Seven days, 1,282 races, 296 first-pick winners. The headline number is a 23.1% first-pick win rate — against a random baseline of 16.7%, that's a consistent lift across a large sample. The first-pick place rate across the week is 44.1%, meaning the top-rated dog ran first or second in nearly half of all races.
The most useful number for assessing model quality is the any-top-3 win rate, which sits at 60.1% for the week. That means in three races out of five, the winner came from the three dogs the model rated highest. The any-top-3 place rate is 83.8%.
The week was not uniform. June 5th was the hardest day — 42 wins from 264 races at 15.9%, well below average — while June 9th produced the best rate at 30.2% from 126 races. Saturdays and Sundays with full cards are typically harder days for the model because grade variability and shorter fields create less predictable conditions. Wednesday's 27.9% from 179 races was among the better single-day returns.
What's working well? The model's composite score system — which weights field-relative speed and first-bend pace for standard-distance races — continues to outperform simpler rating approaches on courses with well-established trap data. Where accuracy dipped, it tended to be on Irish tracks with lower sample sizes and less reliable trap bias data. The 23.1% seven-day average is solid. The model is doing its job.