Fair odds by pure mathematics — the Ghost model's win probabilities, calibrated and rebuilt into a 100% book. The tote board with the takeout removed.
| # | Horse | Ghost Win % | Win (fair) | Place | Show | Morning line | Edge vs market |
|---|
The Ghost / Gallant model outputs a raw win probability for each contender. Left alone it's overconfident — recent favorites carried ~0.97 but won ~0.30.
A Platt correction fit on realized results pulls probabilities toward honesty: p' = σ(a·logit p + b), read live from einstein_calibration.
The rated field is rescaled to sum to exactly 100%. No overround, no takeout — a true probability, not a market price.
Win odds are 1 / p. Place, Show, Exacta and Trifecta all fall out of the same win probabilities by the Harville chain — P(A→B) = pₐ · p_b/(1−pₐ).
Reading the board. "Fair" odds are what the math says the bet is worth with zero takeout, so they'll usually look shorter than the tote. Edge vs market compares the Ghost fair probability to the morning-line implied probability: a positive edge means the model rates the horse more likely than the crowd is pricing — where value lives. Toggle Raw model to see the numbers before calibration.
Source (live). Ghost/Gallant win probabilities from starpreds.agent_predictions (status=current) · calibration einstein_calibration · program numbers + morning lines from starraces.race_entries. The model rates its top 5 per race; unrated runners are shown for the full field but excluded from the fair book. Fair-odds math is illustrative, not a wagering guarantee.
Ghost True Odds is a restricted prediction board. Sign in as an authorized user to view.