TrapMetricsThe System Builder (and why there is no button that finds the best rule)
Sixty-four elements, five years of results, and a back-tester built to refuse to flatter you: holdout, walk-forward, intervals, Trim, Frontier, staking, and no Goal Seek by design.
Every back-tester can make a rule look good. Add a filter, watch the return climb, add another. Five filters later you have a system that made 40% on the past and will make nothing on the future, because you found it by searching the past for the thing that happened to work. The greyhound betting world calls this a system. Statisticians call it fitting the noise.
We built ours to make that harder to do to yourself.
What you can build with
Sixty-four elements in eleven groups, every one knowable before the off. Race conditions: track, distance, grade, trap, going, field size, month. Track geometry: circuit size, run-up to the beam. The dog: our speed rating, early and finishing speed over its last three runs. Connections and breeding. Then the model's own measurements, opened up as rules: times and splits against par over the last one, three and six runs, form rank in the field, career record, days since the last run, win rate when clear and when in trouble, early pace, slow starts, class moves, trainer form, litter-mates' record and more.
The trap element is the one most people reach for first, so it pays to know what the record says before you build on it. At Romford, trap 1 wins 19.0% of races against a fair share of 17.3%, +1.8 points on 7,048 runs over two years, which is a clear bias; trap 6 sits 1.3 points under its fair share, which is inside noise. Every track page carries the same table, with the noise called noise, so a rule built on a trap can start from a measured number rather than a hunch.
On the model era, since June 2026, you can also build with the model's own view: its win chance, its rank, its value against the Timeform and Racing Post forecasts, and the named selection classes.
Every rule is tested against the official record, 1,512,527 runs since June 2021, settled flat at the Betfair closing price after commission.
What it refuses to hide
The holdout. The most recent year is scored separately and by default. If your rule is green on the development window and red on the holdout, the page says the rule is fitted to the past.
The walk-forward. The same rule scored in every period, half-year by half-year, each with its own interval, and a column that answers the harder question: if you had adopted this rule at the start of that period, what would you have got from then on. A rule that paid in four of twelve half-years is not an edge, whatever the total says, and the verdict chip says so.
The intervals. Every return carries a bootstrap interval and a probability that the whole thing is luck. A return whose interval spans zero is shown that way.
Since saved. Save a system and from that day its record is scored on its own. It is the only window hindsight cannot have shaped, and once it has thirty bets it is the number to trust. Edit a rule and the window resets, because the record belongs to the rule you wrote down.
What it offers instead of a search
Trim lifts each rule in turn and scores the bets it excludes against the bets it keeps. If the excluded bets return the same, the rule is not doing work and can go. If they return clearly less, keep it. Trim only ever removes or loosens a rule. It never adds one.
Frontier takes one numeric rule and walks it across its own thresholds, drawing the return on the development window as a solid line and the holdout as a dashed one. A threshold that peaks on the solid line and sags on the dashed one was found, not discovered.
Staking shows what the record would have done to a bank under level stakes, a percentage of bank, and fractional Kelly, with two hundred reshuffles of the bet order to show how much of the drawdown is luck. It is labelled an illustration, because it is one.
The starters, and what they show
The opening page offers starter systems to load and change. Two are the platform's own signals with real records. Two are labelled as found by searching, so you can see what a good-looking curve looks like and then watch its since-saved window. The rest show the elements at work, and most of them show the honest truth of the official record: an actual-over-expected near 1.00, which means the market already prices that idea in. That is not a failure of the tool. It is the tool working.
The rule we will not break
There is no Goal Seek. No auto-build. Nothing that searches rules against outcomes on your behalf. Searching the past for what worked is how we produced our own mirage in July, a system that made 40% until the day we started following it. The Builder exists so that you can test an idea you had for a reason, and be told the truth about it.
Questions this piece answers
Why is there no Goal Seek?
Searching the past for the rule that happened to work produces a system that fits the noise. We produced our own that way as a test, a system that made 40% ROI until the day we followed it. The Builder tests an idea you had for a reason and tells you the truth about it.
What is walk-forward testing?
The same rule scored in every period of the record, half-year by half-year, each with its own interval, plus a column for what you would have got from that period on. A rule that paid in four of twelve periods is not an edge, whatever the total says.
What can I build rules with?
Sixty-four elements in eleven groups, all knowable data points before the off: race conditions, track geometry, our speed rating, connections and breeding, and the model's own measurements such as time against par, form rank, days since the last run and trainer form etc.
What is the since-saved window?
Save your own system, and from that day its record is scored on its own. It is the only window hindsight cannot have shaped. Once it has thirty bets it is the number to trust.