Expected Saves vs Goals Prevented Explained

Expected saves estimates how many attempts a reference goalkeeper would save from a defined set of goalkeeper-tested shots. Goals prevented measures how many fewer or more goals the actual goalkeeper conceded than the post-shot model expected.

On one identical shot set, the measures are directly connected: expected saves equal eligible on-target attempts minus summed post-shot goal probability, while goals prevented equal actual saves minus expected saves. The same goals-prevented result can also be calculated as post-shot expected goals faced minus goals conceded.

The distinction matters because expected saves describes the model baseline; goals prevented describes the goalkeeper's observed result relative to that baseline.

Expected Saves vs Goals Prevented: The Short Answer

For a set containing only shots recorded as either a goalkeeper save or a goal:

• Expected saves asks: how many of these attempts would the model expect the reference goalkeeper to stop?
• Actual saves records how many the goalkeeper stopped.
• Goals prevented measures the difference between actual and expected saves.

Expected saves is a transparent derived quantity in this article. It should not be assumed to be a standardised official field across every provider. Some datasets instead publish expected save percentage, goals saved above average or another related measure.

The Shared Shot-Stopping Inputs

The calculation requires four consistent inputs:

GoalIQAI's guide to Expected Goals on Target and post-shot xG explains how post-shot models assess an attempt after measurable execution information becomes available. Unlike pre-shot expected goals, these models can account for where the ball travels within the goalmouth and, depending on the provider, other information such as trajectory or velocity.

Stats Perform's xGOT explanation describes goals prevented as xGOT faced minus goals conceded. Hudl StatsBomb's post-shot methodology similarly uses saved and scored shots to estimate save difficulty, while emphasising that model calibration affects aggregate goalkeeper evaluation.

How to Calculate Expected Saves

If a shot has a modelled goal probability of pᵢ, its expected save probability is:

Expected save probability = 1 − pᵢ

Add those save probabilities across all eligible attempts:

Expected saves = Σ(1 − pᵢ)

Because there are N eligible attempts, this simplifies to:

Expected saves = N − PSxG

A five-shot workload with combined post-shot xG of 2.20 therefore produces 2.80 expected saves:

Expected saves = 5 − 2.20 = 2.80

This is an aggregate model estimate, not a claim that a goalkeeper can make a fraction of a physical save.

How to Calculate Goals Prevented

First calculate actual saves from the same shot population:

Actual saves = N − G

Goals prevented can then be expressed in either of two equivalent ways:

Goals prevented = actual saves − expected saves

Goals prevented = PSxG − goals conceded

The equivalence follows directly from the shared shot set:

(N − G) − (N − PSxG) = PSxG − G

A positive result means the goalkeeper conceded fewer goals than the model expected. A negative result means the goalkeeper conceded more. Neither result independently proves permanent ability or fault.

Worked Expected-Saves and Goals-Prevented Example

Consider five goalkeeper-tested attempts with the following post-shot probabilities and outcomes:

The model expected 2.20 goals and 2.80 saves. The goalkeeper actually made four saves and conceded one goal.

Goals prevented = 4 actual saves − 2.80 expected saves = +1.20

The alternative calculation produces the same answer:

Goals prevented = 2.20 PSxG − 1 goal conceded = +1.20

The actual save percentage was 80%, while the expected save percentage was 56%:

Actual save percentage = 4 ÷ 5 = 80%

Expected save percentage = 2.80 ÷ 5 = 56%

The difference is 24 percentage points, which is also goals prevented per eligible shot:

1.20 ÷ 5 = 0.24

This small example demonstrates the calculation. It is far too small to establish repeatable shot-stopping skill.

Why Expected Saves and Goals Prevented Answer Different Questions

Expected saves describes workload difficulty. A high figure means the model judged that many of the attempts should have been saved. A lower figure means the goalkeeper faced a more difficult collection of shots, assuming the model is well specified.

Goals prevented describes performance against that workload. It compares what happened with the model baseline. A goalkeeper can record many saves but little positive value if most attempts were easy. Another can make fewer saves yet produce a positive goals-prevented result by stopping unusually difficult shots.

