Business Statistics: The People and Risks Hidden Behind an Average

A single average compresses a dataset, but it does not describe the people, transactions or risks hidden inside it. Reliable interpretation requires three additional questions: what shape does the distribution have, how were observations selected, and how much uncertainty surrounds the estimate?

Reference report · No directional market call

The numerical examples below are teaching inputs from the prepared material. Calculations are shown in their original units so the reader can reproduce them. None of the examples identifies an employer, client or individual.

Start with what a number measures

Data type constrains interpretation.

Measurement types and the comparisons they support

These distinctions are guides rather than substitutes for modelling judgement. Coding cities as 1, 2 and 3 does not make the distance from city 1 to city 3 twice the distance from city 1 to city 2. Categorical predictors can enter a model through properly defined indicators. Similarly, an average of ordinal scores can sometimes be a useful descriptive convention, but it should not be treated as proof of exact psychological distances.

Data cleaning needs the same care. A zero may mean no activity, missing information or a failed capture. A blank may mean not applicable rather than unknown. Before computing a mean or a growth rate, resolve these meanings and preserve the rules used.

Mean, median and the salary example

Consider ten monthly salaries in thousands of rupees:

Monthly salary observations in thousands of rupees — Observation: Salary

The median is ₹42,000 and the mode is ₹42,000. The supplied values sum to ₹645,000, so their arithmetic mean is ₹64,500. The final ₹280,000 observation accounts for approximately 43.4% of total payroll despite representing only one of the ten people. That concentration explains why the mean sits above the middle of the distribution without implying that the large observation is invalid.

The mean answers a valid question: total payroll divided by ten people. The median answers a different question: the middle of the ordered distribution. Neither is universally more honest. For budgeting total payroll the mean is directly relevant; for describing a typical employee the median is more representative. Removing the high observation simply because it is influential would discard economically meaningful information.

The essential rule is to name the quantity and the purpose. “Average salary” is incomplete if the audience does not know whether it refers to a mean, median, full-time equivalent, annualised amount or monthly cash payment.

Spread measures reveal different risks

The prepared comparison uses two businesses with average order value of ₹1,240. One has a standard deviation of ₹85; the other ₹620. Their coefficients of variation are approximately 6.85% and 50% respectively, using standard deviation divided by the positive mean.

The second business has much greater relative dispersion. That may affect stocking, delivery capacity, working capital and customer segmentation. It does not by itself make the business worse: a profitable premium segment can create a wide distribution. The question is whether the operating model understands and can serve the variation.

For observations $x_1,\ldots,x_n$, the sample mean is $\bar{x}$, and the sample standard deviation is:

$$s=\sqrt{\frac{\sum_{i=1}^{n}(x_i-\bar{x})^2}{n-1}}.$$

Here $n$ is the number of observations and $i$ identifies each observation. Squaring prevents positive and negative deviations cancelling; taking the square root restores the original units. The denominator $n-1$ is used for the usual sample-variance estimator, while a complete finite population can use a different denominator for its descriptive variance.

Standard deviation does not require a symmetric distribution merely to exist or be calculated. Symmetry and distributional assumptions become important when converting it into probability statements. An interquartile range, defined as the 75th percentile minus the 25th percentile, describes the middle half and is less sensitive to extreme observations. It should be paired with tail measures when extreme outcomes drive cost.

Shape changes the business question

Distribution shapes reveal different operational questions

The material gives a customer-value example with mean ₹12,000 and median ₹3,400. These figures indicate that high values pull up the mean, but they do not determine the exact revenue share of the top decile. That share must be calculated from the actual observations. A separate example associates 4% of very late deliveries with 65% of complaints and 80% of refund requests. Those percentages are scenario inputs demonstrating concentration, not generally applicable operating ratios.

A useful distribution view combines a histogram, selected percentiles, the mean and median, and a count of observations. For service processes add threshold breaches: the proportion exceeding an agreed limit may be more decision-relevant than the overall average.

Normality is a model, not a default entitlement

For a normal distribution, approximately

Normal-model coverage at the stated distances from the mean

of observations fall within one, two and three standard deviations of the mean.

The percentages are conditional on the distributional assumption. Applying them to skewed claim sizes, fat-tailed returns or mixed customer populations without checking the shape can materially understate risk.

The teaching example uses a manufactured dimension with mean 10 centimetres and standard deviation 0.2 centimetres. A normal model places roughly 99.7% between 9.4 and 10.6 centimetres. Whether that range is acceptable depends on engineering tolerances, which are separate from the statistical distribution. A process can be stable and consistently outside the customer’s specification.

