Business Analytics for Decision Support

Introduction

Business intelligence (BI) becomes valuable when information improves a decision. Dashboards, reports and data warehouses describe what has happened, but decision support requires a broader analytical process: understanding causes, estimating what may happen next, and identifying appropriate actions. Business analytics provides this process by combining data, statistical reasoning, business knowledge and decision models.

The four commonly used analytical perspectives are descriptive, diagnostic, predictive and prescriptive analytics. They are related but not interchangeable. Descriptive analytics summarizes historical and current data. Diagnostic analytics investigates why an outcome occurred. Predictive analytics estimates future or unknown outcomes. Prescriptive analytics compares feasible actions and recommends, or supports, a course of action.

A sound decision-support system does not replace managerial judgement. It makes assumptions visible, quantifies uncertainty where possible, identifies trade-offs and provides evidence that decision makers can challenge. This chapter explains how these forms of analytics support decisions involving trends, segmentation, forecasting, classification, optimization and scenarios.

Learning objectives

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

• distinguish descriptive, diagnostic, predictive and prescriptive analytics;
• select suitable analytical techniques for business questions;
• interpret trends, segments, forecasts and classification results;
• explain basic optimization and scenario-analysis results;
• recognize assumptions, uncertainty, bias and limitations;
• communicate analytical findings as actionable decision support; and
• evaluate governance, ethical and implementation issues in analytics.

The analytical progression

A decision-support project normally begins with a business decision rather than a technique. For example, “Which customers should receive a retention offer?” is more useful than “Can we build a clustering model?” The decision determines the outcome measure, time horizon, constraints, acceptable risk and required action.

The four analytical questions can be summarized as follows:

These perspectives form a progression, not a strict one-way sequence. A prescriptive model may require descriptive data, diagnostic investigation may reveal a new descriptive measure, and predictive results should be monitored after implementation. More advanced analytics is not automatically better: a simple, well-understood rule may be preferable to a complex model when the decision is low-risk, data are limited or explainability is essential.

From data to decision

A practical decision-support workflow contains seven linked activities:

1. Frame the decision. Define the decision owner, action alternatives, objective, constraints and timing.
2. Define measures. Specify the target variable, performance indicators, population and unit of analysis.
3. Acquire and prepare data. Check completeness, accuracy, consistency, timeliness, duplicates and unusual values.
4. Explore and explain. Examine distributions, trends, segments and relationships before modelling.
5. Model or evaluate alternatives. Select a method appropriate to the question and document assumptions.
6. Communicate uncertainty and action. Present implications, confidence, limitations and recommended next steps.
7. Monitor outcomes. Compare actual results with expectations and revise data, models or decisions.

A useful distinction is between a prediction and a decision. A prediction estimates an outcome, such as a 0.72 probability that an account will churn. A decision determines whether to contact the account, perhaps by comparing the expected benefit with contact cost and customer-experience risk. The prediction is an input; it is not the decision itself.

Descriptive analytics: establishing what happened

Descriptive analytics organizes historical or current observations into meaningful summaries. Typical measures include counts, sums, averages, medians, rates and percentiles. A mean is calculated as:

Here, xᵢ is observation i, n is the number of observations and x̄ is the arithmetic mean. The observations must be measured on a meaningful numerical scale. The mean is sensitive to extreme values; the median, the middle ordered observation, can better represent a highly skewed distribution.

Rates require careful denominators. A conversion rate, for example, is:

The numerator must be included in the denominator, and the period and population must be consistent. Reporting “sales increased by 20%” without stating whether this means revenue, units, transactions or margin can mislead decision makers.

Trend interpretation

A trend is a sustained direction or pattern over time. Analysts should distinguish trend from seasonality, cyclical movement, one-off events and random variation. A moving average can smooth short-term fluctuation:

Here, yₜ is the observation at time t, k is the number of periods in the window and MAₜ is the k-period moving average. All observations must use comparable definitions and intervals. A larger k produces greater smoothing but responds more slowly to recent changes.

