How to build a greyhound betting system that isn't just curve-fitting

Anyone with a few million historical results and a filter tool can find a rule set that made money in the past. The skill is telling the difference between a real edge and a shape you drew around random noise. Here is a method that keeps you honest.

1. Start with a reason, not a screen

The reliable way to curve-fit is to try hundreds of filter combinations and keep whichever one shows the best number. You will always find one. It usually means nothing.

Instead, start with a hypothesis you can say out loud: "dogs dropping a grade are over-faced less often, so the market under-rates them", or "wide runners at short sprint tracks are systematically slow away and over-bet". Then test that specific idea. If it fails, you have learned something. If you skip this step, a positive result tells you nothing, because you were always going to get one.

2. Make sure the sample is big enough to mean anything

A 3% return over 400 runners is indistinguishable from luck. As a rough guide:

• Be sceptical of anything with fewer than ~1,000 qualifying runners. Prefer several thousand.
• Price matters as much as count. A strike rate built on backing 1.4-shots stabilises far faster than one built on 12.0-shots — a handful of big-priced winners can swing a small-sample lay or back system entirely.
• Look at how the P&L is distributed. If removing the ten best results turns the system negative, you don't have a system, you have ten lucky days.

3. Test out of time, every time

Split the archive. Build and tune on the older years, then look at the recent years once, without touching the rules.

• Build on 2020–2023, confirm on 2024–2026.
• If the edge only exists in the years you built on, it is curve-fit. Bin it.
• SportsEdgePro shows year-by-year P&L for every filter, so you can see at a glance whether a rule works in five separate years or leans on one freak season.

4. Treat every extra filter as a cost

Each condition you add is a degree of freedom, and each one can only ever improve the historical number — that's what makes stacking filters so seductive and so dangerous.

Two or three conditions with a clear rationale beat seven that "optimise" the back-fit. If you can't explain why a filter should work, it is decoration, and decoration is where curve-fitting hides.

A good test: remove each filter in turn. If taking one out barely changes the result, it was doing nothing but shrinking your sample — drop it. If taking one out collapses the edge and you have a reason for it, keep it.

5. Judge by drawdown and shape, not headline ROI

Two systems: one returns +8% but includes a 45-point losing run over three months; the other returns +4% and never draws down more than 12 points. The second is better, because you will still be betting it when the first has been abandoned in disgust.

Before you trust a rule set, look at:

• Longest losing sequence — in points and in time. Could you sit through it?
• Month-by-month P&L — is it a grind, or three good months carrying nine flat ones?
• Turnover — a 10% ROI over 60 bets a year is not a living; a 2% ROI over 8,000 bets a year might be.

6. Model the costs and keep staking boring

Settle everything at Betfair BSP, net of commission. Evaluate at level stakes — variable staking can flatter or wreck a backtest for reasons that have nothing to do with whether the selection method works. Once the method is proven, then think about staking.

A workflow you can repeat

1. Write the hypothesis in one sentence.
2. Build the broad filter that expresses it. Check the qualifier count and the price spread.
3. Split the P&L by year. Reject anything that only works in-sample.
4. Remove every condition that has no rationale or no effect.
5. Check the longest losing run and the monthly shape.
6. Run it forward on today's cards with Bet Finder and paper-trade for a few weeks before a penny goes on.

Most ideas die at step 3 or step 5. That is the process working, not failing.