1. Start with the decision
A branch manager sees complaints rising from 30 to 42 in a month and proposes adding staff. The increase is twelve complaints, or forty percent of the original count. That arithmetic is correct, but the staffing conclusion is premature. If completed transactions rose from 10,000 to 20,000, the complaint rate fell from three to 2.1 per thousand transactions. If the new complaints concern a software outage, extra counter staff may not address the cause. Statistical thinking connects a number to its denominator, context and decision. It asks what was measured, how it was measured, and what comparisons are justified.
Statistics has two common meanings. In the plural it refers to numerical facts, such as transaction counts and average balances. In the singular it refers to methods for collecting, organising, analysing and interpreting data. Descriptive statistics summarises the observed records. Inferential statistics uses a sample and assumptions to learn about a wider population. A table describing all applications received yesterday is descriptive. Estimating the average processing time for all branches from a sample introduces inference. Neither activity automatically establishes why an observed difference occurred. Causal explanations require additional design and evidence.
2. Units, populations and variables
Define the unit of observation before collecting figures. An account, customer, application and transaction are different units. One customer may hold several accounts, and one application may generate many transactions. Counting transaction rows and calling them customers inflates the apparent customer base. A population is the complete set of units relevant to the question. A sample is the subset actually examined. The sampling frame is the operational list used to select that sample. The frame may fail to match the intended population when some branches have not uploaded records.
A variable is a characteristic measured for each unit. Branch code, loan amount, processing time and decision status are variables. A record should attach each value to the correct unit and time. A data dictionary states the variable name, definition, unit, permissible values, missing-value rule and source. For turnaround time, specify calendar hours or business hours and the exact starting and ending events. If one branch starts its clock at first enquiry and another starts at receipt of complete documents, their means do not describe the same process. Precise definitions precede meaningful calculation.
3. Functions and practical importance
Statistics condenses many observations into manageable summaries. A frequency distribution reveals how many applications fall into each delay band. A median describes a typical experience. A high percentile identifies a tail that an average may hide. Comparisons across periods reveal changes, while comparisons across similar branches help identify where investigation is useful. Statistical forecasts support cash, staffing and capacity plans. These uses make information easier to act on, provided that the summary matches the decision and preserves material differences.
Suppose two counters both average eight minutes per customer. At the first, most visits take seven to nine minutes. At the second, many take two minutes and some take thirty. Equal averages do not imply equal queue experience. Dispersion and the proportion exceeding a service target add essential information. Similarly, a portfolio's total overdue amount and percentage of overdue accounts answer different questions. Large accounts dominate the first measure, while every account has equal weight in the second. A manager should choose measures deliberately instead of treating a convenient dashboard number as the whole problem. Statistical presentation should reduce confusion, not remove the distinctions needed for a sound decision.
4. Limits of statistical evidence
Statistical methods depend on data quality and the assumptions used to analyse it. A precise calculation on duplicated or selectively omitted records can be misleading. Aggregates may conceal local problems. Association does not demonstrate causation. A rise in digital usage alongside a fall in branch visits is consistent with substitution, but changing customer composition or branch opening hours could also matter. Uncertainty remains even when an estimate is reported to several decimal places. More digits describe calculation precision, not necessarily knowledge of the real system.
Statistics also requires comparable definitions. A change in the complaint classification rule can produce a jump without any change in customer experience. A census removes sampling error for the enumerated finite population, but it does not remove recording mistakes, nonresponse or ambiguity. Qualitative evidence remains valuable. Customer interviews may explain why a numerical pattern appears. Good analysis combines measurement with knowledge of the process. When presenting a result, identify the population, period, source, exclusions and relevant limitations. This is part of interpreting the evidence rather than an optional disclaimer attached after the recommendation.
5. Primary and secondary data
Primary data is collected specifically for the current investigation. Examples include timing customer visits, interviewing applicants or conducting a service survey. It offers control over definitions and collection procedures, but requires time and resources. Secondary data already exists for another purpose, such as account records, published reports or historical call logs. It may be economical and extensive, yet its definitions and coverage may not match the new question. The distinction depends on purpose, not whether data is digital or printed.
Before using a secondary source, examine who collected it, why it was collected, the reference period, measurement units, revisions and coverage. An applications table created for disbursement accounting may omit rejected cases. It therefore cannot directly describe approval rates among all applicants. Before collecting primary data, pilot the questions and procedures. Ask whether respondents interpret each question consistently. A question combining speed and courtesy into one rating is difficult to interpret when service is fast but discourteous. A neutral, specific question produces more useful evidence than wording that invites agreement with the bank's preferred answer. Record the questionnaire version so later comparisons remain understandable.
6. Collection methods and their tradeoffs
Direct observation records behaviour or events as they occur. It can measure queue time accurately when the clock rules are clear, but observation may alter staff behaviour. Interviews allow clarification but can introduce interviewer effects and socially desirable answers. Self-completed questionnaires reach many people cheaply, though response rates and misunderstanding can be problems. Administrative extracts use routine records efficiently, but require reconciliation and knowledge of operational exceptions. None is universally best. Choose a method that measures the variable reliably for the intended population.