A trend should not be interpreted causally merely because two lines move together. For example, advertising expenditure and sales may rise together because both are higher during seasonal peaks. Descriptive analysis identifies a pattern; it does not by itself establish its cause.

Diagnostic analytics: explaining variation

Diagnostic analytics asks why performance differs across time, products, locations, customer groups or processes. Common techniques include drill-down, variance analysis, cohort comparison, contribution analysis, correlation analysis and process investigation.

Suppose actual revenue is below budget. A diagnostic analysis may decompose the variance into changes in volume, price, product mix, returns or exchange rates. This is more useful than a single red indicator because each cause may require a different intervention.

Correlation measures linear association between two variables. The population correlation is commonly represented as:

Cov(X,Y) is covariance and σₓ and σᵧ are the standard deviations of X and Y. The value lies between −1 and +1 when both standard deviations are positive. Correlation is unit-free and describes association, not causation. Confounding variables, reverse causality, selection effects and common time trends can all produce misleading relationships.

Root-cause analysis should combine data with process knowledge. A useful sequence is to identify the observed problem, segment it by relevant dimensions, compare affected and unaffected cases, test plausible explanations and verify the explanation against operational evidence. Analysts should avoid declaring a “root cause” when the evidence only supports a contributing factor.

Segmentation and classification

Segmentation

Segmentation divides a population into groups that are relatively similar for a stated purpose. Business segments may be based on geography, value, behaviour, needs or risk. A segment is useful only if it is measurable, sufficiently stable, interpretable and connected to a different action.

RFM analysis is a common practical framework:

• Recency: how recently a customer purchased;
• Frequency: how often the customer purchased; and
• Monetary value: how much the customer spent.

RFM scores can support targeted communication, but they do not prove customer motivation. Clustering methods can identify groups algorithmically, yet the number and meaning of clusters depend on scaling, variables, distance measure and algorithm choice. Analysts should validate clusters for stability and business usefulness rather than treating software output as an objective truth.

Classification

Classification assigns an observation to a predefined category, such as “likely to renew” or “high credit risk.” A classification model may output a class label or a probability. The decision threshold should reflect the cost of errors.

A confusion matrix distinguishes four outcomes:

Important measures include:

TP, TN, FP and FN denote true positives, true negatives, false positives and false negatives. Accuracy can be misleading when classes are imbalanced. Precision answers, “Of the cases predicted positive, how many were truly positive?” Recall answers, “Of the truly positive cases, how many were found?” A bank investigating fraud may value recall, while a costly sales intervention may prioritize precision.

Classification performance must be evaluated on data not used to fit the model. A train-test split or cross-validation helps estimate performance on unseen cases, but it cannot remove changes in customer behaviour, data leakage or biased labels. A model can be statistically accurate while producing unfair or commercially harmful decisions.

Predictive analytics and forecasting

Predictive analytics estimates future or unknown outcomes from historical patterns and explanatory variables. Regression is often used for numerical outcomes; classification is used for categories; time-series forecasting emphasizes temporal dependence.

A simple linear model is:

Here, y is the outcome, x is a predictor, β₀ is the intercept, β₁ is the expected change in y for a one-unit increase in x under the model, and ε is the unexplained error. The interpretation assumes that the relationship is adequately represented by the model and that other relevant conditions are not distorting the estimate. A fitted relationship is not automatically causal.

Forecast accuracy can be evaluated with measures such as mean absolute error:

Here, yₜ is the actual value, ŷₜ is the forecast and n is the number of forecast periods. MAE is expressed in the same units as the outcome and treats errors linearly. A lower MAE is better when the same data, horizon and evaluation procedure are used. Forecast comparisons must use a realistic time-based validation design; randomly mixing future observations into training data can create leakage.

