Algorithm UpdateAdvanced16 Apr 2026· 7 min read

Algorithm Update — Dunstall Park Changes

We've rebuilt the Dunstall Park model from the ground up after a deep dive into the data. Here's what we found, what we changed, and why it matters for your betting.

Why we went back to the drawing board at Dunstall Park

Dunstall Park — Wolverhampton's new venue, which replaced Perry Barr last September — has been one of the trickier tracks in our dataset. Our sprint-distance models settled in quickly, but the standard 480m grades were frustrating, particularly A4 through A6. The picks were losing at a rate that didn't match the venue's overall predictability. So we pulled everything apart.

We went through every race we'd tracked at Dunstall Park: over 1,300 races and 5,500 individual runs, broken down by distance, grade, and trap. What we found was that the features predicting winners at most UK tracks were actively misleading at Dunstall 480m in certain grade bands. That's not a small problem — it meant the model was confidently wrong.

The core finding: suitability was anti-predictive at mid-grades

At A4-A6 — the single biggest segment at 480m, covering over 400 races — our suitability scores were doing the opposite of what they should. Dogs with better suitability records (stronger track, distance, and trap history at Dunstall) were actually less likely to win than dogs with weaker suitability. That's the reverse of what happens at almost every other UK venue, and it was dragging the model down badly.

The explanation is in the grading. At A4-A6 at a newer venue like Dunstall, the field quality is inconsistent — these dogs are less settled in their grade, form is less reliable, and past track experience at this specific venue is sparse for most of them. In that environment, raw speed is a better signal than track history. The model needed to know that.

The solution: a grade-split model

We've built something we've had success with at Romford and Nottingham before — a grade-split model. Instead of applying the same prediction logic to every 480m race at Dunstall, the algorithm now adjusts its approach based on the quality tier of the race.

At **A1**, the top of the ladder, suitability and first-bend ratings still matter — the best dogs are consistent, and their track record tells you something real. The model keeps those signals and is running at 29% strike rate. At **A2-A3**, it's a balanced blend with a small career win bonus layered in. But at **A4-A6 and A7+**, we've stripped out suitability entirely and shifted to a speed-led model with career, rest, and weight multipliers.

Trap bias at Dunstall Park: the numbers

We've also added trap multipliers for the first time at Dunstall Park — previous testing showed they hurt performance, but with the grade-split model in place, they add genuine value. If you're betting at Wolverhampton, these are the numbers to know:

At **480m**: Trap 1 wins 25.7% of races while Trap 6 wins just 18.9% — a 6.8 percentage point spread that compounds over time. At **270m sprints**: inside bias is even stronger, with Trap 1 and Trap 4 both around 31%, while Trap 5 sits at 17.3%. At **660m**: Trap 3 is extraordinary at 38.1% while Trap 4 wins barely 5% — one of the sharpest single-trap biases in the whole UK dataset.

Career, rest, and weight: the new multipliers

At mid-grades and below, we've introduced three new multipliers that the data consistently supported. A dog with a career win rate above 30% gets an 8% score boost — proven winners keep winning, even in inconsistent grade bands. A dog coming back after 6-14 days' rest gets a 6% boost — that's the sweet spot for freshness at this venue. A dog that's lost weight since its last run gets a bump too, because leaner dogs tend to run faster.

The penalties work in reverse: poor career records, long layoffs, and weight gain all cut the score. In our backtest at A4-A6, these three multipliers alone improved Win P&L by over 40 units. That's significant on a segment of just 400+ races.

Sprints and marathons

At the 270m sprint distance, we've flipped the D-grade model (which makes up 91% of Dunstall sprints) from suitability-led to speed-led. The data was clear: pure speed produces a 24% strike rate versus 20.7% for the old suitability approach. For OR sprints the suitability approach still holds, so that's unchanged.

At 660m — the marathon distance — we've rebuilt around trap suitability, which has a +13.23 winner differential at this venue. Course-and-distance intelligence is added on top. This is a race type where the draw and track knowledge matter enormously, and the model now reflects that.

What this means for your betting at Dunstall Park

The headline backtest numbers are a modest lift: 21.9% to 22.1% win strike rate across all 1,345 races, with Win P&L improving by 5.5 units overall. But the headline understates the real improvement — the gains are concentrated in the segments that were bleeding. A7+ at 480m jumped from 29% to 35.5% strike rate over the last 30 days; 660m went from 20.8% to 29.2% all-time.

If you're regularly betting at Wolverhampton: respect the trap bias (Trap 1 at 480m, Trap 3 at 660m), weight speed ratings more heavily in the A4-A6 band, and look for dogs with strong career records coming back after a short rest. The Dunstall Park track page has the current per-distance breakdown and live form.