About PaddocksEdge

    A data-driven second opinion on every UK and Irish race card, built so you spend less time reading form and more time on the runners that matter.

    The short version

    Every runner across today's UK and Irish cards is analysed by a calibrated machine learning model trained on 18 months of historical racing. The model considers 669 data points per runner - form, conditions, connections, breeding, race context, layoff windows - and produces a real probability of winning for each. Where the probability is strongest, the runner surfaces as a selection. Everything else stays off the page.

    Where it came from

    We used to do this by hand every morning, pulling apart form lines, checking trainer and jockey records at the course, cross-referencing going preferences, working out which runners were genuinely well-placed and which ones were just popular because of a name on the saddle.

    It was good. It was also slow. On a busy race day those hours add up fast.

    The earlier version of PaddocksEdge automated the way we thought about it manually. It used eight broad factors — form, going, distance, class, trainer, jockey, connections and freshness — each scored independently and combined into a single rating. That earlier model worked, but it also showed us its limits. Eight inputs is a fraction of what is actually in the data, and the weighting between them was something we had to decide by hand rather than letting the data tell us.

    V4 is what happens when you start over with proper machine learning, proper historical data, and proper validation. The model has learned for itself, from 196,000 runners across 19,600 races, which patterns matter most. The result is a calibrated probability model that knows things we didn't know to look for.

    What the model considers

    The model analyses six broad categories of data for every runner, every day. Each category contributes hundreds of individual data points. The relative importance of each is not decided by us - it's learned from the data.

    Form Patterns

    Recent finishing positions, weight-adjusted ratings, improvement trajectories, and form trends across each horse's career. Recency matters; the model weights newer runs more heavily, but the exact weighting was learned from the data, not assigned by hand.

    Conditions

    Every runner is analysed against today's specific conditions - going, distance, class, course, race type. The model has identified which horses thrive in which conditions across thousands of examples. A horse who has won on heavy ground three times is treated differently to one who has avoided it.

    Connections

    Trainer form, jockey form, and the trainer-jockey combination effect across multiple time windows. Course-specific records. The model picks up subtle changes in form across these signals that a single statistic would miss.

    Breeding

    Sire and damsire patterns across thousands of offspring. Some breeding lines excel at specific distances, age groups, or going types. The model has learned these patterns from results, not pedigree guesswork.

    Race Context

    The shape of today's field. Field size, depth, race type, and the relative position of each runner against the others on the day. A 6/1 shot is a different proposition in a field of eight than in a field of eighteen.

    Time Since Last Run

    Days since last run, mapped to performance patterns at different layoff lengths. Quick returns, normal breaks, and long layoffs all carry different signals depending on the horse.

    The model combines these signals using weights it has learned from 18 months of historical results. We do not publish the exact internal weights - that is the part that took longest to refine - but the inputs are listed openly here.

    When selections release

    Selections are updated from 10:30am UK time, refreshing automatically as declarations come in throughout the morning. Most days that means a small handful of qualifying runners. Some days, none. We would rather show nothing than surface a runner the model is not confident in.

    What "calibrated probability" actually means

    When the model says a horse has a 35% chance of winning, the data shows that roughly 35% of similar runners actually win. This is the hidden discipline behind everything PaddocksEdge does.

    Most racing services give you a score, a star rating, or a confidence figure with no way to verify what it means. PaddocksEdge gives you a probability you can audit. The track record proves the probabilities are honest.

    Full transparency

    Every selection is tracked. Every outcome is published. The hits and the misses, side by side, on the Performance page. No deleting bad weeks. No cherry-picking the good ones. If the model gets a call wrong, it shows.

    That is the deal. Full transparency, every day.

    Keep exploring

    Hand-picked articles and pages to take you deeper.