Forecasts are uncertain. A point forecast of 10,000 units should not be presented as a guaranteed result. Prediction intervals, ranges, probability distributions or alternative scenarios can communicate uncertainty. Uncertainty generally increases with forecast horizon, structural change and limited historical data.

Worked example: interpreting a forecast

A retailer forecasts weekly demand of 1,200 units. Actual demand is 1,260 units. The absolute error is 60 units. If four weekly absolute errors are 60, 40, 80 and 20 units, then:

The result means that the average absolute forecast error over those four weeks was 50 units. It does not mean every future forecast will be within 50 units, nor does it reveal whether errors are systematically too high or too low. Bias should therefore be examined alongside MAE.

Prescriptive analytics: choosing actions

Prescriptive analytics evaluates actions subject to objectives and constraints. Optimization expresses a decision problem using decision variables, an objective function and constraints. A general form is:

x is the vector of decision variables, f(x) is the objective, gⱼ(x) represents constraint j, bⱼ is its limit and F defines feasibility conditions such as non-negativity or integer requirements. Units must be consistent: for example, a labour-hours constraint should not be compared directly with a monetary amount without a valid conversion.

A linear programming example is a product mix problem. Let x₁ and x₂ be units of products A and B. If contribution margins are $30 and $20, the objective is:

If products require 2 and 1 labour hours, with 100 available hours, and 1 and 2 machine hours, with 80 available hours, constraints are:

A solver can identify the feasible plan with the greatest contribution, but the result depends on reliable margins, capacities, demand limits and divisibility assumptions. If products must be whole units, integer restrictions may be required. If demand, prices or capacity are uncertain, scenario or robust approaches may be more appropriate than a single optimized answer.

Scenario and sensitivity analysis

Scenario analysis examines outcomes under alternative assumptions, such as low, central and high demand. It is not the same as assigning a probability to each scenario. Sensitivity analysis changes one or more inputs to determine which assumptions most affect the result.

A decision table can compare actions across states of the world:

If the organization is risk-averse, it may focus on the worst outcome; if probabilities are credible, it may calculate expected value. Expected monetary value is:

pᵢ is the probability of state sᵢ and V(a,sᵢ) is the value of action a in that state. The probabilities must sum to 1, values must use a comparable unit and the decision maker must accept the expected-value framework. EMV can conceal severe downside risk, so it should be supplemented by range, downside and sensitivity information.

Integrated illustrative case: FreshCart delivery planning

Background. FreshCart, an illustrative online grocery retailer, experiences late deliveries and idle delivery capacity. Management wants to improve service without increasing the fleet substantially.

Problem. The business needs to identify high-risk delivery periods, understand causes, forecast demand and allocate drivers.

Analysis. Descriptive dashboards show lateness by hour, zone and product category. Diagnostic analysis reveals that lateness is concentrated in two zones during Friday evenings, when order volume and traffic both increase. The relationship between order volume and lateness is positive, but analysts avoid claiming that volume alone causes lateness because weather and driver availability also vary.

A classification model estimates the probability that an order will be late. Its recall is preferred over accuracy because missing a genuinely late order creates substantial customer dissatisfaction. However, the model is tested for performance across zones and customer groups to detect unequal error rates. Operations staff review a sample of false positives and false negatives.

A time-series forecast estimates demand by zone and half-hour interval. The forecast is reported with ranges rather than as a single certain number. The prescriptive model assigns available drivers to intervals and zones, subject to driver hours, vehicle capacity, delivery windows and minimum service levels. Scenarios test ordinary traffic, severe traffic and a promotional weekend.

Approach and solution. FreshCart uses the forecast as an input, not a commitment. When predicted demand exceeds capacity, the decision system compares temporary drivers, delivery-slot restrictions and customer incentives. Managers select a plan after reviewing cost, service impact and scenario sensitivity. The model is monitored weekly, with actual lateness and forecast error fed back into the process.

