Using AI as a Horse Racing Handicapping Tool: What It Actually Does

FYI - This application is not for sale. The project is strictly for research purposes and will not be sold to the public. I will gladly answer any questions

There has been a lot of noise about artificial intelligence and what it might mean for handicappers. Most of it falls into two camps: breathless hype about AI picking winners, or skeptical dismissal from experienced players who have watched too many systems come and go. This piece is meant to land somewhere more useful — an honest look at what an LLM-based handicapping application actually does, what its genuine advantages are, and where its limitations still lie.

What the Application Is
This is a browser-based handicapping tool that connects to a large language model (LLM) through an API and is built specifically around Premium past performances. The workflow is straightforward: a user loads the day's PP data, the LLM reads and extracts the relevant figures for each horse, and those figures are pre-loaded into a structured handicapping interface. From there, the user works through races one at a time — running a simulation, reviewing AI-generated analysis, and assembling wager selections.
How
This Differs from Typical AI Racing Tools
Most AI tools marketed to handicappers are either rule-based systems dressed up in AI language, or black-box neural networks trained on historical racing data whose logic cannot be inspected or questioned. An LLM sits in a different category: it is a reasoning engine that can be given a specific framework — pace figures, Prime Power ratings, speed consistency — and apply that framework to the data in front of it, producing an explanation of its conclusions in plain English.
That transparency is the key difference. When a black-box model tells you to bet a horse, you have no way to understand why or disagree with specific reasoning steps. When an LLM produces a handicapping opinion, it produces prose that can be read, interrogated, and pushed back on. A handicapper can note that the AI missed a troubled trip or that the pace projection ignores a likely scratch, and that context can be fed back in. The conversation is iterative.

Because the LLM holds multiple factors in mind simultaneously — surface and distance switching, trainer patterns, pace scenario adjustments for a specific field shape — it handles nuance that pure statistical models struggle with.

The Simulation Engine and Bet Sizing
The application includes a Monte Carlo-style simulator that runs multiple iterations to produce win, place, and show probabilities for each horse. Prime Power carries the heaviest weighting, with speed figure consistency and pace scenario making up the remainder. The weights are adjustable through sliders, so handicappers can tune the model to reflect their own methodology rather than accepting a fixed formula.

Kelly Criterion bet sizing is integrated into the wagering module. Given a horse's simulated win probability and current odds, Kelly outputs a mathematically defensible fraction of bankroll to commit. The math is real — but the output is only as good as the probability estimate going in, which depends entirely on the quality of the handicapping inputs.