
xG per shot is total expected goals divided by the number of shots in the same sample. It describes the average modelled value of those attempts. It does not measure attacking quality on its own: a team taking fewer, better chances can have a higher average while generating less total threat.
Read the average beside shot volume, total xG and the circumstances in which the shots occurred. The useful question is whether a side created more attempts, better attempts, or a different mixture of both. Those are related questions, but they need different denominators.
Keep the numerator and denominator together
Suppose a team produces 2.00 xG from 20 shots. Its average is 0.10 xG per shot. Another produces 1.60 from eight, averaging 0.20. These invented figures give the second side a higher average chance value but the first a higher total.
Neither column makes the other redundant. The first team may be creating plenty of modest chances. The second may be reaching fewer but stronger shooting positions. A match review should preserve both observations instead of declaring a winner from whichever number looks more impressive.
Wyscout’s glossary explains the shot-level probability model behind its xG values. That establishes what its estimates concern; it does not establish that every provider uses identical inputs. Our arithmetic examples use invented values and do not reproduce a real match or OptaTips performance.
Use one provider and the same event coverage for both columns. Dividing a first-half xG total by full-match shots creates a meaningless hybrid. Mixing a total including blocked attempts with a denominator excluding them can also inflate the apparent average.
An average is not the probability of every shot
Five shots with values 0.02, 0.03, 0.05, 0.10 and 0.30 sum to 0.50 xG and average 0.10. None except the fourth needs to have an individual probability of 10%. The average describes the group, not a typical location that every attempt shared.
Those five shots might come from very different positions and phases. Reporting only their mean removes that distribution. If your explanation concerns repeated access to close-range chances, inspect the individual events or a suitable chance-quality breakdown.
A single large chance can move a short-sample mean sharply. Add one invented 0.50 attempt to the five-shot group and the new average becomes 1.00 divided by six, approximately 0.167. The original five attempts have not improved. The composition of the sample has changed.
This is why a high average over one match is a starting point for review rather than evidence of a permanent attacking characteristic. Look for the situations that produced the attempts and whether those situations recur.
Penalties require a matched adjustment
If you remove penalty xG, also remove the corresponding penalty attempts from the denominator. A non-penalty numerator divided by all shots is not the same metric as non-penalty xG per non-penalty shot.
Consider an invented total of 2.10 xG from 15 attempts, including one penalty valued at 0.76 under the example’s provider convention. The all-shot average is 2.10 divided by 15, or 0.14. Removing that penalty leaves 1.34 from 14 attempts, approximately 0.096 per attempt.
Dividing 1.34 by 15 instead gives approximately 0.089. That number is arithmetically possible but represents a different denominator. Label it accurately rather than presenting it as the usual non-penalty average.
The non-penalty xG guide explains why this adjustment does not automatically isolate open play. Corners, free kicks and other non-penalty situations can remain. Use a phase breakdown if the question specifically concerns open-play creation.
Several shots can belong to one attacking sequence
A blocked shot followed by a rebound can create two attempt records close together. A table of shot averages does not show whether the attempts came from separate possessions or from one scramble.
Keep that distinction visible when writing about sustained chance creation. Ten recorded attempts are not necessarily ten independent attacks. Nor can you automatically translate a sum of xG into the probability that a team scored at least once.
The calculation for at least one success under an independence assumption is different from adding the probabilities. Two invented 0.20 attempts would give 1−0.80×0.80=0.36 under that assumption, rather than 0.40. A real rebound sequence may not justify treating its attempts as independent trials at all.
That example explains a mathematical distinction; it is not a replacement match forecast. Preserve the model’s intended use and avoid deriving additional probabilities from a summary without a suitable method.
Compare samples before comparing teams
Venue, opposition, score state and dismissals can change the opportunities a side receives. A team chasing a deficit may take speculative attempts; a leading opponent may wait for transitions. Their averages can reflect those conditions as well as their players’ choices.
Separate comparable periods where the data support it. State how many matches and attempts remain after a filter. A narrow split can clarify context while leaving so few shots that its mean becomes unstable.
The home-and-away comparison guide offers another way to investigate unequal samples. Apply such splits because they answer a specific question, not because one produces a flattering number.
When aggregating games, distinguish the mean of match averages from the pooled shot average. A match with one shot otherwise receives the same weight as one with 20. Pooling compatible xG totals and shot counts answers the average-per-attempt question across the combined sample.
Zero attempts are not zero-quality attempts
With no shots, xG per shot is undefined because the denominator is zero. A dash, blank or other documented display convention is preferable to interpreting 0.00 as the quality of attempts that never happened.
The absence of attempts is itself useful information. It might indicate a failure to progress into shooting positions, but the average cannot explain why. Check territory, possession progression and the match circumstances before assigning a cause.
Similarly, missing event coverage is not a zero-shot match. If a provider did not supply the data, mark the observation unavailable. Converting missing entries to zero can distort totals and give an inaccurate impression of a complete record.
Use a review statement that preserves the trade-off
A defensible summary of the opening example says that A created more total modelled shot value, while B created fewer attempts with a higher average value. It then investigates where the attempts came from and which match conditions helped produce them.
For a goalkeeper or finishing question, a different measure may be needed. The post-shot xG explanation distinguishes the quality of the original chance from information about the delivered attempt.
For a future-match claim, retain a separate forecasting method and evaluation. Historical xG per shot describes recorded attempts. It does not guarantee how many chances will occur next time, what their quality will be, or whether a quoted price represents value.
Work through a two-match aggregation
Here is a second hypothetical example. In the first game a team records one attempt worth 0.40 xG. In the second it records 19 attempts worth 1.90 in total. The match averages are 0.40 and 0.10. Their simple average is 0.25.
Across the 20 actual attempts, however, the combined average is 2.30 divided by 20, or 0.115. The simple mean answers what the equally weighted match averages were. The pooled calculation answers what the average recorded shot was worth. It gives the larger shot sample more influence.
Neither weighting should be hidden. For an article about the kinds of chances a team typically takes, state whether your unit of observation is a match or an attempt. If you want each game to carry equal weight, explain why rather than calling the result the pooled average.
Rounding can introduce smaller discrepancies too. Published match totals may already be rounded to two decimals, while the underlying event values retain more precision. Recomputing from displayed totals may differ slightly from the provider’s own aggregate. Do not diagnose a substantive data fault from a tiny rounding difference without checking the raw fields.
Retain an audit row with match identity, period, provider, eligible shot count, total value and exclusions. A penalty adjustment should list both the removed attempts and their values. That small ledger makes the comparison reproducible without requiring a reader to accept an unexplained dashboard number.
If one game lacks complete coverage, distinguish the average for observed shots from an estimate for the whole period. The former can be calculated; the latter requires a method for missing data. Quietly substituting league averages changes the exercise and needs separate disclosure.
For adults aged 18 and over. All worked examples are hypothetical. Provider references explain definitions and do not imply a data-supply relationship with OptaTips.
Comments 0
No comments yet.