Outcome. In this illustrative case, the major improvement is not simply a more accurate model. It is the integration of measurement, explanation, forecasting, optimization and operational review. The case demonstrates that an analytical recommendation is credible only when the data, assumptions, constraints and implementation responsibilities are explicit.

Assumptions, uncertainty and limitations

Analytics can fail for reasons that are not mathematical. Important limitations include:

• Data quality: missing, duplicated, outdated or inconsistent records can distort results.
• Representativeness: historical data may not represent new customers, new markets or unusual events.
• Measurement error: proxies may not measure the intended business concept.
• Model assumptions: linearity, stable relationships, independence or constant variance may be inappropriate.
• Concept drift: customer behaviour, competitors, regulations and processes change over time.
• Confounding: an observed association may reflect an omitted variable.
• Overfitting: a model may fit historical noise and perform poorly on new cases.
• Optimization limits: the mathematically optimal plan may be operationally infeasible or ethically unacceptable.
• Human and organizational factors: users may misunderstand, ignore or strategically manipulate indicators.

A professional analyst states what is known, what is estimated, what is assumed and what would change the conclusion. Sensitivity analysis, out-of-sample testing, independent review and monitoring are practical controls.

Governance, ethics and responsible use

Decision-support systems often process personal, financial or operational data. Organizations should apply data minimization, purpose limitation, access control, retention rules and appropriate security. Legal requirements differ by jurisdiction, but privacy and data-protection obligations commonly require a lawful purpose, transparency and safeguards for personal information.

Fairness requires more than removing sensitive attributes. Proxies may reproduce disadvantage, and unequal error rates may affect groups differently. Analysts should assess data provenance, label quality, subgroup performance, explainability and avenues for review or appeal. High-impact decisions should retain meaningful human oversight; “the algorithm decided” is not an adequate accountability mechanism.

Governance should assign ownership for data, models, decisions and monitoring. Model documentation should record intended use, exclusions, training data, performance, limitations, version and approval. Generative AI can assist with querying, summarizing or code drafting, but outputs require verification because generated content may be inaccurate, confidential or biased. AI should not be allowed to conceal uncertainty or bypass established controls.

Best practices for decision-support analytics

1. Begin with a decision and action, not a fashionable algorithm.
2. Define measures, denominators, time periods and data ownership.
3. Establish a baseline before claiming improvement.
4. Separate training, validation and test information appropriately.
5. Use business-relevant metrics and error costs.
6. Present ranges, assumptions and sensitivity, not only point estimates.
7. Validate findings with domain experts and affected users.
8. Document reproducible data and modelling procedures.
9. Monitor outcomes, drift, fairness and unintended consequences.
10. Make recommendations proportionate to the quality and stakes of the evidence.

Chapter Summary

Business analytics supports decisions through four connected perspectives. Descriptive analytics summarizes what happened; diagnostic analytics investigates why; predictive analytics estimates future or unknown outcomes; and prescriptive analytics evaluates actions under objectives and constraints. Trends require separation of seasonality and random variation. Segmentation is useful only when groups support meaningful, measurable actions. Classification must be evaluated with metrics suited to error costs, while forecasting requires time-aware validation and explicit uncertainty. Optimization can clarify trade-offs but is only as sound as its data, assumptions and constraints. Scenario and sensitivity analysis reveal how recommendations change under uncertainty. Responsible decision support also requires privacy, fairness, explainability, governance, human oversight and continuous monitoring.

Key Terms

Descriptive analytics: Analysis that summarizes historical or current performance.

Diagnostic analytics: Analysis that investigates possible reasons for an observed outcome.

Predictive analytics: Analysis that estimates a future or unknown outcome.

Prescriptive analytics: Analysis that evaluates or recommends actions.

Decision variable: A controllable quantity selected in an optimization model.

Objective function: The quantity an optimization model maximizes or minimizes.

Constraint: A condition limiting feasible decisions.

Segmentation: Division of a population into groups for analysis or action.

Classification: Assignment of observations to predefined categories.

