Business Analytics Methods

Introduction

Business intelligence becomes valuable when data is converted into a defensible basis for action. A dashboard may display sales, customers, costs, or operational events, but the analyst must still decide what question is being asked, what comparison is appropriate, what evidence supports an interpretation, and what action is justified. Business analytics methods provide a disciplined way to make those decisions.

Four forms of analytical thinking are commonly distinguished:

• Descriptive analytics asks what happened.
• Diagnostic analytics asks why it happened or what factors are associated with it.
• Predictive analytics asks what is likely to happen.
• Prescriptive analytics asks what action should be considered under stated objectives and constraints.

These categories are related but not interchangeable. A forecast is not an explanation of cause, and an association is not proof that one variable caused another. Power BI users should therefore combine visual analysis, data modelling, statistical reasoning, domain knowledge, and careful communication.

Learning Objectives

By the end of this chapter, you should be able to:

1. Distinguish descriptive, diagnostic, predictive, and prescriptive analytics.
2. Apply segmentation, comparisons, trends, variance, contribution, and Pareto analysis.
3. Construct and interpret cohort, correlation, forecasting, what-if, and scenario analyses.
4. Select suitable measures and visualisations for business questions.
5. Explain uncertainty, assumptions, data quality, and the limits of causal claims.
6. Use Power BI features responsibly to support managerial decisions.

1. An Analytical Reasoning Framework

A reliable analysis normally follows a sequence:

1. Define the decision. Specify who must decide, what decision is under consideration, and when it must be made.
2. Define the measure. Clarify the numerator, denominator, time period, population, currency, and unit of analysis.
3. Establish a baseline. Compare the current result with a target, prior period, budget, benchmark, or control group.
4. Explore patterns. Use segmentation, trends, contribution, and diagnostic comparisons.
5. Assess uncertainty and alternative explanations. Check data quality, confounding variables, selection effects, and changes in definitions.
6. Model possible futures. Use forecasts, what-if parameters, or scenarios when assumptions are explicit.
7. Recommend and monitor. Connect an action to an expected outcome and define how success will be evaluated.

This sequence prevents a common error: starting with an attractive chart and searching for a story after seeing it. Analysis should begin with a business question and an appropriate comparison.

Four analytical modes

The modes often form a progression, but an organisation may need to return to an earlier mode. For example, a recommendation may reveal that the underlying sales measure was incorrectly defined.

2. Descriptive Methods for Business Dashboards

Segmentation

Segmentation divides observations into meaningful groups so that aggregate results do not conceal important differences. Common segmentation variables include customer type, geography, product category, channel, sales representative, industry, order size, and acquisition month.

A useful segment should have a clear business interpretation and sufficient observations. Too few segments produce little insight; too many create a fragmented dashboard and increase the risk of finding accidental patterns. Segments may be:

• Predefined, such as regions or product categories.
• Rule-based, such as high-value customers defined by annual spend.
• Derived, such as customer groups created by recency, frequency, and monetary value.
• Behavioural, such as users grouped by feature adoption or purchase sequence.

Segment definitions must be stable and documented. If “active customer” changes from one report to another, comparisons become unreliable.

Measures, rates, and denominators

Counts are often misleading without a denominator. For example, 100 complaints may represent poor service in a small branch but excellent performance in a branch serving 100,000 transactions. A rate is generally calculated as:

The numerator and denominator must refer to the same period and compatible population. A conversion rate, for example, might be:

The denominator should not be silently changed from opportunities to website visitors. Percentages can also be unstable when the denominator is small, so dashboards should show volume alongside rate where appropriate.

Comparisons

Comparisons give descriptive results meaning. Important comparison types include:

• Time comparison: current month versus previous month or same month last year.
• Target comparison: actual result versus budget or service-level target.
• Peer comparison: one store, region, or team versus comparable units.
• Index comparison: values expressed relative to a base period.
• Like-for-like comparison: comparison using the same stores, products, or customers in both periods.

For a current value X₁ and comparison value X₀, the absolute change is:

The percentage change, when X₀ ≠ 0, is:

The base must be identified. A rise from 10 to 15 is a 50% increase, whereas a fall from 15 to 10 is a 33.3% decrease. Percentage change is not symmetric.

Trends and time series

A trend is a sustained direction in a measure over time. It should be distinguished from:

• Seasonality: a recurring pattern related to calendar periods.
• Cyclicity: longer, irregular movements associated with economic or business cycles.
• Noise: short-term random variation.
• Step change: a sudden and persistent shift, often following a policy, system, or market change.

