Building your own scoring model

Building your own scoring model: what does that actually mean?

Every runner The Racing Bot scores is marked across 15 measurable signals - things like horse form, trainer form, jockey form, course record, class, going and distance suitability, among others. By default those signals are blended together using the bot's own settings, from Off (barely counted) through to Triple (heavily weighted), to produce a single score per horse. Building your own model simply means changing those dials yourself, so the score reflects what you think matters most rather than the default balance.

This isn't about guessing at a magic formula. It's about testing an idea. Maybe you believe course record matters more for jumps racing than the default weighting gives it credit for, or that trainer form should count for more than jockey form in big handicaps. The Simulator lets you turn those hunches into an actual re-weighted score, then check what would have happened if you'd used it.

How re-weighting the 15 signals works

Each of the 15 signals - form, class, going, distance, course record, trainer and jockey form and the rest - has a slider running from Off to Triple. Off removes that signal from the score entirely. The default sits in the middle. Triple roughly trebles its influence relative to the others. Nothing here changes the underlying data; it changes how much weight each piece of data carries when the signals are combined into one score.

This matters because different signals suit different types of racing. Course record might mean more over jumps at a track with unusual fences than it does in a five-furlong sprint on the flat. Going suitability might matter more in a soft-ground slog than on quick summer turf. Re-weighting lets you build a model tuned to the kind of racing you actually bet on, rather than relying on one set of weights to cover every code, distance and ground condition.

Why backtesting on the full settled record beats picking a good week

It's easy to convince yourself a set of weights works by looking at a handful of recent results. A good run of five or ten races tells you almost nothing reliable - it could just as easily be luck as method. Racing is volatile enough that any set of rules, good or bad, can look brilliant over a short run and terrible over the next one.

That's why the Simulator doesn't test your weights against a cherry-picked week. It replays them against the entire settled archive - currently 15,107 picks across 418 days - and shows you the strike rate your rules would actually have produced across all of that, with a 25% overall winner rate and 50% placed as the baseline record to compare against. Seeing your weighted model's results against a sample that size, spanning well over a year of racing across all sorts of tracks, going and class levels, tells you far more than any short hot streak could.

This is also why the record is kept public and auditable rather than just summarised. A backtest only means something if you can trust the numbers behind it, and if you can go back and check them yourself against every settled pick rather than take a headline figure on faith.

Turning a backtest into a saved bracket

Once you've re-weighted the signals and run them against the settled archive, you'll have a strike rate for that exact combination. If it looks like an improvement worth using - or simply a better fit for the type of racing you follow - you can save that set of weights as a bracket. A bracket is just your rules stored as a named filter, ready to be applied to future racing rather than only tested on the past.

From that point on, the bracket does the sorting for you. It can drive which picks appear on your dashboard, which emails you receive, and which alerts get triggered, all based on the exact weighting you built and tested rather than the site-wide default. You can save more than one bracket if you want different weights for different situations - one for competitive handicaps, another for small-field racing, for example - and switch between them or run them side by side.

Keeping it honest

A backtest, however large, describes what has already happened - it's a measure of how a set of rules performed across the settled record, not a promise about what's coming next. Building your own model is a way of testing ideas against real, auditable history rather than relying on a hunch or a tip, and that's a genuinely useful discipline for any punter. But racing has 25% winners and 50% placed across the whole 15,107-pick record for a reason: most runners lose, favourites get beaten, and no combination of weightings changes that basic fact.

The value of the Simulator is in the process, not in chasing a single higher number. Re-weight the signals, run the backtest across the full archive, save the bracket if it earns its place, and keep checking it against new results as they settle in. That loop - test, save, review - is what turns a model from a one-off guess into something you can actually stand behind, because the whole record it's measured against is there for you to audit at any time.