Methodology & Backtesting

Raw sportsbook prices include the vig (typically 4-6%). We strip it using the Shin (1993) insider-trading model instead of naive proportional devigging, because Shin better accounts for informed money on the underdog.

p_true = (√((z·(1-z))² + 4·(1-z)·p_book²) - z·(1-z)) / (2·(1-z))

Full Kelly is theoretically optimal but too volatile for real bankrolls. We apply a quarter-Kelly fraction and further shrink stakes when our model's recent Brier score suggests miscalibration.

stake = bankroll · 0.25 · (b·p - q) / b · shrink(brier)

Books price same-game parlays at the independent product. We compute the true joint probability using the standard bivariate correlation adjustment:

P(A∩B) = pA·pB + ρ · √(pA(1-pA)·pB(1-pB))

Positive ρ (QB pass yds + WR rec yds) means the book is under-pricing. Negative ρ (opposing team ML) means over-pricing.

Backtests are trivially cherry-picked. Instead, every AI pick is logged to pick_price_log at the moment it fires, then auto-graded against real ESPN outcomes. The public track record shows every graded pick from the last 7-30 days — wins, losses, and CLV.

We report CLV (Closing Line Value) in addition to win rate because CLV is a leading indicator: beating the close consistently predicts long-term ROI even during variance-driven cold streaks.

• No bet placement, no sportsbook account linking, no wagering.
• No cherry-picked "best of" records — every pick is logged pre-outcome.
• No hidden model params — Shin z, Kelly fraction, and ρ matrix are all above.
• No fake volume — visitor and pick counts are real database queries.