A line chart is useful for trends, but the time grain must be appropriate. Daily data may reveal operational variation, while monthly data may be more suitable for strategic planning. Analysts should check missing dates, inconsistent fiscal calendars, and partial current periods.

A moving average can smooth short-term fluctuation. A simple moving average over k periods is:

Here, X_t is the value at time t, and k is the number of periods. The method gives equal weight to the included observations and is descriptive rather than automatically predictive.

3. Diagnostic Analysis

Diagnostic analysis investigates patterns associated with a result. It does not necessarily establish why the result occurred. Useful methods include drill-down, decomposition, variance analysis, contribution analysis, Pareto analysis, cohorts, and correlation.

Variance analysis

Variance analysis compares actual performance with a reference value. For actual value A and budget or target B, the simple variance is:

For a positive measure such as revenue, a positive result may be favourable. For a cost measure, a positive result may be unfavourable. Therefore, the dashboard should label the interpretation rather than relying only on colour.

This formula requires B ≠ 0. If the budget is zero, use an absolute variance or a clearly labelled “not meaningful” result. For volume and price analysis, a revenue change can sometimes be decomposed approximately into price and volume effects. If revenue is R = P × Q, where P is price and Q is quantity, then a basic bridge may compare changes in price and quantity. The exact decomposition depends on the chosen convention and whether interaction effects are separately allocated; analysts should document the convention.

Contribution analysis

Contribution analysis identifies how much each component contributes to a total. For component value x_i and total T, its share is:

This is meaningful only when the components are mutually exclusive and collectively represent the same total. A product’s share of revenue is not the same as its share of profit. The metric must therefore be named precisely.

Pareto analysis

Pareto analysis ranks categories by contribution and calculates cumulative contribution. It is often associated with the 80/20 rule, but 80/20 is an empirical pattern, not a universal law. A Pareto chart combines descending bars with a cumulative-percentage line. It can reveal that a small group of products, customers, or defect types accounts for a large share of the total.

The method is especially useful for prioritisation, but it can mislead when categories have different strategic importance, margins, risk, or controllability. A small customer segment may be strategically important even if its current revenue contribution is modest.

Cohort analysis

A cohort is a group sharing a relevant starting event or period, such as customers acquired in the same month. Cohort analysis follows each group through time and is useful for retention, repeat purchase, product adoption, and subscription behaviour.

A basic retention rate for cohort c at period t can be expressed as:

The denominator must remain defined consistently. Cohort tables should distinguish calendar age from calendar date. A customer acquired in January and one acquired in March may both be at month-three age, but they experienced different market conditions.

Correlation

Correlation measures the strength and direction of a linear association between two quantitative variables. Pearson’s correlation coefficient is:

Here, x_i and y_i are paired observations, x̄ and ȳ are sample means, and r ranges from −1 to +1. The coefficient is unitless and measures linear association. It can be distorted by outliers, non-linear relationships, restricted ranges, or mixed subgroups.

Correlation does not establish causation. Sales and advertising may rise together because both respond to seasonal demand. A third variable, such as market growth, may influence both. A causal claim requires stronger design or evidence, such as randomisation, a credible quasi-experiment, temporal ordering, control of confounders, and a defensible causal model.

4. Predictive Methods

Forecasting

Forecasting estimates future values using historical data and assumptions about the data-generating process. Forecasts may use naïve benchmarks, moving averages, exponential smoothing, regression, or more advanced time-series models.

A naïve forecast sets the next forecast equal to the latest observed value:

It is simple and often useful as a benchmark. A seasonal naïve forecast uses the value from the corresponding previous season. More complex models should demonstrate improvement over such benchmarks.

Forecast accuracy can be evaluated using errors. If actual values are A_t and forecasts are F_t, the mean absolute error is:

where n is the number of evaluated periods. MAE is expressed in the same units as the measure. Mean absolute percentage error is:

MAPE is undefined when A_t = 0 and can be unstable for values close to zero. Forecasts should therefore include prediction intervals or other uncertainty communication where available. A forecast is not a promise; it is an estimate conditional on data and assumptions.

Predictive interpretation

A predictive model can identify variables associated with a future outcome without proving that changing those variables will change the outcome. For example, a churn model may find that low usage predicts cancellation. Increasing usage may or may not reduce churn because low usage could be a symptom of dissatisfaction rather than its cause.

