How we price a bet — and how we check whether the price was any good

Winning a bet doesn't tell you the bet was right

A bet either wins or it doesn't. That single result tells you almost nothing about whether taking it was the correct decision.

Take a coin flip at 2.10. That is a good bet — you are being paid more than the odds of the event. Flip it once and you lose about half the time. Flip it ten times and losing overall is completely ordinary. The bet was correct on every one of those flips. The scoreboard just hadn't caught up.

This runs both ways, and the second direction is the dangerous one. A bad bet wins constantly in the short run. If you judge your decisions by last week's profit, you will keep the habits that happened to pay and drop the ones that happened not to — which is close to the worst possible way to learn.

So the question is not "did it win". The question is whether the price you took was better than the thing you were betting on deserved. That is measurable, and it is measurable much sooner than profit is.

What a price actually is

Odds are a probability wearing different clothes.

Divide 1 by the decimal odds and you get the probability the price implies. A price of 2.00 implies 50%. A price of 4.00 implies 25%. A price of 1.25 implies 80%.

Now add up the implied probabilities for every outcome in a market. They should total 100%. They never do.

That extra 2.7% is the margin — the overround. It is the bookmaker's cut, built into the price, and it is entirely normal. Every book has one. It is how the business works.

What matters for our purposes is the consequence: because the margin is baked in, the price in front of you is not the market's honest estimate of the probability. It is that estimate, shaded. Any comparison that ignores the shade is measuring the wrong thing.

What we compare a price against

A price cannot judge itself. To say a price is good, you need something to call it good relative to.

Not all prices carry the same amount of information. A market with a thin margin and high limits has been pressure-tested — anyone who thinks it is wrong can bet into it, in size, until it moves. A market with a fat margin and low limits has not. The first is a much better estimate of what is actually going to happen.

So the reference we care about is the sharper market: the one with tight margins and real limits, where informed money is free to correct the price. Exchange prices work on the same principle for the same reason — they are what people were willing to lay, not what a book was willing to offer.

There is a failure mode here we would rather name than hide. A reference price can be wrong. It can be stale, left up after news the market hasn't absorbed. It can be mismatched, where two sources are quietly describing different events — different competition, different teams, same names. Compare against a broken anchor and you get a number that looks like a large edge and is in fact a measurement error.

Filtering those out is a real and ongoing part of the work, not a solved problem. When a figure looks too good, the first assumption should always be that the comparison is broken rather than that the market is.

Closing line value: the scoreboard we actually use

Here is the idea that does most of the work.

Between the moment a market opens and the moment it closes at kickoff, it absorbs everything: team news, weather, injuries, and every opinion anyone was willing to back with money. The closing price is the market's most informed estimate of that event. It is the last word.

So: compare the price you took against the closing price. If you took 2.60 and it closed at 2.40, you got a better price than the fully-informed market eventually settled on. That is closing line value.

Why it is the better scoreboard is simple. Profit is a slow, noisy signal — outcome variance drowns it for a long time. CLV strips the outcome out entirely. It doesn't care whether the bet won. It asks only whether the price was good at the moment it was taken, judged against the best available answer. That converges far faster, because you get a reading on every single bet instead of waiting for results to average out.

CLV is a proxy, and proxies have limits. It can be measured against the wrong closing price. It can be flattered by only ever betting markets that drift. It says nothing about whether you could actually get the stake on. A number that looks good in a backtest can fail out of sample — that is the normal outcome for a model, not an unusual one.

We hold ourselves to it anyway, because a flawed measure of decision quality is still worth more than a clean measure of luck.

Stake sizing is a separate decision

Which bet to take and how much to put on it are two different questions, and conflating them is one of the more expensive mistakes available.

The same selection can be correct at one stake and reckless at another. If a price is genuinely in your favour, that tells you the direction of the decision. It tells you nothing about size. Size is a function of your bankroll and how much variance you can absorb without being forced to stop — and being forced to stop is the only true failure state, because a strategy you can no longer run has an expected value of zero.

Which brings up the part nobody enjoys. Losing runs are not a malfunction. They are a structural feature of betting positions with small margins. A stretch of losing bets is what this looks like from the inside even when everything is working, which is exactly why judging the method by the last fortnight is a trap.

We are not going to tell you what fraction of your bankroll to stake. That depends on your circumstances, and we are not in a position to know them.

How much evidence is enough

Not as little as most people assume.

The reason is the size of the effect relative to the noise. A meaningful edge in betting is small — low single-digit percentages — while the outcome swings are enormous by comparison. When the signal is that quiet and the noise is that loud, it takes a lot of observations before you can tell them apart. A few dozen bets is not a track record. It is a rounding error with a story attached.

CLV converges faster than profit does, for the reason above — it reads out on every bet rather than waiting for outcomes to average. But faster is not instant, and the honest version of this is that the threshold is a judgement call, not a bright line. There is no sample size at which uncertainty switches off. There is only a point where the number starts carrying more information than noise, and reasonable people put that point in different places.

Ours is roughly 150 settled bets. Below that we will show you the running numbers, but we will not make a claim off them — and we will always print the sample size next to the figure so you can discount it yourself. We would rather state the threshold in advance and be held to it than pick one afterwards, once we have seen which number it produces.

What we will and won't say about our own results

We publish the running figures now, as they stand. They are on the front page: profit, ROI, closing line value, and the number of settled bets they came from. They update as bets settle. Right now the profit line is negative.

We show them because withholding data you already have is not transparency, and because a number is close to meaningless without the sample size printed beside it. So the count always travels with the figures.

What we will not do until the sample passes roughly 150 settled bets is turn those numbers into a claim. No "our members make X". No annualised projection. No screenshot of a good fortnight offered as evidence of anything. The figures are there to be read, not to be sold with — and a positive CLV reading off a sample this size is genuinely not strong enough to carry an argument yet.

When the sample does get there, what changes is the claim, not the disclosure:

• The full sample, dated, with the losing stretches still in it.
• CLV alongside profit — the decision measure and the outcome measure, together.
• The sample size, stated plainly, as it is now.
• The method for measuring it, so the number can be argued with.

Figures elsewhere on this site are labelled for what they are — the five-season back-tested set is marked as a backtest, and a backtest is weaker evidence than live results. If you ever find a number here you cannot trace to a stated method and sample, tell us and we will fix it.

We are not putting a date on this. It happens when the sample supports it.

What this page does and does not tell you

Plainly, so there is no room for a misreading:

• This describes a method, not a track record. Nothing here is evidence that the method works. It is an account of how we make and check decisions, and how we intend to be judged.
• Nothing here is advice. Not betting advice, not financial advice. It is an explanation of how prices and closing lines work.
• Nothing here predicts your results. Even a genuinely positive-expectation approach loses over stretches, and how it goes for you depends on your stakes, your discipline, and which prices you can actually get on.
• We could be wrong. The method rests on the closing line being a good estimate of the truth. That is a widely-held view and we think it is correct. It is still a view.

If something on this page is unclear or you think it is wrong, we would rather hear it: hello@profitpanda.co.uk. Corrections get made on the page.