Likewise, a control limit is not the same as a contractual service limit. The first describes unusual variation under a process model; the second describes an external requirement. Confusing them can lead a team to accept poor service because it is statistically normal for its own process.

Sampling determines what an estimate can describe

An estimate is only as representative as its selection process. A survey of enthusiastic repeat customers cannot automatically describe former customers who left. A database of surviving firms cannot establish the expected outcome of every firm that began operating. A voluntary survey can over-represent respondents with unusually strong experiences.

The advanced exercises distinguish survivorship bias, self-selection, sample versus population, and inference audits. These are not small technical adjustments. They determine the target of the inference. More observations from a systematically biased sample can produce a more precise answer to the wrong question.

Record eligibility, sampling frame, invitations, responses, exclusions and missing outcomes. Compare respondents with the population on observable characteristics. Weighting can help when the relevant selection variables are measured, but it cannot fully repair selection on unobserved factors. State the remaining limitation rather than hiding it behind a large sample size.

Standard deviation and standard error answer different questions

Standard deviation describes variation among observations. Standard error describes the sampling variation of an estimator. For independent observations from a population with finite standard deviation $\sigma$, the standard error of the sample mean is:

$$SE(\bar{x})=\frac{\sigma}{\sqrt{n}}.$$

Here $\sigma$ is the population standard deviation and $n$ is sample size. When $\sigma$ is unknown, its estimate and the relevant inferential method must be used. Four times as many independent observations approximately halve this standard error. Four times as many highly correlated observations need not provide that improvement.

A confidence interval adds uncertainty around an estimate. Under its assumptions, a 95% confidence-interval procedure covers the fixed population parameter in 95% of repeated samples. It does not mean that the realised interval has a frequentist 95% probability of containing the fixed parameter, nor that 95% of individual customers fall inside it.

Mix shifts can reverse the apparent trend

An aggregate outcome is a weighted combination of subgroup outcomes. If the weights change, the total can improve even when each subgroup deteriorates. The teaching material illustrates this with recovery rates changing from 78% to 81% while the case mix becomes less severe. Without subgroup denominators and rates, the aggregate cannot establish improved treatment performance.

The same issue appears in conversion rates, loan defaults, branch productivity and customer satisfaction. Compare like with like through stable cohorts, relevant stratification or explicit standardisation. Do not choose the segmentation after seeing results solely because it produces a preferred story.

Seasonality creates a related trap. A strong month may reflect a recurring calendar effect, a promotion or a change in exposure rather than a persistent improvement. Examine appropriate comparisons and separate trend, seasonal effects and irregular variation. A decomposition is a model whose assumptions need review; it does not identify causation automatically.

Keep workbook calculations reproducible

The exercise collection includes separate tabs for data types, central tendency, outliers, spread, variance, quartiles, distributions and confidence intervals. Their common value is that a reader can move from raw observations to a result and then ask what the result means. Preserve that sequence when adapting the method to a business dataset.

Record the exact range included in each formula. A mean calculated over forty observations and a standard deviation accidentally calculated over thirty-nine can appear plausible while describing different samples. Check that percentages use the intended denominator, missing values are handled consistently and filters do not silently change which records contribute.

Different spreadsheet percentile conventions can produce slightly different quartiles in small samples. Name the convention when the boundary affects a decision. Similarly, sample and population variance functions answer different descriptive or inferential purposes. The formula should follow the purpose, not whichever function name happens to be familiar.

Finally, distinguish a calculated result from a typed explanatory sentence. If the dataset changes, the formula may update while the text remains stale. Review both together, especially where the text recommends an action or identifies a segment as unusually risky.

Communicate precision in proportion to evidence

Reporting a long decimal is not evidence of accuracy. Input measurement error, sampling error, model uncertainty and scenario uncertainty can dominate the final digits. A forecast should show the variables that matter, a central estimate when justified and a range tied to explicit assumptions.

The practical sequence is straightforward: establish definitions, inspect the distribution, audit selection, quantify uncertainty and then make the comparison relevant to the decision. Each step prevents a different form of false confidence. Together they turn summary statistics into evidence that a reader can interpret and challenge.

Useful sensitivity analysis changes one or more important inputs and shows the effect on the decision. It is different from inventing a confidence interval around a scenario. Where probabilities cannot be estimated credibly, label scenarios as scenarios rather than attaching arbitrary likelihoods.