Totals also combine efficiency with opportunity. A goalkeeper facing many attempts has more scope to accumulate positive or negative goals prevented than one facing a small workload. Rates can support comparison, but they do not remove model uncertainty or small-sample noise.

Why the Exact Shot Set Matters

The algebra works only when every component uses the same eligible attempts. Do not combine a headline shots-on-target count from one source with post-shot probabilities from another.

Provider definitions can differ. For example, the published Opta event definitions state that shots on target can include an attempt blocked by a last-line defender, while a save specifically involves a goalkeeper preventing the ball from entering the goal. A goalkeeper calculation may therefore require a narrower saved-or-scored population than the displayed team-level shots-on-target total.

Before calculating the measures, check:

• whether every event ended as a goalkeeper save or goal;
• whether penalties are included;
• how own goals and defender blocks are treated;
• whether rebounds are recorded as separate attempts;
• whether all probabilities come from the same provider and model version;
• whether rounding has changed the published aggregate.

Model Sensitivity and Provider Differences

Post-shot probabilities are model estimates, not observed facts. Providers may use different training samples, definitions and features. One model may include shot speed or goalkeeper position while another may not.

If another model assigned the worked example a combined post-shot value of 1.90 rather than 2.20, expected saves would rise to 3.10 and goals prevented would fall to +0.90. The observed four saves and one goal would not change; only the model benchmark would.

That sensitivity is why figures from different providers should not be added or ranked as if they came from one scale. Even within one provider, a model update can affect comparisons with historical values.

Common Calculation and Interpretation Errors

Using incompatible denominators

Expected saves, actual saves, goals conceded and post-shot xG must describe the same attempts. A population mismatch breaks the mathematical relationship.

Treating expected saves as an independent forecast

Expected saves is conditional on the shots already included. It does not independently predict how many on-target attempts a goalkeeper will face in the next match.

Reversing the subtraction

For a higher-is-better goals-prevented measure, use post-shot xG minus goals conceded, or actual saves minus expected saves. Reversing the order changes the sign.

Assuming every positive result is goalkeeper skill

Deflections, sightlines, defensive interference, positioning, model omissions and randomness can affect the outcome. The GoalIQAI guide to sample size in football analytics explains why a few dramatic saves should not carry the same evidential weight as a larger, comparable workload.

Ranking complete goalkeepers with one metric

Expected saves and goals prevented evaluate shot stopping within a defined model. They do not directly measure cross claiming, sweeping, distribution, communication or tactical fit.

How GoalIQAI Interprets the Measures

GoalIQAI would use a staged process:

1. Define the population. Confirm which saved and scored attempts are included.
2. Verify the model. Record the provider, version, features and treatment of penalties or own goals.
3. Reconstruct the calculation. Derive expected saves, actual saves and goals prevented from the same shot set.
4. Compare totals and rates. Separate accumulated contribution from per-shot efficiency.
5. Inspect the distribution. Check whether one unusually difficult attempt dominates the total.
6. Review context and video. Consider sightlines, positioning, deflections and defensive interference.
7. Express uncertainty. Treat the result as evidence about the observed workload, not proof of future performance.

This approach helps separate signal from noise in football data. A persistent result across meaningful, comparable samples may justify further investigation, but forecasting should still account for regression, team context and uncertainty.

Key Takeaways

• Expected saves estimate how many attempts a reference goalkeeper would stop from a defined saved-or-scored shot set.
• Expected saves equal eligible shots faced minus summed post-shot goal probability.
• Goals prevented equal actual saves minus expected saves.
• The equivalent calculation is post-shot expected goals faced minus goals conceded.
• A high raw save count does not establish above-average shot stopping if most attempts were easy.
• The formulas require identical shot populations, provider definitions and model versions.
• Positive goals prevented can reflect ability, variance, context or model limitations and should not be projected forward unchanged.

Related Guides

• Goalkeeper Analytics Explained

Stay Ahead of the Market

Subscribe to GoalIQAI for evidence-based football predictions, betting-market analysis and practical football analytics guides.