HOW IT WORKS

Every morning we gather today’s racing cards: who’s running where, the race conditions, barrier/stall, jockeys, weights etc.

We then bundle this with each runner’s career history: strike rate, place rate, race class, distance performance, form and recent form. In total, about 25 features per horse are fed into our ML models.

We use machine learning models used across competitive ranking problems and modern tabular prediction to rank the statistical chances of each horse. Rather than guessing each horse’s win probability in isolation, it learns from the order finishers came in across thousands of races and produces a ranking for the whole field, most-likely winner at the top, working down.

The models are retrained every Sunday on the latest results, so it never goes stale. UK and AUS each have their own model. The racing is too different to share one.

Every pick on the page carries a 1–5 banana rating. More bananas means the model is more confident — that’s the only signal. Five bananas means the model is as sure as it gets; one banana means it has a view but the race is genuinely open.

The clever bit: the top pick doesn’t automatically get 5 bananas. The banana count for the highest pick is anchored on the model’s absolute confidence, not just its rank. That means a clear favourite in a tight race lights up differently than a clear favourite in a flat, low-conviction race. From there, each lower-ranked horse steps down by one banana.

UK and AUS each use their own confidence thresholds (UK races are about twice as “hot” on the calibration scale at every quantile), so a 4-banana pick means the same thing in both regions: an above-average level of conviction by that region’s historical standards.

One-banana races — we defer to the market. When the model has no real edge in a race (a single banana), our pick historically does no better than the crowd’s. So in both UK and Australian racing, a one-banana race switches the headline selection to the market favourite, and the model’s own picks slide to 2nd and 3rd. We back the model when it’s confident and step aside to the market when it isn’t. It’s a higher-strike-rate call, not a promise of profit — short-priced favourites win more often but rarely beat the bookmaker’s margin.

What you see on the predictions page are top picks: the model’s view of who’s most likely to win. That isn’t the same as a value bet.

A value bet is one where the price on offer is higher than the true probability justifies: finding a horse the bookies have underrated, not just the most likely winner. The two often overlap; they’re not the same thing. A short-priced favourite can be the model’s top pick AND terrible value at the same time.

The page tells you who Punt Monkey thinks will win. What you do with that (back it, lay it, dutch it, leave it alone) is your call.

• The top 3 picks for every race. That’s it. We don’t flood the page with the rest of the field.
• Each pick’s banana rating and a win probability.
• Bookmaker deeplinks, so you can compare odds quickly. We don’t place bets ourselves.
• Daily refresh: UK picks at 07:00 BST, AUS at 20:00 BST.

• We don’t place bets on your behalf. Punt Monkey is a predictions service, not a bookmaker.
• We don’t store your stakes, your wins, or your losses.
• We don’t sell your data. See our privacy policy for the full picture.
• We don’t pretend the model is always right. Even at our best venues, the top pick wins fewer than half the time. Treat the bananas as an edge, not a guarantee.

Punt Monkey is for adults aged 18 and over. Predictions are for informational purposes only. They are not financial or betting advice. Past performance does not guarantee future results.

If gambling stops being fun, please reach out for support: BeGambleAware.org (UK) or GamblingHelpOnline.org.au (Australia).