For a queue study, a practical protocol records arrival, service start and service finish using anonymous visit identifiers. Waiting time is service start minus arrival. Service time is finish minus service start. Total time is their sum. Record abandoned visits separately rather than silently excluding them. If only completed visits are analysed, long-waiting customers who leave may disappear from the result. A pilot day can expose timestamp errors and unusual events before a longer collection begins. Access should be limited to the information required for the analysis. Educational examples can use invented records, avoiding customer identifiers while preserving the statistical structure.
7. Census, sample and error
A census attempts to collect information for every population unit. A sample examines selected units. A census is attractive for a small population or when the records already exist in reliable form. Sampling is useful when observation is expensive, time is limited or testing is destructive. The choice is about the decision and collection process, not a belief that a larger file is always more accurate. A carefully designed sample can outperform a poorly administered census for a particular estimate.
Sampling error arises because different samples yield different results. Nonsampling errors include undercoverage, nonresponse, incorrect answers, coding mistakes and duplicate records. Increasing sample size generally reduces sampling variation under the same valid design. It does not automatically repair systematic exclusions. If a mobile survey reaches only app users, collecting another ten thousand responses still does not directly represent customers who never use the app. Distinguish missing units from missing values within observed units. Both affect interpretation, but they require different remedies. Chapter A02 develops probability sampling and standard errors. Here the key lesson is that collection quality determines what later formulas can legitimately tell us.
8. Types and scales of data
Qualitative data describes categories, such as payment channel or occupation group. Quantitative data describes numerical amounts or counts. Discrete variables take countable values, such as the number of failed transactions. Continuous variables are modelled as taking values along an interval, such as waiting time or balance. Recorded precision does not change the underlying concept. Waiting time rounded to whole minutes is still a measurement of a continuous duration. Account numbers contain digits but function as labels, so averaging them has no useful interpretation.
Nominal categories have no intrinsic order. Ordinal categories have an order, such as low, medium and high service ratings, but equal numerical distances are not guaranteed. Interval scales support meaningful differences without a natural zero, while ratio scales have a meaningful zero and support ratio comparisons. An elapsed duration of twenty minutes is twice ten minutes. A numerical code of four for a rating is not automatically twice as favourable as a code of two. Choose summaries consistent with the scale. Frequencies suit categories, medians can suit ordered values, and arithmetic operations require a defensible numerical meaning. These distinctions also guide chart selection.
9. Classification before tabulation
Classification groups observations by shared characteristics. Geographical classification groups branches by region. Chronological classification organises observations by time. Qualitative classification groups by categories such as channel. Quantitative classification groups by numerical magnitude, such as transaction value bands. Classes should be mutually exclusive, so each eligible observation enters one category, and collectively exhaustive, so no eligible observation is lost. Provide an explicit unknown category where necessary rather than forcing uncertain records into a convenient class.
For elapsed minutes, intervals 0 to under 5, 5 to under 10, and 10 to under 20 are exclusive classes. An observation of exactly five enters the second class. Inclusive integer classes such as 0 to 4 and 5 to 9 are suitable for integer counts. If rounded measurements are treated as continuous, class boundaries may need adjustment, for example 4.5 between recorded values four and five. State the convention. Open-ended classes such as fifty and above can be useful for reporting, but their midpoint is unknown. They therefore complicate grouped calculations of the mean. Retain raw data whenever possible so grouping choices can be revised.
10. Construct a frequency distribution
Consider twenty application times grouped into 0 to under 5, 5 to under 10 and 10 to under 20 minutes. Suppose the frequencies are six, nine and five. The total frequency is twenty. Relative frequency divides each count by twenty, producing 0.30, 0.45 and 0.25. Percentage frequency multiplies these proportions by one hundred. The resulting percentages sum to one hundred, apart from possible rounding. The distribution shows that the largest count occurs in the middle band, while one quarter of the observations take at least ten minutes.
Less-than cumulative frequencies are six, fifteen and twenty at the upper boundaries five, ten and twenty. Thus fifteen of twenty observations, or seventy-five percent, are below ten minutes. A more-than cumulative distribution works in the opposite direction. At the lower boundaries zero, five and ten the counts are twenty, fourteen and five. Verify that cumulative counts move in the expected direction and end at the total. A cumulative percentage is not the percentage within a single class. Subtract adjacent cumulative values to recover class frequencies. These basic reconciliation checks detect many transcription and denominator mistakes before more sophisticated analysis begins.
11. Tables that answer a clear question
A useful table has a descriptive title, labelled rows and columns, measurement units, a period, totals where meaningful and a source note. A simple table classifies observations by one characteristic. A two-way table cross-classifies two characteristics, such as channel and late-payment status. The cells must refer to the same unit and population. Do not combine transaction counts in some cells with customer counts in others. Footnotes should explain exceptions that materially change interpretation, such as whether reversals are included.