Precision: The proportion of predicted positive cases that are truly positive.

Recall: The proportion of actual positive cases correctly identified.

Confusion matrix: A table of actual and predicted classification outcomes.

Forecast: An estimate of a future value or event.

Moving average: An average calculated over a rolling time window.

Mean absolute error: Average absolute difference between actual and predicted values.

Scenario analysis: Examination of outcomes under alternative assumptions.

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

Overfitting: Fitting historical noise rather than generalizable patterns.

Concept drift: Change in the relationship between inputs and outcomes over time.

Model governance: Policies and controls for responsible model development and use.

Review Questions

Short Answer Questions

1. Distinguish descriptive, diagnostic, predictive and prescriptive analytics.
2. Why can a high classification accuracy be misleading for an imbalanced dataset?
3. What is the difference between a forecast and a decision?
4. State two limitations of correlation analysis.
5. What is the purpose of sensitivity analysis?

Long Answer Questions

1. Explain a complete workflow for developing a business decision-support solution, from framing the decision to monitoring outcomes.
2. Discuss how a manager should interpret a forecast while recognizing uncertainty and model limitations.
3. Explain the role of decision variables, objectives and constraints in an optimization model.
4. Compare segmentation and classification, including their purposes, methods and evaluation concerns.
5. Discuss privacy, fairness, explainability and human oversight in analytical decision support.

Analytical / Application Questions

1. A retailer has high model accuracy but very low recall for fraud. Explain why this may be unacceptable and recommend additional evaluation measures.
2. A forecast shows increasing demand during a promotional period. Design a scenario analysis that considers capacity, cost and service risk.
3. An optimization model recommends a plan that violates an informal operational practice. Explain how you would investigate the conflict and decide whether to revise the model or the practice.

Multiple Choice Questions with Answers

Question 1

Which analytical perspective primarily asks why an outcome occurred?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: B. Diagnostic analytics

Question 2

A forecast has absolute errors of 10, 20, 30 and 40 units. What is its MAE?

A. 20 units

B. 25 units

C. 30 units

D. 100 units

Correct Answer: B. 25 units

Question 3

Which measure answers: of the cases predicted positive, how many were truly positive?

A. Recall

B. Accuracy

C. Precision

D. Specificity

Correct Answer: C. Precision

Question 4

What is a central risk of randomly mixing future observations into a forecasting model’s training data?

A. Under-counting

B. Data leakage

C. Lower dimensionality

D. Seasonal adjustment

Correct Answer: B. Data leakage

Question 5

In an optimization model, what specifies the quantity to be maximized or minimized?

A. Constraint

B. Decision variable

C. Objective function

D. Validation set

Correct Answer: C. Objective function

Question 6

Which statement about correlation is correct?

A. It always proves causation

B. It measures only categorical variables

C. It describes association and does not by itself prove causation

D. It must always be greater than zero

Correct Answer: C. It describes association and does not by itself prove causation

Question 7

What is the main purpose of scenario analysis?

A. To hide uncertainty

B. To examine outcomes under alternative assumptions

C. To guarantee a forecast

D. To remove all constraints

Correct Answer: B. To examine outcomes under alternative assumptions

Question 8

Which issue occurs when a model fits historical noise and performs poorly on new data?

A. Overfitting

B. Segmentation

C. Aggregation

D. Normalization

Correct Answer: A. Overfitting

Question 9

Why might recall be especially important in fraud detection?

A. Missing genuine fraud can be costly

B. False positives are impossible

C. Fraud data are always balanced

D. Recall measures monetary profit directly

Correct Answer: A. Missing genuine fraud can be costly

Question 10

Which practice best supports responsible analytical decision-making?

A. Treating model output as unquestionable

B. Removing all human review

C. Documenting assumptions, limitations and monitoring requirements

D. Using the most complex model available

Correct Answer: C. Documenting assumptions, limitations and monitoring requirements