What are we trying to estimate?

A statistical calculation needs a defined target. For identical one-unit independent rounds, X can represent the total payout per round and its mean μ the theoretical return per unit. The sample mean estimates μ. Multiplying by 100 expresses that mean as an RTP percentage.

Real data may include changing stakes, linked free spins, several modes or different game versions. A simple average of payout multiples may no longer equal the monetary payout ratio. First define a completed round, a stake basis and a homogeneous dataset. Weighted or dependent observations require methods appropriate to those features.

Before choosing a formula, ask whether the outcome ledger is complete. Selecting only winning streams or sessions shared by volunteers can bias the result even if the number of observations is large.

For mixed monetary stakes, first use a turnover-weighted return calculation.

How payout variation affects uncertainty

For independent identically distributed observations with finite known standard deviation σ, the standard error of the sample mean is σ/√n. It measures variation between possible sample means under the model. It is not the standard deviation of individual payouts and does not limit what a particular session can return.

With an estimated standard deviation and suitable assumptions, a t-based interval uses the sample standard deviation and an appropriate t critical value. Normal approximations need justification, especially for strongly skewed payouts and rare prizes. The NIST reference explains confidence limits for a mean; applying the method to a particular slot requires checking those assumptions.

Reference: NIST: confidence limits for the mean. The calculations in this section are illustrative unless a named product is explicitly identified.

Worked interval under an explicit hypothetical model

A 95% confidence procedure is designed to cover the fixed true parameter in 95% of repeated samples under its assumptions. It does not mean that a particular future session has a 95% chance of paying within the interval.

How many observations would a chosen precision need?

Under the same normal-approximation assumptions, a target half-width e for the mean suggests n ≈ (1.96σ/e)². Units matter: one percentage point of RTP is 0.01 payout units per one-unit wager, not 1 payout unit.

When a reassuring interval can still be wrong

  • The dataset omits losses, cancelled rounds or delayed feature payouts.
  • The estimated spread misses a rare prize that materially contributes to the mean.
  • Several modes, changing stakes or versions are pooled without a suitable model.
  • Rounds are dependent because of stateful features.
  • The analyst repeatedly checks results and stops at a convenient point without accounting for the stopping method.

A conventional interval formula does not repair bad sampling or a mismatched definition. A narrow interval around a biased dataset can be confidently misleading.

What a responsible monitoring report should include

Identify the title, version, mode, stake treatment, full period, data coverage, round count, payouts, wagers and statistical assumptions. Where the game model supplies exact theoretical probabilities, model-based testing may be more informative than treating a small player history as an unknown-population experiment.

The UKGC’s monitoring guidance discusses theoretical return, actual return and volatility in an operational setting. That is a reason to request the relevant evidence, not to treat every losing account as an anomaly. Use the evidence checklist to document a specific concern without wagering more to enlarge a sample.

Reference: UKGC: monitoring theoretical and actual RTP. The calculations in this section are illustrative unless a named product is explicitly identified.

Place the statistical claim within the wider evidence matrix for RTP claims.

Common questions

Is 1,000 spins enough to check RTP?

There is no universal answer. The payout distribution, dependence, completeness of the data and desired precision determine how informative the sample is.

See the explanation: What are we trying to estimate? →

Explore the explanation: How payout variation changes the uncertainty →

What does a 95% RTP confidence interval mean?

It describes an estimation procedure’s long-run coverage under stated assumptions. It is not a guarantee about a future session’s payout.

See the explanation: Worked interval under an explicit hypothetical model →

How does halving the error margin change sample size?

In the stated standard-error model, it requires four times as many independent observations, with the same payout variation and confidence multiplier.

See the explanation: How many observations would a chosen precision need? →

Can a large dataset still give misleading RTP?

Yes. Selection bias, missing rounds, mixed versions or an unsuitable statistical model can invalidate an estimate regardless of the record count.

See the explanation: When a reassuring interval can still be wrong →

Sources & further reading

Primary sources checked on 17 November 2026. The examples above use the stated rules and assumptions. Product pages describe published versions; they do not verify every operator’s configuration. Regulatory guidance applies within its stated scope.