Models should be assessed using data not used for fitting, with attention to data leakage, changing populations, class imbalance, calibration, and fairness. In Power BI, a model output should be accompanied by its prediction date, refresh date, scope, and known limitations.

5. Prescriptive, What-If, and Scenario Analysis

Prescriptive analysis evaluates possible actions against objectives and constraints. It may use business rules, sensitivity analysis, optimisation, simulation, or decision models. At diploma level, transparent scenario analysis is often more useful than an opaque optimisation model.

What-if analysis

A what-if analysis changes one input while holding other assumptions constant. For example, a manager may assess the effect of a 5%, 10%, or 15% price increase. If demand changes with price, the analyst must state whether the demand response is assumed or estimated.

A simple profit model is:

If revenue equals price times quantity and variable cost equals unit variable cost times quantity:

where P is unit price, Q is quantity, V is variable cost per unit, and F is fixed cost for the relevant period. The units must be consistent, such as dollars per month and units per month. This model assumes linear unit economics over the analysed range and excludes other costs unless included explicitly.

If quantity is assumed to respond to price using a constant elasticity e, a simplified relationship may be written as:

This is an assumption-based model, not a universal law. It is most defensible over a limited range and requires P₀ > 0, P₁ > 0, and a clearly interpreted elasticity.

Scenario analysis

Scenario analysis changes several related assumptions together. For example:

• Base case: expected demand, current price, normal costs.
• Upside case: stronger demand and improved conversion.
• Downside case: weaker demand, higher input costs, and delayed delivery.

Scenarios should be internally coherent. An upside case should not combine low demand with unusually optimistic revenue unless there is a specific reason. Scenario labels should communicate that they are constructed possibilities, not assigned probabilities, unless probabilities have been estimated.

Sensitivity and decision quality

Sensitivity analysis changes one or more assumptions to identify which variables most affect the result. A tornado chart can rank the impact of inputs. Decision-makers should consider not only expected profit but also risk, cash flow, legal obligations, service quality, employee effects, and reversibility.

6. Power BI Implementation Principles

Power BI can support these methods through a well-designed semantic model and carefully defined measures.

Model before visualise

A star schema is generally suitable for business reporting: fact tables contain events such as sales or transactions, while dimension tables describe dates, products, customers, and locations. A marked date table supports consistent time intelligence. Relationships, filter direction, granularity, and missing values must be checked before creating measures.

Measures should express business logic consistently. For example, revenue and margin should be measures rather than manually repeated calculations in multiple visuals. KPI names should specify whether they are actual, target, rate, cumulative value, or forecast.

Visual selection

Interactivity should support a question rather than create distraction. Slicers should show their current state, and drill-through pages should preserve context. Tooltips can provide definitions, sample size, and data freshness.

Validation and governance

Before publishing, reconcile totals to an authoritative source, test filters, inspect edge cases, and confirm that refresh timing is suitable. Row-level security should be tested using representative user roles. Sensitive customer or employee data should be minimised and protected. Automated insights should be reviewed by a knowledgeable analyst before being used in a consequential decision.

7. Illustrative Case: Retail Customer Retention

This is a fictional illustrative case. Northstar Home, a multi-channel retailer, observes that overall repeat-purchase rate has declined from 34% to 29% over two quarters. Management asks whether a new loyalty programme should be expanded.

Background and problem

The headline rate combines customers acquired through stores, the website, and marketplaces. Marketing initially concludes that the loyalty programme caused the decline because the decline began after a programme redesign. This is a causal claim that the available dashboard does not yet support.

Analysis

The analyst first defines repeat purchase as at least one additional completed order within 90 days of the first order. Cancelled orders are excluded. The denominator is customers with a completed first order who have had at least 90 days of observation.

The dashboard then applies four analyses:

1. Segmentation: repeat purchase is split by channel, product category, region, and acquisition month.
2. Cohorts: customers are grouped by first-order month and followed for 90 days.
3. Comparison: programme participants are compared with eligible non-participants, while noting that participation is voluntary.
4. Diagnostic review: delivery time, discount level, first-order category, and customer-service contacts are examined.

The cohort view shows that customers acquired during a major marketplace promotion had lower repeat purchase, regardless of programme participation. These customers received heavy first-order discounts and had longer delivery times. Programme participants have a higher observed repeat rate, but they also tend to have larger first orders and may be more engaged before joining.

