A club keeps beating strong opponents. Another has an expensive squad. Which offers the better clue to its next result? Gergely Csurilla and László Csató examined that question in a football research preprint submitted to arXiv on 18 September 2026. Their analysis covers Champions League and Europa League matches from 2020/21 to 2024/25.
The question behind any quoted probability is how the number was obtained. That applies when reading a match preview or a game description in an online casino. The event and the assumptions behind the calculation need to be stated before a percentage can be assessed.
The Five-Season Test
The researchers compared performance-based Elo ratings with squad market valuations. Their paper, “Match forecasts in UEFA club competitions: Elo ratings versus Transfermarkt valuations”, reports almost identical forecasting accuracy when the indicators are used separately on test data.
Combining them helped a little. The best way to combine them varied with the outcome being predicted, so the paper offers no single formula that works best for every forecasting task.
That finding is interesting because the inputs come from different places. An Elo rating tracks competitive performance through a calculation. A squad valuation puts an estimated market value on the players available to a club. The researchers tested how much each contributed to forecasting completed European fixtures.
A second source of information can be useful even when it adds only a small improvement. The size of that improvement matters, though. Adding another column to a forecast is easy; showing that it improves results requires a comparison on matches held back for testing.
The work is a preprint. Its findings should remain attributed to Csurilla and Csató, with the competitions and five-season period attached.
A Squad Valuation Needs a Date
Imagine an analyst preparing a season forecast in August. Nine months later, the squad’s valuation has changed after a full season of performances. Both figures concern the same club, but only the August figure was available for the original forecast.
Using the later figure in a reconstruction of that forecast would introduce information from the future. The result might look convincing while answering a question the analyst could never have answered at the time.
This is an easy detail to miss in a spreadsheet. A club name can stay unchanged across several rows, while the information beside it comes from different dates.
A research report should explain which version of each input it used. It should also explain how the test matches were kept apart from the material used to develop the method. A chronological test, with earlier information used for later fixtures, helps preserve the order in which events happened.
What a 60% Forecast Allows
Suppose a model gives the home side a 60% chance of winning. A draw remains possible, as does an away win. The forecast has left 40% for those other outcomes.
There are two ways to review what follows. One is to count whether the model’s preferred outcome happened. Another is to examine many fixtures given similar probabilities and see how often home wins occurred.
Those estimates become more convincing when home wins appear at roughly the stated frequency across a large collection of comparable forecasts. One Saturday’s results offer much less evidence.
The underlying count needs to accompany any summary. “60% correct” leaves several questions open: how many matches were assessed, which competitions were included, and what qualified as a correct prediction?
There is also a difference between correctly naming a match winner and predicting an exact score. A report comparing methods should use the same target for both. Otherwise, the headline percentages can conceal two very different tests.
The Shot Behind an xG Figure
Expected goals, or xG, concerns a shot’s estimated chance of becoming a goal. A model may consider the shooting position, the angle and how the opportunity developed, using earlier attempts to estimate how often such chances are converted.
Take an illustrative shot valued at 0.25 xG. That figure assigns the attempt an estimated 25% scoring chance. It says nothing by itself about the team’s probability of winning the match.
This is why an xG figure needs its label when copied into a preview. A value for one shot, a team’s total from a completed match and a pre-match win probability refer to different events.
Providers can also use different inputs. If two reports give different xG values for the same attempt, the model and its features are worth checking before treating either figure as a transcription error.
Game Modes and the Probability Calculation
A probability quoted for Thimbles needs the selected mode beside it. The review should establish which positions count as successful and what assumptions are being made about their likelihood. The game’s instructions are the place to check those conditions.
For a purely hypothetical example, imagine three equally likely positions with one ball hidden among them. Choosing one position gives a one-in-three chance of finding it. With two balls occupying different positions, that becomes two in three. These numbers illustrate the calculation; they are not verified probabilities for the linked game.
A payout figure adds another piece of information. If a successful ten-unit stake returns twenty-five units including the stake, the multiplier is 2.50. That number does not establish the probability of a successful selection.
A game review needs enough information to explain both figures. Where rounds are independently generated, previous results cannot identify the next successful position. Spending and time limits should be set separately from any probability calculation.
The Details Behind a Published Forecast
A compact research note can make a forecast much easier to revisit. Useful fields include:
- The fixture and precise outcome being predicted
- The date the forecast was recorded
- The dates of the ratings or valuations used
- The matches included in the test
- The measure used to judge the results
- The study’s stated limitations
For the September paper, that note would name Csurilla and Csató, identify the two indicators and retain the five-season European sample. It would record their finding of similar accuracy and the modest improvement from combining the inputs.
The next question is how a published method performs when applied to further matches with information available beforehand. Keeping those forecasts allows a later review to compare the original estimates with the results, without updating the earlier predictions after the scores are known.