Suppose the digital channel has twenty late payments among two hundred payments, while the branch channel has thirty late among one hundred. Raw late counts are twenty and thirty. Within-channel late rates are ten percent and thirty percent. The combined late rate is fifty divided by three hundred, or about 16.67 percent. Averaging ten and thirty percent gives twenty percent, which wrongly gives equal weight to unequal denominators. A row percentage answers a question conditional on the row category. A column percentage conditions on the column category. Always identify the denominator before comparing rates. Displaying counts next to percentages makes this choice visible to the reader.
12. Select an honest visual display
A bar chart compares categories with separated bars. A histogram displays a numerical frequency distribution with adjacent intervals. When all class widths are equal, count heights give areas proportional to counts. With unequal widths, use frequency density, calculated as frequency divided by class width, so area remains proportional to frequency. For the earlier classes of widths five, five and ten, densities are 1.2, 1.8 and 0.5. Drawing the final bar at height five would exaggerate its visual importance relative to its wider base.
A frequency polygon joins values at class midpoints and can help compare distributions with consistent classes. An ogive plots cumulative frequency against class boundaries. It supports approximate percentile reading. A line chart is appropriate for measurements ordered in time, provided that spacing reflects the time scale. Pie charts require parts of one meaningful whole and become difficult to read with many categories. Label axes and units. A truncated vertical axis can magnify small differences in bars, where length carries the comparison. If a focused line-chart scale is used, make it conspicuous. Visual emphasis should communicate the evidence rather than manufacture a dramatic conclusion.
13. Cleaning and reconciliation case
A branch export contains 1,020 rows. Twenty are duplicate transmissions of already recorded applications. After removing verified duplicates, there are 1,000 unique applications. Forty have missing decision timestamps and sixty remain pending at the reporting cutoff. The remaining nine hundred have measurable completed turnaround. Their mean is a description of completed measurable applications, not automatically the mean for all thousand applications. Pending cases may be slower, making this exclusion important. Report each reconciliation step and preserve the original extract for auditability.
Do not replace missing turnaround with zero, because zero is a valid and very short duration. A missing-value code should remain distinguishable from a measurement. Investigate impossible negative times, unusually large values and inconsistent time zones. An extreme value may be a genuine long delay rather than a data error. Correct it only when evidence supports the correction. Maintain a transformation record that identifies rules, counts affected and the resulting population. Once cleaned, compare totals with the source system and ensure every classified record appears exactly once. Reliable analysis is a chain from the original operational event to the final table, with each change explainable.
14. Interpretation, composition and a learning check
Imagine two branches handling routine and complex applications. Branch North completes ninety percent of routine cases within target and sixty percent of complex cases. Branch South completes eighty-five percent and fifty-five percent respectively. North performs better within both categories. If North handles mostly complex cases and South mostly routine cases, South may nevertheless have the higher overall completion percentage. The aggregate comparison reflects both performance and composition. Compare like with like or standardise to a common mix before attributing the aggregate difference to branch quality.
To check your understanding, return to the complaints example. The first month has thirty complaints and ten thousand transactions. The second has forty-two and twenty thousand. Calculate both counts and rates, state that the rate fell, and then identify what is still unknown. The severity, cause and distribution of complaints may have changed. A lower overall rate does not rule out a serious problem in one channel. A useful recommendation is to investigate categories and causes with consistent definitions. The statistical result informs that investigation. It does not, by itself, select a staffing solution or prove that customer experience improved in every respect.
Guided exercise: recover information from a table
An audit presents less-than cumulative counts of eight below five minutes, twenty below ten minutes, thirty-five below twenty minutes and forty below thirty minutes. Recover the ordinary class frequencies by subtraction. The first class contains eight observations. The second contains twenty minus eight, or twelve. The third contains thirty-five minus twenty, or fifteen. The final class contains forty minus thirty-five, or five. These counts sum to forty, providing a reconciliation check. The percentage taking at least ten minutes is twenty divided by forty, or fifty percent, assuming all observations are nonnegative and below thirty.
The third class has the largest frequency, but its width is ten minutes while the second class is only five minutes wide. Their histogram densities are fifteen divided by ten, or 1.5, and twelve divided by five, or 2.4. The second class therefore has the taller density bar. This does not contradict the larger count in the third class because the third bar is twice as wide. Area, rather than height alone, represents frequency. If a report calls the tallest density bar the class containing the most observations, correct the interpretation by examining both width and height. This exercise connects tabulation, cumulative counts, percentages and visual representation in one internally consistent dataset, while showing why chart labels and class boundaries are essential to sound reading.
Frequency-table laboratory
Use nonnegative integer counts and positive class widths. Counts and widths must have matching lengths.
Practice and review
25 questions. One answer per question, no negative marking. Explanations remain readable without JavaScript.
Local browser storage only; no account synchronisation.
Ready: 25 questions.
References and scope
- IIBF CAIIB Rules & Syllabus 2026, PDF page 5: Official Module A scope; A01 to A09 are Clerk2Officer editorial identifiers.
- NIST statistical handbook: quantitative techniques: Statistical definitions and links to methods.
Original exposition, cases and questions. Formula conventions are stated in the lesson. Automated checks support arithmetic and functionality; they are not independent academic certification.