The Odds Lab

What Odds Movement Before Kick-off Actually Tells Us

Prices move. Everyone who has watched an exchange market knows that. The interesting question is whether that movement carries information beyond what the final price already contains.

The efficient-market answer is no. If a selection drifts from 3.00 to 3.40, the 3.40 should already embody every reason the price moved. The history is baked in. Under that view, odds movement before kick-off is a curiosity, not a signal.

We ran the numbers on our own recorded prices for August 2026. The answer we got was not the textbook one — and the honest reading of it is a weak, provisional hint rather than anything resembling a finding you could build on.

What we recorded

The method is simple. For every selection we take the first traded price we recorded and the last traded price before kick-off. We measure the percentage change between them. Then we take the final price, strip out the overround so the book sums to 100%, and treat the resulting number as the market’s implied probability. Compare that implied probability against what actually happened, aggregated across a group of selections, and you get an expected win count to test the observed win count against.

That is the whole test. It does not ask whether a price “should” have moved. It asks whether the group of selections that shortened won more or fewer times than their own closing price predicted.

The liquidity filter

Thin markets produce noise. A €40 bet can shift a price 5% in an illiquid tennis market at two in the morning, and that movement means nothing at all.

So we restricted the analysis to markets with more than €3,000 matched. That left 247 markets out of nearly five thousand matches — a small slice, which tells you something about how concentrated exchange liquidity really is in the pre-match window.

What the data showed

Three groups, split by how far the price moved between first and last recorded trade.

Selections that shortened by more than 3%

Expected wins from their own closing prices: 99. Actual wins: 85.

z = −2.03, p = 0.04.

Backed selections won less often than their final price implied. The money that came in pushed the price past the point the results justified.

Selections that drifted by more than 3%

Expected wins: 61. Actual wins: 75.

z = +2.20, p = 0.03.

The mirror image. Selections the market abandoned won more often than the closing price suggested.

Control group: prices stable within 3%

z = +0.10.

Nothing. No deviation at all. Which is exactly what you want to see in a control — it suggests the probability normalisation isn’t systematically biased and the method isn’t manufacturing a result out of thin air.

The reading: overreaction, not information

The pattern is internally consistent. Money arrives, the price moves, and it moves too far. Steamers get overbought. Drifters get oversold. The control group, where no money arrived in size, prices correctly.

This is a well-known effect in financial markets and it has an obvious behavioural mechanism. A price move is itself a signal to other participants. Someone sees a selection shortening, infers that informed money is behind it, and follows. Layers pull their offers. The move feeds on itself past the point where the underlying information warranted.

If our numbers are describing something real, the closing price on a heavily-moved selection is not the best available estimate of the true probability — the mid-point between opening and closing might be closer.

That is the interesting version of the story. Here is why we are not telling it with any confidence.

Why this is an indication, not a proof

One month is not a season

Thirty-one days. That is the single biggest problem and no amount of statistical framing fixes it.

Worse, the month is August. Early-season football is the least stable environment in the calendar: squads are half-assembled, transfer windows are open, form lines from the previous campaign are unreliable, and market-makers are pricing teams they have not seen play competitively in three months. If there is a period where pre-match money is likely to be noisier — and where the market’s initial prices are likely to be further from the truth — August is it.

An effect that appears in August may be an August effect, not a market effect. We would need November, February and April data before anyone could claim otherwise.

We tested more than one thing

The p-values look presentable in isolation. They are less impressive in context.

We tested three groups. We also chose a 3% threshold, which is one choice among many — 2%, 4% and 5% were all available. Every additional cut you make at the data raises the chance that one of them clears p = 0.05 by luck alone.

A crude Bonferroni correction for three groups turns p = 0.04 into roughly p = 0.12. That is not significant by any conventional standard. The threshold choice makes it worse still.

The two significant results pointing in opposite and symmetrical directions is mildly reassuring — a pure fluke would not usually be so tidy. But “mildly reassuring” is as far as it goes.

Sample sizes are small

Expected win counts of 99 and 61 imply groups running to a few hundred selections each, not thousands. A difference of 14 wins is the kind of gap that a single unusual week of results can create or erase.

The numbers exclude commission

Everything above is calculated on gross outcomes. Exchange commission is not in it.

That matters enormously if anyone tries to translate the observation into a strategy. On the drifting group, 75 wins against 61 expected looks like a substantial gross edge on that subset. Strip out 2–5% commission on net winnings, then account for the fact that you cannot always get matched at the last traded price — the price you see and the price you actually get are different things, particularly in the final minutes — and a large fraction of the apparent gap disappears.

An edge that only exists gross of costs is not an edge.

The capture cadence is coarse

Four snapshots a day means the “first” price for a given selection could be anything from three days to twelve hours before kick-off, depending on when the market opened relative to our capture times. And the final recorded trade is the last one before kick-off in our data, which is not necessarily the true closing price — late team news and the final flurry of money can move things after our last observation.

This is a genuine measurement problem. A finer capture interval would produce a cleaner test, and it might produce a different answer.

What a real test would look like

If someone wanted to establish whether this holds, the design writes itself:

  1. Six to twelve months of data, spanning at least one full season, including mid-season and run-in periods.
  2. A pre-registered threshold. Pick 3% before looking, and do not go fishing for the cut that produces the prettiest z-score.
  3. Separate football and tennis. They are different markets with different liquidity profiles and different information flows. Pooling them is convenient, not correct.
  4. Higher-frequency capture, ideally hourly in the final 24 hours, so movement can be attributed to a window rather than a whole day.
  5. Net-of-commission accounting from the outset.
  6. Out-of-sample validation — build on one period, test on another that was never examined during construction.

Until that exists, what we have is a single month of August data that mildly contradicts the efficient-market assumption on 247 markets.

The honest summary

Odds movement before kick-off may carry residual information that the closing price does not fully absorb. Our data leans that way. The control group behaving correctly gives the result a bit more credibility than a bare p-value would.

But a one-month sample from the least representative month of the season, with multiple groups tested and commission excluded, is a reason to keep collecting data. It is not a reason to believe anything yet.

Most published claims about steamers and drifters rest on considerably less than this, and are stated with considerably more confidence. That asymmetry is worth noticing.

This article is for informational purposes only. It does not constitute betting advice or a recommendation to gamble. Gambling can be addictive — 18+.