Get Early Access
Methodology

Why 97.8% AUC Is Real

Understanding what this number actually measures.

Different Sports, Different Statistics

Most AI sports prediction models tackle a binary outcome: Team A versus Team B. With a coin-flip baseline at 50%, the strongest team-sport models in the world land somewhere between 55% and 70% AUC. That's a hard problem, and those are honest numbers for it.

RaceHP solves a fundamentally different problem. The question isn't whether one of two teams wins — it's which competitor wins from a field of 8 to 20+ runners. The model's job is to rank the actual winner above the entire field. With rich, structured data on every competitor — form, fitness, conditions, historical performance — separating winners from non-winners is a statistically different problem than picking between two football teams.

Comparing our AUC to a team-sport AUC is like comparing a sprinter's time to a swimmer's time. Same unit of measurement. Completely different event.

Why Higher AUC Is Expected

In a typical 12-horse race, one horse wins and eleven do not. Many of those eleven are clearly outclassed — a first-time starter is not beating a Grade 1 stakes winner. AUC measures the model's ability to rank the actual winner above non-winners. When most non-winners are obviously weaker competitors, that ranking task produces higher scores than picking between two evenly matched football teams.

Every prediction model in the world watches one sport. URIN watches five. Not because more coverage is better — because the patterns transfer. How a dominant horse separates from its field follows the same mathematics as how a leading car pulls away from the pack. Pace, conditions, field strength, competitor separation — the data tells the same story whether the competitor has four legs or four wheels.

15.8 million samples. Five disciplines. One neural network that sees connections no single-sport model ever could. That’s what 97.8% represents.

Random Chance
50%100%

AUC measures ranking accuracy. 50% = random guessing. Drag to explore.

Pick any winner and any non-winner from any race. AUC is the probability that the model ranks the winner higher.

At 50%, the model is guessing — a coin flip with no predictive power. At 97.8%, URIN identifies the correct ranking 978 times out of 1,000.

That gap represents the training data, 144 features per competitor, and transferable patterns learned across five racing disciplines.

Common Questions

Isn't 97.8% AUC a sign of data leakage?

Data leakage means the model accidentally sees the answer during training. Our model uses only pre-race information — speed figures, class ratings, post position, historical performance. No finishing times, no margins, no post-race data. And the definitive proof: every prediction is SHA-256 locked before the race starts. A model with data leakage cannot produce correct predictions on races that have not happened yet.

Isn't this just overfitting?

An overfit model memorizes training data but fails on new data. Our model has been verified live across hundreds of races in five sports since early 2026 — on events the model had never seen. Overfitting does not produce consistent results on unseen live events across five different racing disciplines.

How is this different from other sports AI?

Most sports AI platforms build one model for one sport. RaceHP built a single neural network that processes horse racing and motorsport simultaneously — five disciplines under one architecture. The cross-domain training means the model learns patterns from 15.8 million samples that single-sport systems structurally cannot detect.

The Proof Is in the Predictions

Every RaceHP prediction is SHA-256 hashed and timestamped on the Bitcoin blockchain before post time. The hash cannot be altered after the fact. Nobody — including us — can go back and change what was predicted.

33 / 35
94.3%
Top-3 in Horse Racing
11 / 11
100%
Perfect card at Aqueduct — May 9, 2026
13 / 14
92.9%
Kentucky Derby Day
5 / 1
Five sports · one network
Verified in real time across every discipline

No amount of overfitting or data leakage produces correct predictions on live races that haven't happened yet. The blockchain proves the predictions existed before the outcomes were known.

View Full Track Record →