Top-pick record
How often the first study selection matches the recorded winner, with the race count beside it.
Horse-Oracle Analytics
Horse-Oracle performance records with named comparisons, race counts, time windows, and uncertainty.
The record so far
Four figures, read from the latest dated receipt.
What Horse-Oracle has measured
This page is not a statistics glossary. It is a reading guide to the specific results Horse-Oracle has earned, the comparisons that remain inconclusive, and the signals that are not allowed into the live model.
[VERIFIED] · 1,785 held-out races
Analytics scorecards
Calibration
Whether stated probabilities behave like probabilities, not whether a single race went the expected way.
Against a baseline
Every comparison names its control—often the morning-line control—so a number has a fair reference.
What the sample covers
Dates, races included, exclusions, and limits stay next to the result.
Probability quality improved
The probability view improved log loss by 0.0079 versus the overround-removed market reference across four expanding walk-forward folds. This measures the quality of the percentages, not whether the top horse won more often.
[VERIFIED] · 1,048 accepted-set races
The late tote beat the model
The model top choice won 29.3% versus 37.5% for the T-5 tote favorite. Difference: −8.2 percentage points; 95% interval: −11.3 to −5.1. The late market is the required opponent for future claims.
[CLAIMED] · replication pending
BRIS Prime Power stays passive
The research signal was positive in a 1,538-race full-field BRIS set with 1,048 out-of-sample races, but September coverage was zero and the signal is not in the live feature set. It is a research candidate, not a live advantage.
[INCONCLUSIVE] · 2,852 paired races
Competitor comparison is close
The limits are part of the result
- A probability improvement is not automatically a winner-rate improvement.
- An earlier market comparison cannot be substituted for the T-5 tote.
- A candidate signal cannot be called live when coverage disappears.
- A confidence interval crossing zero is reported as inconclusive.
- Paper research does not prove betting profit.
PAPER MODEL — NOT BETTING ADVICE · $0 STAKED