How to read a tipster's record
We don't sell tips, which makes this easier to write. "Is that record any good?" is one of the most common questions in racing, and the answer is nearly always in what the record leaves out. Here is what to check — including on ours.
1. At what price was it settled?
Every record is a set of selections plus a decision about what price to score them at. That second half does most of the work, and it is usually the half that isn't stated.
The same selections can be settled at:
- the price showing at the moment the selection was published — which may have lasted minutes
- the best price available anywhere at that moment, across every bookmaker
- a price with best-odds-guaranteed applied, so it takes the better of two outcomes
- the starting price — the one almost anyone can get, and the one a reader arriving an hour late would have had
Those are not small differences. A set of selections can show a healthy return on the first basis and a losing one on the last, without a single selection changing. The gap is largest exactly where records look most impressive: short-priced runners whose odds move fast.
Ask: at what price is this settled, and could I have got it? A record that doesn't say isn't a record — it's a claim.
2. How much of it is just variance?
Short runs are mostly noise. A run of selections over a few weeks describes a few weeks; it does not establish that anything is repeatable. Racing produces winning streaks and losing streaks from nothing but chance, and a record is often first published precisely because a streak happened.
The related trick is the chosen window. "Since May" is a decision someone made, and the question is whether it was made before May or after it.
Ask: how many selections is this figure based on, over what period, and was that period picked in advance?
3. What happened to the records you can't see?
Every visible record is a survivor. The ones that went badly tend not to be on display — not always because anyone deleted anything, but because services that go badly quietly stop being promoted.
Inside a single record, the same effect shows up as an archive that starts suspiciously recently, or a reset explained as a new model or a fresh start. A record that begins after the bad period is a record of the period after the bad period.
Ask: is every selection still visible, losers included? Has the archive ever been reset? And is this the only thing this source publishes, or the one that happened to work?
4. What does the headline rate hide?
A strike rate on its own means nothing, because it has no comparison attached. The number you need is what random selection would have produced over the same races — and that is arithmetic you can do yourself.
Worked example. A service says its shortlist of 3 runners contains the winner 45% of the time.
In a field of 10, three runners picked at random contain the winner 3 ÷ 10 = 30% of the time.
So the claim is 45 ÷ 30 = 1.5× random — not 45 times better than guessing, which is how a bare percentage tends to read.
Do the same division on any rate you're shown. Anything close to 1.0× is doing what chance does. And note the direction the field size pushes: a shortlist looks far better in small fields than large ones, so a headline rate averaged across both flatters itself.
Ask: what would random have done here, and what is the ratio?
5. Does its confidence match what happened?
Anything that publishes a score, a probability or a confidence band is making a testable promise. If the band it labels most confident predicts one thing and delivers materially less, it is overconfident exactly where you are being asked to trust it most.
This is separate from whether it makes money, and it is worth keeping separate. A model can be well calibrated and still not beat the market, because the market is already priced. Calibration is not profit.
Ask: can I see the curve — predicted against actual, across every band, not just the top one?
The five questions, together
- Price: at what price is it settled, and could I have got it?
- Sample: how many selections, over what period, chosen when?
- Survivorship: are the losers still there, and has the archive ever restarted?
- Baseline: what would random have done, and what's the ratio?
- Calibration: does its stated confidence match the outcome?
Our own answers
It would be a poor page that asked those questions and dodged them. We publish selections before the off and settle them in the open, losers included, with a hash chain you can recompute in your browser so you can check that no earlier day was edited after the fact.
And the finding we'd rather not have: we back-tested every scoring variant against tens of thousands of real results, and the honest read is that the model ranks a race sensibly but tracks the market rather than beating it. So there is no profit claim here to check — we don't make one, because we don't have one.
The numbers behind all of that, including the ones that don't flatter us, are on Performance — rendered from the published data rather than typed in by hand, so they can't quietly go stale.
This page is about how to evaluate a claim. It is not betting advice, it is not a recommendation for or against any service, and it names none. 18+ · BeGambleAware