ML Systems vs Dog Selector Systems — Which One Should You Use?
RateThatDog supports two kinds of betting system. Dog Selector systems filter on live runner data; ML systems filter on the model's prediction features. Here's the difference and when each one wins.
What's the difference between ML and Dog Selector systems?
Both kinds of system save filters and track P&L. The difference is what they filter on. **Dog Selector (DS) systems** filter on live runner data — form figures, trainer, trap, grade, recent winning times. **ML systems** filter on the model's prediction features — composite score, suitability components, field speed rank, first-bend rank.
DS systems answer the question "what does this dog look like on paper today?". ML systems answer the question "how does this dog rank against today's field on the metrics that predict winners?". Different questions, different filters, sometimes very different selections.
When should I use a Dog Selector system?
When your hypothesis is about real-world facts: "in-form trainer at this track", "dog dropping in grade", "won at this distance recently", "top of the speed column". DS systems are the right tool when you're encoding a piece of human horse-sense — something a tipster would say at the track.
They're also better for narrow scenarios where the ML feature set doesn't cover the angle. "Trainer X with dogs over 480m at Hove" is a clean DS filter; it would be awkward to express through ML composite ranks.
When should I use an ML system?
When you want to bet on the model's confidence rather than a hand-built rule. "Top composite rank in race AND gap to next ≥ 8" is the kind of filter that surfaces dogs the model is genuinely backing. ML systems also let you stack rank filters — top speed AND top suitability AND top first bend — which DS systems can't really do.
ML systems are also the right tool for backtesting. The platform stores 188,000+ pre-race ML snapshots, so you can run a new ML filter set against history and see exactly what the strike rate would have been before you go live.
Which one has the better strike rate?
Both can hit. The validated composite-60 Hot Dogs filter — an ML system in shape — hits 28.34%. Strong DS systems around in-form trainers and trap-specific patterns can hit similar rates over their (typically smaller) qualifying samples. The right comparison isn't ML vs DS in the abstract; it's whether your specific filters express a real edge.
Two practical points. ML systems tend to produce more picks (the rank filters apply across every race). DS systems tend to produce fewer, more concentrated picks (the rules are usually venue- or trainer-specific). Choose based on what you want — broad coverage or tight conviction.
Can I combine ML and DS filters?
Yes. The Dog Selector lets you stack ML-style rank filters (composite, speed, first bend) on top of DS filters (trainer, grade, distance) within a single saved system. The combined approach often gives the cleanest result — "top-3 composite rank in race AND trainer in form AND trap 1 at Hove 500m" is a stack that a pure ML or pure DS system would each only half-cover.
The cost of stacking is sample size. Every filter cuts the universe of qualifying dogs. Aim for 1-5 picks per day and let the system run.
Frequently asked questions
Are ML systems on ratethat.dog better than Dog Selector systems?
Neither is inherently better. ML systems are stronger when you're betting on the model's confidence and want to backtest against 188,000 historical snapshots. DS systems are stronger when you're encoding a specific human-readable rule like in-form trainers or trap-specific patterns.
Can I backtest a Dog Selector system?
Partially — the platform tracks DS systems forward from when you save them, but most of the deeper backtesting features (against the 188k+ snapshot dataset) are designed for ML filter shapes.
What kind of filter is composite score?
Composite score is an ML feature — it's part of the model's prediction output. So filters using composite (e.g. composite ≥ 60) are ML system filters. Both kinds of system can use it.
How many filters should I stack?
Two to four works well. Each filter cuts the qualifying sample, so over-stacking leaves you with too few picks to learn anything.
Where do I see all my saved systems?
On the Systems page. Both DS and ML systems live in the same table, with strike rate, place rate, ROI and daily breakdowns for each.