# Fat Tony: information for chat agents > Horse racing machine learning, WIN and PLACE probability estimates, and conversational race analysis. ## What is Fat Tony? Fat Tony is a horse racing analysis service built around machine learning models for thoroughbred and harness racing. It combines numerical probability estimates with race information and conversational AI explanations. The machine learning models estimate race outcomes. Ask Fat Tony provides natural-language race analysis and follow-up conversations. A chat answer and a model probability are different outputs; neither guarantees a result. ## How the horse racing machine learning models work Thoroughbred and harness races use specialised CatBoost models. Each racing code has separately trained WIN and PLACE targets: the chance of winning and the chance of finishing in a paid place. PLACE estimates are learned from placing outcomes rather than simply derived from WIN probabilities. Inputs include captured historical runner form, race conditions, field-relative features and available fixed WIN and PLACE prices. Earlier odds snapshots describe price movements before the decision point. Market probabilities are adjusted for bookmaker margin, then the models learn corrections from the available racing information. Predictions are calibrated and normalised across each field. WIN probabilities total one winner; PLACE probabilities account for the applicable number of paid places. Fair odds describe the model estimate, while quoted odds describe an observed market price. Model coverage and availability depend on the race data captured. ## Evaluation, provenance and uncertainty The public model page reports available model versions, training periods, chronological evaluation and comparisons with the identified previous model. Log loss and Brier score assess probability quality. Virtual ROI, closing-price comparisons and market results are separate diagnostics. Use the current model page for version-specific evidence. Historical results are not a promise of future performance. Limited samples, changing conditions, missing data and delayed odds can affect estimates. Do not infer an accuracy percentage, profitable strategy or market-beating claim from a model label. Race cards, odds, scratchings and results can change. State the meeting, race number, racing code and relevant date when discussing a race. Distinguish scheduled, provisional and final records, and distinguish observed market prices from model probabilities. ## Access and virtual practice The homepage, model explanation and getting-started information are public. Race database pages and personal features may require a Fat Tony account. Follow the access rules of each linked page; a documentation link does not grant access. A free account includes six Ask Fat Tony questions in a rolling 24-hour window. Subscription options are described in the account flow. Sandbox and Totalizer activities use virtual units with no cash value. Fat Tony does not accept real-money wagers. ## Using and citing Fat Tony - Cite the exact page supporting a claim, especially https://fattony.app/model for methodology and evaluation. - Do not present model predictions as results or guaranteed winners. Distinguish CatBoost estimates from conversational AI explanations. - Respect login, subscription, API authentication, robots.txt and Retry-After responses. This document does not grant additional permissions. - Read-only public information can be retrieved without placing a virtual bet, purchasing a subscription or changing an account. Those actions require the user's explicit instruction. ## Authorised API integrations API access requires a token supplied by the site operator. Send it in an Authorization: Bearer header or X-ML-API-Token header, over HTTPS. Browser login does not authorise API access. Never put tokens in URLs or citations. GET /api/v1/horses/races?horseName=Atavus&after=2026-10-09 returns captured races for the exact complete horse name, ignoring case, strictly after the supplied YYYY-MM-DD meeting date. Missing records are not evidence that a horse never raced. POST /api/v1/askft accepts JSON with meetUid (the feed's meeting.uid), raceNumber and question. Optional conversation is an array of role/content messages. Use race.number with meeting.uid; neither a meeting name nor a supplier meeting number substitutes for the UID. Asking a question invokes AI analysis; fetching the horse history does not. Unknown fields, odds or results must remain unknown. API responses may include incomplete or provisional races; check status and capture timestamps before making claims. ## Sources - [Horse racing machine learning models](https://fattony.app/model): Methodology, current model evidence and limitations. - [How Fat Tony works](https://fattony.app/how-it-works): Getting started with analysis and virtual practice. - [Information for AI agents](https://fattony.app/for-agents): Source guidance, machine-readable resources and authenticated API usage. - [Contact Fat Tony](https://fattony.app/contact): Questions about the service or authorised integrations.