Checking the actual slot return: working methods

0) What can we really check

Theoretical RTP - declared by the provider (there are several versions of the same title).
The actual return of your session is empirical RTP on your sample.
RNG's integrity is not directly verified by the player; we check the RTP version and statistics of our results.

1) We confirm the RTP version with the operator (before tests)

1. Slot screen info: displayed RTP version (number) and provider.
2. Comparison: the same title in another licensed casino - do the RTP numbers and the maximum win (max win) match?
3. Modes: for some games, RTP with Bonus Buy differs from the base game - we fix both values.
4. Promo limits: max bet, excluded games, contribution - so as not to lose winnings "according to the rules."


2) Test plan: how to count the "real return"

Definitions

$ b _ i $ - rate in the back $ i $ (in currency).
$ w _ i $ - payment for spin $ i $ (in currency).
Empirical RTP:

• $$
• \widehat{RTP}=\frac{\sum_i w_i}{\sum_i b_i}imes 100%
• $$

For comparability, also return to the rate: $ r _ i =\frac {w _ i} {b _ i} $.

Separation of modes

Keep separate series: Base/Frispins (payments that fell into the bonus )/Bonus Buy. Mixing breaks the variance estimate.

Fix in advance

Sample size (how many spins/purchases), slot version, bet, error target.

3) Confidence interval and sample size (practical)

We estimate the average $ E [r] $. We take the variance from the data and construct the interval:

• 1. Collect a pilot: ≥ 1,000 spins (or 30 bonus purchases).
• 2. Calculate the sample average of $\bar r $ and Art. deviation of $ s $ at $\{ r _ i\} $.
3. For main batch $ n $ spins 95% -interval:
• $$
• CI \approx \bar r \pm 1{,}96 \cdot \frac{s}{\sqrt{n}}
• $$

4. Required sampling for the desired accuracy $\pmarepsilon $ (in fractions, e.g. 0.01 = 1 pp):
• $$
• n \approx \left(\frac{1{,}96\cdot s}{arepsilon}ight)^2
• $$

Practice guidelines (after the pilot, adjust to your $ s $):
• Low/medium volatility: to meet the ± of 3 percentage points, you need 10-30 thousand spins.
• Medium/high: ± 3 pp require 30-80 thousand spins.
• ± 1 pp for most slots is often a spin 100k-1M.
• For Bonus Buy, count "spin" = one purchase; due to the huge variance of ± 10 percentage points. may require 200-500 purchases or more.

4) How to reduce error and "noise"

Single rate denomination in the series.
Autospin with a fixed pace, without "manual selection" of moments.
Clear blocks (e.g., 500-1,000 spins) with interruptions - to control emotions and pace.
Exclude demo mixes: demos and real money are separate magazines.
Separately count bonuses: payments that came from freespins, flag.

5) Journal: minimum set of columns (copy)



`datetime	slot	rtp_shown	mode(base/fs/buy)	bet	win	r = win/bet	balance_before	balance_after	notes`


One CSV per slot version and casino.
Screen the slot info screen at the beginning of the series (RTP proof).

6) Interpretation of the result

Compare your $\widehat {RTP} $ and its 95% -CI to the declared RTP.
If the entire interval is 2 + percentage points lower than the declared one, this is a reason to suspect a different version of RTP or accounting errors.
If the interval exceeds the declared value, no differences were statistically detected (insufficient data/high variance).
Evaluate different modes (base vs buy) separately and do not draw general conclusions from mixed data.

7) A/B casino comparison (same slot)

Purpose: to understand if the operator has a reduced RTP version.
Order:
• 1. Select one title, same bet value.
• 2. In two licensed casinos, collect 10-20k spins.
3. For each - $\bar r, s, n $; mean difference is tested by normal approximation:
• $$
• Z=\frac{\bar r_A-\bar r_B}{\sqrt{\frac{s_A^2}{n_A}+\frac{s_B^2}{n_B}}}
• $$



$	Z	> 1 {,} $96 → differences are significant (check if different RTP versions/modes are compared).


8) What often breaks the test (and how to fix it)

Mixing modes (base + FS + buy) → enter flags and individual totals.
Change rates within a series → do a sub-series on one bet or ration everything to $ r _ i = win/bet $.
Early stops "when it went/when it didn't go" → fix the size of the series before the start.
Promotional distortions (WR, max bet) → turn off bonuses for the duration of the check or take them into account separately.
Changing the casino/version in the middle of the → series is prohibited; new file = new series.

9) Separately about Bonus Buy

Consider buying an atomic test: $ r _ i =\frac {win _ i} {price _ i} $.
Separate the purchase types (80 ×/100 ×/200 ×).
If there is a gumble in the slot before the start of the bonus with the risk of losing the purchase, select a separate selection for it (this is a different risk mode).
Don't mix buy and base - they have different RTP and variance.

10) Quick checklist "Before you draw conclusions"

The RTP version is displayed and scanned.
Series predefined: $ n $ spins/purchases.
The rate is constant (or everything is normalized to $ r _ i $).
Modes are marked: base/fs/buy.
The pilot for 1,000 spins did - estimated $ s $, calculated $ n $.
Calculated 95% -CI for $\widehat {RTP} $.
Conclusions are drawn by intervals, and not by one number $\widehat {RTP} $.

11) Mini-examples (landmarks)

Example 1 (base, average volatility): pilot 1,000 spins → $\bar r = 0 {,} 94 ,\s = 4 {,} $5 (in bets). We want $\pm3 $ pp (= 0.03).
$ n\approach (1 {,} 96\cdot 4 {,} 5/0 {,} 03) ^ 2\approach 8 {,} 6imes 10 ^ 3 $ → take 10,000 spins.

Example 2 (Bonus Buy 100 ×): pilot 50 purchases → $\bar r = 0 {,} 88 ,\s = 0 {,} $9 (to the purchase price). For $\pm10 $ pp (= 0.10):
• $ n\approach (1 {,} 96\cdot 0 {,} 9/0 {,} 10) ^ 2\approach $311 → target 350-400 purchases.

12) What to do if you suspect the "wrong return"

1. Double check: whether the modes are mixed, whether there was a version change/bet.
2. Repeat the series on the second licensed statement.
3. If the discrepancy is stable and significant, write in support with the data: RTP screen, CSV log, your calculations.
4. If there is no reaction, go to the operator with the confirmed RTP version.

Result

A working check of the "real return" is a confirmation of the RTP version by the operator and statistics of your spins with the correct methodology: a single rate, separate accounting of modes, sufficient sampling, calculation of 95% intervals and comparison with the declared RTP. Any conclusions "by sensation" or 200-500 backs - noise. Make a pilot, plan the volume, count the intervals - and make decisions on the numbers, not on the "hotness" of the slot.