Last updated · By the Clerk2Officer editorial team · Report a correction
1. Describe location and spread together
A branch reports an average queue time of ten minutes. Another reports the same average, but customers at the second branch experience much more unpredictable waits. A measure of central tendency describes a representative location. A measure of dispersion describes how observations vary around or across that location. Shape describes asymmetry and tail behaviour. A useful statistical summary combines these features instead of expecting one average to answer every operational question.
Start by identifying the variable and unit. A mean balance is measured in currency, a mean duration in time, and a mean count in count units even if its numerical value is fractional. A summary should refer to a defined population and period. The choice of average depends on the question. Arithmetic means suit additive totals, geometric means suit multiplicative growth, and harmonic means arise in some rate problems. The median is resistant to extreme magnitudes and can represent a typical customer experience. The mode identifies the most frequent category or value. No measure is automatically best for every dataset. The following calculations use invented banking examples so the arithmetic can be checked independently of changing market or regulatory facts.
2. Arithmetic mean and its properties
The arithmetic mean is the sum of observations divided by their count. For service times four, six, eight, ten and twelve minutes, the sum is forty and the mean is eight. Deviations from eight are minus four, minus two, zero, two and four. Their sum is zero. This balancing property is a useful arithmetic check. If every observation increases by three minutes, the mean increases by three. If each value is multiplied by two, the mean doubles.
For a frequency distribution, multiply each distinct value by its frequency, add the products and divide by total frequency. Values two, four and six with frequencies three, two and five have mean (6+8+30)/10 = 4.4. When only grouped intervals are available, class midpoints may substitute for individual observations, giving an approximation. The quality depends on how observations lie within each class. An open-ended class has no uniquely defined midpoint. Do not silently assign one. The mean uses every magnitude and is sensitive to extremes. Replacing twelve with one hundred in the earlier service-time example raises the mean sharply even though four of five customers have unchanged experiences.
3. Weighted and combined means
A weighted mean is the sum of weight times value divided by the sum of weights. The weights must match the question. A combined mean of groups uses group sizes as weights because each underlying observation contributes equally. If twenty applications average six days and thirty average ten days, the combined total is 20 × 6 + 30 × 10 = 420 application-days. Dividing by fifty applications gives 8.4 days. The simple mean of six and ten, namely eight, incorrectly gives the groups equal influence.
Weighted averages also arise in portfolio yields. If balances are stable over the stated period, a balance-weighted rate can describe the portfolio's average rate. If the question concerns average customer experience, account balances may be inappropriate weights. A branch with a large loan should not automatically dominate the average time experienced by customers. Always explain what one unit of weight represents. When combining percentages, reconstruct numerators and denominators where possible. Ten defaults among one hundred loans and twenty among four hundred yield thirty defaults among five hundred, or six percent. Averaging ten percent and five percent gives 7.5 percent and answers a different question about equally weighted groups.
4. Geometric mean for growth
The geometric mean of positive values is their product raised to the reciprocal of the number of values. For growth calculations, use growth factors rather than raw percentage changes. An increase of twenty percent has factor 1.20. A decrease of twenty percent has factor 0.80. Their product is 0.96, so an initial amount of one hundred becomes ninety-six after two periods. The constant equivalent factor is √0.96, approximately 0.979796. The equivalent growth rate is therefore about minus 2.02 percent per period.
The arithmetic average of plus twenty and minus twenty is zero, but zero growth would leave the original amount unchanged. Compounding explains the difference. For three periods with factors 1.10, 1.05 and 1.08, multiply the factors and take the cube root, then subtract one. All factors must be positive for the usual logarithmic calculation. A total loss produces a zero factor and requires special treatment, while negative values do not support the ordinary positive-data geometric mean interpretation. Distinguish a historical compound growth rate from a forecast. A constant equivalent past rate summarises the observed path but does not establish that future growth will continue at that rate.
5. Harmonic mean for rates
The harmonic mean of positive values is n divided by the sum of their reciprocals. It is appropriate when equal quantities are processed at different rates and the desired combined rate is total quantity divided by total time. Suppose a team processes ten files at ten files per hour and another ten files at twenty files per hour. The first batch takes one hour and the second half an hour. Twenty files in 1.5 hours gives 13.333 files per hour. The harmonic mean of ten and twenty gives the same result.
If equal time rather than equal quantity is spent at each rate, the arithmetic mean is appropriate. One hour at ten files per hour and one hour at twenty produces thirty files in two hours, or fifteen per hour. The correct average follows from the underlying totals. For positive values, arithmetic mean is at least geometric mean, which is at least harmonic mean, with equality when all values are equal. This inequality is a useful check but does not choose the appropriate measure for the business question. In a rate problem, write the numerator and denominator units first. Many mistakes disappear once total files and total hours are explicitly reconstructed.
