Profit path

Unfiltered FanDuel Tier-1 realized ROI across five selected executable edge groups

Unfiltered FanDuel LightGBM: smallest edge group +13.63%, largest -10.85%. XGBoost flipped that pattern.

Confidence warning

Unfiltered FanDuel Tier-1 season-by-season ROI and uncertainty for native and derived markets

Unfiltered FanDuel LightGBM made +12.13% in 2022 and +28.52% in 2023, then slipped just below zero in 2024 and 2025. Same-side agreement did not fade.

The fade

Unfiltered DraftKings Tier-1 season-by-season ROI and uncertainty for native and derived markets

Unfiltered DraftKings LightGBM returned +1.28%, +6.17%, +6.66%, and +12.38%. The uncertainty still crossed break-even.

Steadier, not proven

BlackArbs LLC · From two book results to one shared signal

The next question

What if both models had to pick the same side?

FanDuel faded. DraftKings held up. The next comparison kept only games where both book-specific models agreed on direction, while preserving each book's own line and price.

2Book-specific models.

1Shared direction.

2Separately settled prices.

FanDuel model

Fits on FanDuel market features.
Predicts at the FanDuel line.
Keeps the FanDuel price.

+

DraftKings model

Fits on DraftKings market features.
Predicts at the DraftKings line.
Keeps the DraftKings price.

Take the bet only when both models pick the same side.

FanDuel Over 44.5 and DraftKings Over 45 can agree. Each bet is still judged at its own line.

No price mixing

Both books, same side

+10.23%

294 games · 588 book-bets · +$6,012.98
Sharpe 0.94 · geometric Martin 3.74.

FanDuel prices

+10.05%

294 bets · +$2,955.50
Sharpe 0.91 · Martin 3.34.

DraftKings prices

+10.40%

294 bets · +$3,057.48
Sharpe 0.97 · Martin 4.12.

Same side, two prices. Opposite-side and one-book selections are not in this number.

Main finding

Combined FanDuel and DraftKings agreement portfolio: stepped equity and drawdown

FanDuel returned 10.05% with an 8.73% maximum drawdown; DraftKings returned 10.40% with a 7.84% maximum drawdown. The combined view returned 10.23% with an 8.21% maximum drawdown. Both legs contain the same 294 outcomes, so this is not substantial outcome diversification or a locked 2026 test.

Portfolio view

BlackArbs LLC · Research conclusion

The next clean test

Run both book models separately. Take only totals where they agree on the side. Settle each bet at its own book.

Same-side LightGBM was stronger than either unfiltered book, and it stayed positive every season. That comparison was found after seeing the tape. It still needs a locked 2026 test.

2Independent book models.

1Shared same-side rule.

0Permission to treat a historical result as guaranteed.

This Model Beat FanDuel AND DraftKings on NFL Totals. Then the “Best” Bets Lost Money.

Profit appeared independently at both books.

FanDuel LightGBM

DraftKings LightGBM

Why this matters

FanDuel's unfiltered path was weaker than DraftKings' path.

DraftKings LightGBM built profit across the full test.

A bigger claimed edge did not mean a better bet.

Without the agreement filter, FanDuel faded.

Unfiltered DraftKings stayed positive in every season.

Same direction. Separate models. Separate prices.

FanDuel model

DraftKings model

Agreement made the profitable candidate stronger.

Together, the two book legs ended at $26,013.