A confidence interval is a range calculated from sample data to express uncertainty around an estimate. It is more informative than reporting one number as if the sample revealed the exact population value.
In casino analysis, confidence intervals can be used around estimates such as average theoretical loss, conversion rate, machine uptime, campaign response, observed return, or the proportion of incidents closed on time. They do not turn a short sample into certainty. They show how much precision the available data supports.
The estimate and the interval answer different questions
Suppose a sample of rated trips produces an average theoretical loss of $184. That is the point estimate. A 95% confidence interval of $169 to $199 describes the uncertainty around the estimate under the chosen method and assumptions.
The point estimate answers, “What is our best single estimate from this sample?”
The interval answers, “What range of population values remains reasonably compatible with this sampling process?”
The interval is not a promise that every future trip, table, machine, or player result will fall inside it.
What “95% confidence” actually means
A common but inaccurate explanation says there is a 95% probability that the fixed true value is inside this one calculated interval. In standard frequentist statistics, the population parameter is treated as fixed. The randomness is in the repeated samples and the intervals they produce.
A better interpretation is:
If the same sampling and interval-building procedure were repeated many times under the model assumptions, about 95% of those intervals would contain the true population parameter.
The NIST/SEMATECH Engineering Statistics Handbook gives a concise official explanation of confidence intervals and repeated sampling.
Confidence interval for a mean
When the population standard deviation is unknown, a common interval for a mean uses the Student t distribution:
[ \bar{x}\pm t_{\alpha/2,,n-1}\frac{s}{\sqrt{n}} ]
Where:
- (\bar{x}) is the sample mean;
- (s) is the sample standard deviation;
- (n) is the sample size;
- (t_{\alpha/2,,n-1}) is the critical t value for the selected confidence level and degrees of freedom;
- (s/\sqrt{n}) is the estimated standard error of the mean.
Worked casino example
A casino samples 36 comparable weekday table shifts. Average drop is $48,000 and the sample standard deviation is $9,000. For a 95% interval with 35 degrees of freedom, the critical t value is approximately 2.03.
[ \text{Standard error}=\frac{9{,}000}{\sqrt{36}}=$1{,}500 ]
[ \text{Margin of error}=2.03\times1{,}500=$3{,}045 ]
[ \text{95% interval}=48{,}000\pm3{,}045 ]
So the interval is approximately $44,955 to $51,045.
This does not predict the drop for the next shift. It estimates the underlying mean for the defined population of comparable shifts, assuming the sample and method are appropriate.
Confidence interval for a proportion
Casino reporting often uses proportions: the share of disputes resolved within target, the response rate to a controlled offer, or the percentage of machines available during scheduled hours.
A simple large-sample interval is often introduced as:
[ \hat{p}\pm z_{\alpha/2}\sqrt{\frac{\hat{p}(1-\hat{p})}{n}} ]
Where (\hat{p}) is the observed proportion and (z_{\alpha/2}) is the normal critical value. This “Wald” interval can perform poorly with small samples or proportions near 0 or 1, so analysts often prefer Wilson or other methods.
Suppose 172 of 200 sampled incidents were closed within the target time:
[ \hat{p}=\frac{172}{200}=0.86 ]
Using the simple 95% approximation:
[ 0.86\pm1.96\sqrt{\frac{0.86(0.14)}{200}} ]
The margin is about 0.048, producing an interval of roughly 81.2% to 90.8%. The estimate is 86%, but the sample does not justify pretending the true rate is exactly 86.0%.
Larger samples usually narrow the interval
The standard error for a mean contains (\sqrt{n}) in the denominator. That means precision improves with sample size, but not in a one-for-one relationship.
To cut the standard error approximately in half, the sample size must be multiplied by four.
| Sample size | Relative standard error |
|---|---|
| 25 | 1.00 |
| 100 | 0.50 |
| 400 | 0.25 |
This is why a few additional observations may do little for a very noisy metric. Read sample size and standard deviation together: sample size controls how much data you have, while standard deviation reflects how variable the data are.
A narrow interval can still be wrong
Confidence intervals describe sampling uncertainty under a model. They do not automatically correct:
- biased samples;
- missing or duplicated records;
- inconsistent definitions;
- changes in game mix or operating conditions;
- non-independent observations;
- heavy tails or extreme outliers;
- a faulty mathematical model;
- measurement error.
A very large biased dataset can produce a narrow interval around the wrong answer. Data quality and population definition come before calculation.
For example, an interval for average table hold based only on profitable weekend shifts cannot represent all shifts. An interval for campaign response that excludes customers who opted out cannot be generalized without care. An observed slot return interval does not reveal programmed RTP if the sample is short, selectively chosen, or affected by jackpots and game mix.
Confidence intervals are not prediction intervals
A confidence interval for the mean estimates the uncertainty around the average. A prediction interval addresses where an individual future observation may fall and is usually much wider.
Suppose the estimated mean daily drop is $48,000 with a 95% confidence interval of $44,955 to $51,045. An individual future day might still be far below $44,955 or above $51,045 because day-to-day variation is much larger than uncertainty about the mean.
Managers often confuse these questions and treat a mean interval as an operating limit. Use a control chart, prediction interval, or other method when the decision concerns individual observations rather than the average.
Confidence level and precision involve a trade-off
Using the same data and method:
- a 90% interval is narrower;
- a 95% interval is wider;
- a 99% interval is wider still.
Higher confidence requires a broader range. Choosing 99% does not improve the data; it changes how conservative the interval procedure is.
The confidence level should be selected before looking at the result and should reflect the decision. A routine directional analysis may use 95%. A high-consequence compliance or safety decision may require a different standard, additional evidence, or a method specified by policy or regulation.
Casino uses where intervals help
Game and machine analysis. Observed return, hold, or event rates can be compared with expected ranges, while recognizing variance and jackpot effects.
Marketing tests. A response-rate difference should be reported with uncertainty rather than declared a winner because one percentage is slightly higher.
Operations. Average service time, uptime, staffing demand, and incident closure rates can be estimated with ranges.
Surveillance and compliance. Alert conversion and false-positive rates should include uncertainty, especially when rare-event samples are small.
Simulation. Repeated simulated estimates can produce Monte Carlo error intervals, but the interval only reflects the simulation design and assumptions. Read simulation for that distinction.
Common interpretation errors
- “95% of the data are inside the confidence interval.” Not necessarily; the interval is usually about a parameter, not the spread of observations.
- “A wider interval means the process is worse.” It means the estimate is less precise, which may result from small sample size or high variability.
- “Overlapping 95% intervals prove no difference.” Overlap is not a complete significance test.
- “A confidence interval proves causation.” It does not address confounding or study design by itself.
- “The true value changes when the interval changes.” The interval changes because the sample changes.
- “More decimal places mean more confidence.” Display precision cannot create statistical precision.
A confidence interval is an honesty tool. It makes the uncertainty visible so managers do not mistake a sample estimate for a complete mathematical truth.