How it works
Powerful modeling under the hood, a clear picture on top. Here's the idea, without the homework.
Advanced neural-network analysis
At the core is a deep neural-network model trained to rate every horse in a race against the rest of the field, producing a calibrated probability of winning.
Models for pace dynamics
Races are won and lost on pace. Dedicated models read the likely shape of each race: who wants the lead, how much early pressure there is, and how that sets up the run home.
Projected pace collapses
A separate model flags races where the early pace looks likely to collapse: the scenarios that quietly hand the race to closers and reshape the finish.
An ensemble, not a single guess
Rather than trusting one model, we combine several into an ensemble. Agreement between them is a stronger signal than any one model alone.
Calibrated and backtested
Outputs are calibrated so a stated probability means what it says, and every claim we make is measured against large historical samples using cross-validation.
Turned into clear, usable picks
All of that becomes something simple on your screen: win probabilities, best-bet highlights, and backtested strategy ideas you can take, leave, or adapt.
What the backtests say, and what they don't
Before launch we tested the model against three seasons of historical data, 2023 through 2025. Those tests produced the numbers below. Backtests are a controlled experiment: useful, honest about method, and easier than the real world. That is why they live on this page with full context instead of leading the homepage, and why the numbers we lead with are the ones we grade in public, live, after we publish.
Across three seasons of races the model never trained on, the winner was one of our top three picks 73% of the time, and our top pick 36% of the time.
Horses we rated 40% or better won 50.5% of the time, about 3.5 times the average runner's chance.
A flat bet on every horse gives back about 26 cents of every dollar to the track. A flat bet on our top pick gave back about 13. Nobody beats the takeout by picking winners alone, which is why we flag edges instead.
About a quarter of all starters get under a 5% chance from us. Together they won fewer than 1 race in 20.
The count grows every night, and every one of them is on the calibration page.
Except for the live race count, figures reflect historical backtested results on races the model never trained on. Past performance does not guarantee future results. Provided for informational and entertainment purposes only.
Ready to handicap with an edge?
Start free on a selection of races, then unlock every track.