A correlation between delivery time and repeat purchase is negative, but this does not prove that delivery time caused the outcome. Customers in remote regions may both experience longer delivery and have different purchasing patterns.

Approach and solution

Management decides to run a controlled pilot. Eligible customers are randomly assigned, where operationally and ethically appropriate, to receive either the redesigned loyalty communication or the existing communication. The team predefines the primary outcome, observation window, exclusions, minimum sample size, and monitoring rules. It also tracks delivery time and acquisition channel so that imbalances can be inspected.

The Power BI report contains:

• a definition page for measures and exclusions;
• a cohort retention matrix;
• a channel and category decomposition;
• a pilot results page showing treatment and comparison groups;
• a scenario page estimating financial impact under different adoption and repeat-rate assumptions.

Outcome

The pilot finds a modest improvement in repeat purchase among eligible customers, but the financial benefit is smaller than expected because incentive costs are high. A targeted version aimed at high-margin categories is recommended, subject to continued monitoring. The report does not claim that the original redesign caused the overall decline; instead, it identifies plausible factors and uses a stronger design for the next decision.

Learning points

The case illustrates that a dashboard should distinguish observation from explanation, association from causation, and forecast from commitment. It also demonstrates the importance of cohort definitions, comparable groups, cost-inclusive scenarios, and decision-focused metrics.

8. Limitations, Ethics, and Best Practice

Common analytical failures

• Simpson’s paradox: an aggregate relationship reverses or disappears after subgroup analysis.
• Selection bias: observed participants differ systematically from non-participants.
• Survivorship bias: analysis excludes failed, inactive, or lost entities.
• Data leakage: predictive models use information unavailable at the prediction time.
• Multiple comparisons: examining many segments increases the chance of finding an apparently unusual result by chance.
• Base-rate neglect: a high accuracy percentage may conceal poor performance on a rare outcome.
• Overfitting: a model describes historical noise rather than generalisable structure.
• Metric gaming: people optimise a measured target while harming the underlying objective.

Ethical and governance requirements

Analytics should be lawful, proportionate, transparent, and respectful of affected people. Personal data should have an appropriate purpose and access control. Sensitive attributes should not be used to produce unfair treatment, and proxy variables should be investigated. Automated recommendations should have human oversight when decisions affect employment, credit, insurance, healthcare, or access to essential services.

Analysts should retain a record of data sources, definitions, transformations, assumptions, model versions, refresh dates, and approvals. A result that cannot be reproduced or explained is difficult to govern.

Best-practice checklist

1. Start with a decision and define the population.
2. Document numerator, denominator, time window, and exclusions.
3. Show counts with rates and expose small samples.
4. Use like-for-like comparisons where possible.
5. Inspect segments before making aggregate claims.
6. Treat correlation and model importance as non-causal unless stronger evidence exists.
7. Benchmark forecasts against simple alternatives.
8. Make scenario assumptions visible and test sensitivity.
9. Separate actuals, estimates, forecasts, and targets visually and semantically.
10. Validate totals, permissions, refreshes, and interpretation before publication.

Chapter Summary

Business analytics methods turn dashboard data into structured reasoning. Descriptive analytics establishes what happened through measures, comparisons, segments, trends, contribution, and Pareto analysis. Diagnostic analytics investigates differences and associations using variance analysis, cohorts, decomposition, and correlation, while recognising that association does not prove causation. Predictive analytics estimates future outcomes and must communicate error, uncertainty, and assumptions. Prescriptive analysis evaluates actions through what-if analysis, scenarios, sensitivity, and explicit constraints.

Effective Power BI analysis depends on a sound semantic model, consistent definitions, suitable visuals, transparent measures, and governance. Analysts should check denominators, cohort eligibility, data quality, selection effects, and forecast benchmarks. The central professional responsibility is not merely to produce a persuasive chart, but to communicate what the evidence supports, what remains uncertain, and what decision can responsibly follow.

Key Terms

Descriptive analytics: Analysis that summarises what has happened.

Diagnostic analytics: Analysis of patterns and factors associated with an outcome.

Predictive analytics: Estimation of likely future or unknown outcomes.

Prescriptive analytics: Evaluation of possible actions under objectives and constraints.

Segmentation: Division of observations into meaningful groups.

Baseline: A reference value used for comparison.

Variance: Difference between an actual result and a target, budget, or reference.

Contribution: The amount or share supplied by a component of a total.

Pareto analysis: Ranking components and examining cumulative contribution.

Cohort: A group sharing a defined starting event or period.

