Trusting the test
Read the verdict like a pro
Proof is the lie detector. Before a single dollar goes to real money, it replays history walk-forward — train on the past, test on a future it never saw — and asks one question: did this model actually beat the closing market out of sample? It hands back a verdict. Here's how to read it.
The four verdicts
Out-of-sample pari-mutuel ROI is positive with the confidence interval clearing zero. This is the only verdict that unlocks real money. Real, paid, repeatable.
Positive signal but not yet significant — or ML→close drift with flat ROI (the line moved; the pool still paid the close). Keep it on practice and gather more races.
It does not beat the close. No threshold or tweak fixes this — the features need work. Do not bet.
Too few settled bets to judge. The honest answer is "we don't know yet." Get more data.
The numbers behind the verdict
The bottom line — profit per $1 if paid at final odds. This is what makes EDGE real. Watch the CI: the lower bound must clear 0.
Morning-line movement toward the close. Not ticket CLV. Does not cash in the win pool. The field name is mlDriftToClose.
Can the model rank winners as well as the market? 0.5 = coin flip. The closing market is ~0.77 — that's the wall.
When we say 20%, do they win 20%? Low ECE = honest probabilities. High = overconfident (the classic failure).
Why ML→close can be green while ROI is red
You'll see it: drift +12%, ROI −2%. Not a contradiction. The morning line moved toward the close; you are still paid the close. That number is not CLV. True CLV needs a tote or ticket price you actually booked.