How It Works

How We Taught a Machine 50 Years of Kiwi Horse Racing Intuition

Every good punter develops a sixth sense — knowing when a track bias has shifted, which jockey owns a particular course, or when a trainer is about to strike. We turned those instincts into data, and taught a machine to see them all at once.

1. Class Relief Modelled Through Domestic Rating

A horse drops back from open company to rating 75 and suddenly looks impossible to beat. Every punter scans the form for this — but class relief isn't binary. A horse coming back from open class is different from one dropping from rating 85 to rating 75.

We baked the horse's domestic rating directly into the model — the same NZTR rating that determines what grade the horse races in. We also computed its rating versus the field: the difference between this horse's rating and the average rating of every other horse in the race.

This single feature captures class relief in one number. When a rating-85 horse lines up against an average field rating of 70, the model sees a +15 gap and adjusts the win probability accordingly.

2. Track Bias Changes With the Season

Trentham in autumn rides differently than Trentham in spring, even at the same official track rating. A "Dead 5" in September is not the same as a "Dead 5" in March.

We built a track bias engine that analyses every race's first-400m sectional time, leader win rate, and on-pace strike rate. It tracks this per track and per condition band, using a 60-day half-life so recent meetings carry more weight than months-old form.

The model knows that Cambridge Synthetic in June rides like a Firm-Dead track — not the same as Heavy 10 Trentham in winter.

3. Recency-Weighted Jockey Performance

Some jockeys just suit certain tracks. A form guide might tell you a jockey's lifetime strike rate — but what you really want to know is: how have they been going lately at this specific track?

We compute a jockey's lifetime win rate at each track, jockey-trainer combination stats, and a recency-weighted form line using an Exponentially Weighted Moving Average (EWMA). A jockey who has been riding winners lately gets a much bigger lift than one whose last good run was 18 months ago.

4. Continuous Distance Profiling

You see a horse stepping up to 2000m for the first time. If it has the breeding for a stayer and closed off strongly over 1400m last start, you know it's ready.

We created distance-profiling features that look at a horse's entire career — its win rate in each distance band, its best (peak) winning distance, and a distance-specialist score measuring how much better it performs at today's distance than it does overall.

The model doesn't just check "has this horse won at 2000m before?" It computes a continuous distance-aptitude score from every start at every distance.

5. Trainer Profile Features

A trainer who hasn't had a winner in 60 days is either out of form or out of ammunition. But a blanket "cold trainer" label misses the nuance.

Every horse gets nine trainer profile features: career win and place rates, track win rate, distance win rate, class-level win rate, seasonal strike rate, and recency-weighted form. The model treats a trainer who is 3-from-8 in the past month very differently from one who is 3-from-60.