Retention rate: The proportion of an initial cohort remaining active at a later time.

Correlation: A numerical measure of linear association between two variables.

Confounder: A variable associated with both an apparent explanatory factor and an outcome.

Forecast: An estimate of a future value based on data and assumptions.

MAE: Mean absolute error, a forecast error measure in the original units.

What-if analysis: Assessment of an outcome after changing one input assumption.

Scenario analysis: Assessment of outcomes under internally consistent sets of assumptions.

Sensitivity analysis: Investigation of how results change when inputs vary.

Data leakage: Use of information in model development that would not be available at decision time.

Causal claim: A claim that changing one factor produces a change in an outcome.

Review Questions

Short Answer Questions

1. Distinguish descriptive analytics from diagnostic analytics.
2. Why must a rate be accompanied by a clearly defined denominator?
3. What is the purpose of cohort analysis?
4. Why does correlation not, by itself, establish causation?
5. State one difference between what-if analysis and scenario analysis.

Long Answer Questions

1. Explain how segmentation, contribution analysis, and Pareto analysis can be combined to prioritise business action. Include limitations.
2. Describe a robust process for conducting variance analysis in a Power BI dashboard.
3. Explain the assumptions and limitations of forecasting, including the role of benchmark models and error measures.
4. Discuss how a star schema, documented measures, and governance improve analytical reliability.
5. Explain why causal claims require stronger evidence than descriptive or correlational findings.

Analytical / Application Questions

1. A retailer’s conversion rate rises from 4% to 5%, but qualified opportunities fall from 20,000 to 8,000. Calculate the percentage change in the conversion rate and explain why the dashboard should show both the rate and the volume.
2. A manager observes that loyalty-programme members purchase more frequently than non-members and proposes expanding the programme. Identify at least three alternative explanations and design a stronger evaluation approach.
3. A business has a monthly forecast of 1,000 units and actual demand of 920 units. Calculate the absolute forecast error and explain what additional information is needed before judging forecast quality.

Multiple Choice Questions with Answers

Question 1

Which question is primarily descriptive?

A. What happened to revenue by region last quarter?

B. Which intervention will maximise profit?

C. What will demand be next month?

D. Did the loyalty programme cause retention to increase?

Correct Answer: A. What happened to revenue by region last quarter?

Question 2

A rate is most interpretable when the analyst also reports its:

A. Chart colour

B. Denominator and underlying volume

C. Number of dashboard pages

D. Font size

Correct Answer: B. Denominator and underlying volume

Question 3

Actual sales are 120, while the budget is 100. What is the percentage variance from budget?

A. 16.7%

B. 20%

C. −16.7%

D. −20%

Correct Answer: B. 20%

Question 4

Which method groups customers by their acquisition month and follows them over time?

A. Pareto analysis

B. Cohort analysis

C. Naïve forecasting

D. Correlation analysis

Correct Answer: B. Cohort analysis

Question 5

Pearson’s correlation coefficient is designed primarily to measure:

A. Linear association

B. Causal impact

C. Forecast error in currency

D. Cumulative contribution

Correct Answer: A. Linear association

Question 6

If actual demand is 920 units and the forecast is 1,000 units, the absolute forecast error is:

A. 80 units

B. −80 units

C. 920 units

D. 1,920 units

Correct Answer: A. 80 units

Question 7

Which statement best describes a what-if analysis?

A. It changes an input assumption to examine the resulting outcome.

B. It proves that an input caused a historical outcome.

C. It removes all uncertainty from a forecast.

D. It ranks categories by cumulative contribution.

Correct Answer: A. It changes an input assumption to examine the resulting outcome.

Question 8

Which issue occurs when a predictive model uses information unavailable at the time a prediction would have been made?

A. Seasonality

B. Data leakage

C. Contribution analysis

D. Like-for-like comparison

Correct Answer: B. Data leakage

Question 9

The 80/20 pattern in Pareto analysis should be treated as:

A. A universal mathematical law

B. An empirical rule that may or may not fit the data

C. A replacement for segmentation

D. Proof of causation

Correct Answer: B. An empirical rule that may or may not fit the data

Question 10

Which design generally provides stronger evidence for whether an intervention caused an outcome?

A. A randomised controlled comparison, where feasible

B. A single before-and-after chart with no controls

C. A correlation calculated from voluntary participants

D. A dashboard showing only the successful cases

Correct Answer: A. A randomised controlled comparison, where feasible