6. Median and quartiles
The median is the middle ordered observation when the count is odd. With an even count, the conventional numerical median is the mean of the two central observations. For two, four, five, nine and thirty, the median is five. The arithmetic mean is ten because the extreme value thirty pulls it upward. The median uses order rather than every magnitude, so replacing thirty with three hundred leaves it unchanged. This resistance is valuable when describing a typical outcome in a skewed distribution.
Quartiles divide ordered data into quarters. Different sample quartile conventions can produce different numerical answers for small datasets, so state the convention used. Under the median-of-halves convention excluding the overall median for odd n, the data two, four, six, eight, ten, twelve, fourteen have median eight, lower quartile four and upper quartile twelve. For grouped data, a common interpolation estimates the median as L + [(N/2−c)/f]h, where L is the median-class lower boundary, c the cumulative count before it, f its frequency and h its width. This assumes an approximately uniform spread within that class. The corresponding quartile formulas replace N/2 with N/4 or 3N/4.
7. Mode and grouped location
The mode is the most frequent value or category. It can be meaningful for nominal data, such as the most frequently used payment channel. A dataset may have one mode, several modes or no uniquely useful mode. For values two, two, three, four, four and five, both two and four are modes. The arithmetic mean is not required to coincide with either. Grouping affects the apparent modal class, so the grouping rule should accompany a grouped interpretation.
For equal-width grouped classes with a clear modal class, a common interpolated mode is L + [(f₁−f₀)/(2f₁−f₀−f₂)]h. Here f₁ is the modal-class frequency and f₀ and f₂ are adjacent frequencies. With frequencies five, twelve and eight for classes zero to under ten, ten to under twenty and twenty to under thirty, the estimate is 10 + (7/11) × 10, approximately 16.36. This is an estimate based on the histogram shape, not a known observed value. Unequal class widths require attention to frequency density. The empirical relation mode approximately equals three medians minus two means is only a rough relation for some moderately skewed distributions, not a universal identity.
8. Range and quartile dispersion
Range equals maximum minus minimum. For balances of two, four, six, eight and ten units, the range is eight. It is simple and sensitive to extremes because it uses only the endpoints. For positive endpoints, the coefficient of range is (maximum−minimum)/(maximum+minimum). Here it is 8/12, or two thirds. The denominator can be zero or the interpretation inappropriate for some signed or non-ratio measurements. Always check the measurement scale before applying relative coefficients.
The interquartile range is Q3−Q1 and describes the spread of the middle half of the distribution. Quartile deviation, also called semi-interquartile range, is half that difference. With Q1 = 10 and Q3 = 30, the interquartile range is twenty and quartile deviation ten. The coefficient of quartile deviation is (Q3−Q1)/(Q3+Q1), giving one half here. It is less affected by extreme tails than the range, but deliberately ignores some information. Two datasets can have identical quartiles and very different worst delays. A service dashboard may therefore show both the middle-half spread and an upper percentile. Robustness is useful, but it does not eliminate the need to inspect operationally important extremes.
9. Mean deviation and variance
Mean absolute deviation around a specified centre is the average absolute distance from that centre. For two, four and six around their mean four, absolute deviations are two, zero and two, giving 4/3. Specify whether the centre is the mean or median. Absolute deviations avoid cancellation. Squared deviations provide another approach and lead to variance. For the same three values, squared deviations from four are four, zero and four.
If these three values constitute the entire population, variance is eight divided by three. If they are a sample used to estimate a population variance, the usual unbiased sample variance divides by n−1 and is eight divided by two, or four. Estimating the sample mean uses one degree of freedom, explaining the adjustment. Standard deviation is the square root of variance. Population SD is √(8/3), approximately 1.633, while sample SD is two. Variance has squared units and standard deviation restores the original units. State which convention is used. A calculator that switches denominators without explanation can create apparently contradictory answers even when the underlying observations are identical.
10. Calculating and interpreting standard deviation
A direct variance calculation finds the mean, computes deviations, squares them, adds the squares and divides by the chosen denominator. The identity sum of squared deviations equals sum of squared values minus n times the squared mean provides an alternative. For values four, six, eight, ten and twelve, the mean is eight and squared deviations total forty. Population variance is eight and population SD approximately 2.828. Sample variance is ten and sample SD approximately 3.162. Keep extra precision during calculation and round the final result.
Adding a constant to every observation changes the mean but leaves variance and standard deviation unchanged. Multiplying every observation by a constant multiplies SD by its absolute value and variance by its square. Converting minutes to seconds therefore multiplies SD by sixty and variance by 3,600. Standard deviation summarises overall squared dispersion and gives substantial influence to extreme deviations. It is not a maximum error and does not guarantee that observations lie within a fixed number of standard deviations unless a distributional statement is supplied. In a normal model, familiar percentage rules apply approximately. Without such a model, do not automatically claim that nearly all banking outcomes fall within three standard deviations.
11. Coefficient of variation
The coefficient of variation is standard deviation divided by mean, often multiplied by one hundred to express a percentage. It compares relative variation for positive ratio-scale measurements with a meaningful nonzero mean. Branch A has mean waiting time ten minutes and SD two, giving CV twenty percent. Branch B has mean twenty and SD three, giving CV fifteen percent. Branch B has greater absolute variation but smaller variation relative to its own mean. Neither statement alone establishes which branch provides better service.
