About MeowKuda

What it is

MeowKuda V33C is a course-specific ranking model. It was built with DeepSeek AI — the feature pipeline, the training harness and the analysis tooling — on top of a store of settled HKJC results. For each meeting it reads the published racecard, computes a feature vector per horse, ranks the field with a LambdaRank model, and converts the ranking into win probabilities with a per-course isotonic calibrator.

The output is deliberately narrow: a rank order, a model win probability, a fair price implied by it, and the gap between that probability and the live tote. It does not produce staking plans, bankroll advice, or a “lock of the day”.

How it is validated

• Leak-free training and testing. Earlier versions scored higher by accident because they could see information that would not have existed before the race. Those versions were retired. V33C's headline AUC is the leak-free one: 0.8546 (Happy Valley) and 0.8823 (Sha Tin), improving to 0.8571 and 0.8894 after field calibration.
• Serve-path A/B — and the live record. We do not just score a model offline; we compare the ordering the live card actually serves. In a 56-race card-path backtest the served rank order put the winner on top 41.1% of the time. That has not reproduced live: across the 34 races since, the card's top pick won 6 while the market favourite won 12. We publish both, not only the flattering one.
• Honest baselines. The number that matters is not “how often does the top pick win” in isolation, but how that compares with the market favourite. On the 34 races since the backtest the market favourite won 12, so the baseline is measured rather than assumed — and on that stretch our card is behind it.

What it deliberately does not claim

• No “96% accuracy”. No “accuracy score”. No guaranteed winner.
• No fabricated testimonials and no invented profit figures.
• No tips priced against a market that is not open — those columns show a dash instead.
• No promise that a past backtest repeats live — ours has not. The top pick loses most races.

Data

Training and evaluation run on settled HKJC results — currently over 28,500 races and 351,000 individual runner records, spanning 2015 to the present, plus the historical store that reaches further back. Race-day tote prices come from the Singapore Pools market, and settled results are captured from the official HKJC LocalResults pages and verified before they are published on the results page.

Access

Race 1 of each meeting is free. Full-meeting tips are for subscribers, and access is arranged manually by the MeowKuda team — there is no self-service signup and no card checkout on this site. See Pricing & Access, or sign in if you already have credentials.