The CV is unstable when the mean is near zero and generally unsuitable for quantities whose zero point is arbitrary. It also requires caution for signed returns, where the mean may be small or negative. A low CV does not compensate for an unacceptably high average delay. Compare location, absolute spread and relative spread with the decision objective. If a manager wants predictable staffing, variation in workload may matter. If customers face a fixed service promise, the fraction exceeding that promise may be more relevant. Relative measures help compare scales, but they should not replace the original units that determine costs and customer experience. Always report the mean used in the denominator.
12. Skewness
Skewness describes asymmetry. A right-skewed distribution has a longer or more influential right tail, as often occurs with transaction values or delays. A left-skewed distribution has the corresponding left tail. The mean is often pulled toward the long tail, although familiar mean-median-mode orderings are not universal laws. A histogram or ordered data inspection should accompany a numerical coefficient when possible. A single outlier can strongly affect moment-based skewness.
Pearson's first coefficient is (mean−mode)/SD when a useful mode exists. Pearson's second is 3(mean−median)/SD. For mean twelve, median ten and SD six, the second coefficient is one. Bowley's coefficient uses quartiles: (Q3+Q1−2median)/(Q3−Q1). For Q1 = 4, median = 6 and Q3 = 12, it is one half. Moment skewness standardises the third central moment by the cube of SD. These are different definitions and need not agree numerically. Use the definition requested in the problem. A zero skewness coefficient does not prove normality, and some asymmetric distributions can have zero third moment. Shape assessment should consider more than one numerical feature.
13. Kurtosis and tail interpretation
Moment kurtosis is the fourth central moment divided by the square of the variance. It is dimensionless and gives strong weight to large deviations because they are raised to the fourth power. A normal distribution has kurtosis three. Excess kurtosis subtracts three, so the normal reference has excess zero. Some software reports kurtosis and other software reports excess kurtosis. A value of zero is therefore meaningless without the convention. Sample bias corrections can also cause numerical differences between implementations.
Relative to the normal reference, positive excess kurtosis indicates greater fourth-moment tail weight, while negative excess indicates less. It should not be interpreted only as the height of the central peak. Rare extreme losses can matter greatly even when the centre looks ordinary. A portfolio with occasional very large deviations can have a high kurtosis and make normal-tail approximations unreliable. Conversely, symmetry does not determine kurtosis. A symmetric uniform distribution and a normal distribution have different tail behaviour. When using classroom labels such as leptokurtic, mesokurtic and platykurtic, connect them to the stated kurtosis convention and the normal reference. Avoid treating these labels as complete descriptions of operational risk.
14. An integrated service case
Five observed waiting times are two, four, four, five and twenty minutes. The mean is seven and the median four. The mode is four. The range is eighteen. Squared deviations from seven are twenty-five, nine, nine, four and 169, totalling 216. Population variance is 43.2 and population SD approximately 6.573. If the five observations are treated as a sample, sample variance is fifty-four and sample SD approximately 7.348. The large final wait pulls the mean upward and creates a long right tail.
The operational conclusion depends on the objective. A typical customer waited around four minutes, but one customer experienced twenty. Removing that observation simply because it is inconvenient would hide a possible service problem. Investigate whether it is a valid delay, a different service category or a recording error. A useful report identifies the sample size, median, mean, spread and the reason for the long wait if known. To check your calculations, verify that deviations from the mean sum to zero, that variance is nonnegative, and that the mean lies between the minimum and maximum for positive weights. Use the calculator to compare mean and median after changing only the largest observation.
Guided exercise: grouped location and its uncertainty
Forty service times fall into equal-width classes zero to under ten, ten to under twenty and twenty to under thirty minutes, with frequencies ten, twenty and ten. Midpoints are five, fifteen and twenty-five. The grouped mean estimate is (10×5 + 20×15 + 10×25)/40 = 15 minutes. The median class is the second because the cumulative count crosses twenty there. Interpolation gives 10 + [(20−10)/20]×10 = 15 minutes. Equal estimates do not prove that the original observations were symmetric, because grouping has discarded their exact positions.
The midpoint-based population variance estimate is [10×100 + 20×0 + 10×100]/40 = 50, giving SD approximately 7.071 minutes. This calculation treats all observations within a class as if they occurred at its midpoint. It can differ from the variance calculated from raw times. If the underlying records are available, prefer them for precise computation. If only the grouped table is supplied, label the result as a grouped estimate and state the denominator convention. A useful answer includes the method, units and approximation, rather than presenting extra decimal places as evidence of greater accuracy. Grouping improves readability at the cost of detail, and a careful analyst keeps that tradeoff visible.
Open the interactive chapter laboratory and offline practice
Chapter practice
This practice is separate from timed mock tests. Answers and explanations appear after you choose an option.
Your reading progress
Sign in to save your progress across devices.