2. Business Problem Framing and Analytical Planning

Introduction and Context

Data analysis rarely fails because an analyst cannot calculate an average, fit a model, or produce a chart. It more often fails because the analysis addresses the wrong problem. A manager may ask, “Why are sales declining?” when the real issue is a change in product mix, a stock-out problem, or a decline in customer visits. Similarly, a request to “build a churn model” may conceal a more important question: which customers should the organization contact, through which channel, and at what cost?

Business problem framing is the disciplined process of translating an organizational concern into a clearly defined analytical problem. Analytical planning then specifies how evidence will be obtained, measured, analysed, interpreted, and converted into action. Together, these activities connect managerial judgement with data and prevent analysis from becoming an isolated technical exercise.

This chapter presents practical and conceptual approaches for defining analytical questions, identifying stakeholders, developing hypotheses, selecting metrics, assessing constraints, and designing an analysis plan. The discussion draws on the CRISP-DM framework, the scientific method, logic models, and decision-oriented analytics. The emphasis is not simply on producing correct results, but on producing results that are relevant, credible, ethical, and usable.

Learning Objectives

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

1. Distinguish a business problem from an analytical problem and a data question.

2. Translate managerial concerns into objectives, questions, hypotheses, metrics, and decision criteria.

3. Identify stakeholders, decision owners, constraints, assumptions, and risks.

4. Apply CRISP-DM and the scientific method to plan an analytical investigation.

5. Design an analysis plan that connects data, methods, outputs, and actions.

6. Evaluate the quality and limitations of an analytical plan.

7. Recognise ethical, governance, and organizational issues in problem framing.

From Managerial Concern to Analytical Question

A business problem is an undesirable situation, opportunity, or decision requiring attention. It is expressed in organizational language: customer retention is weakening, delivery costs are increasing, loan applications take too long, or a new service must be evaluated. An analytical problem is a more precise statement of what evidence must be generated to support a decision.

For example, the business problem may be: “The subscription service is losing customers.” An analytical problem could be: “Estimate the factors associated with cancellation during the next 30 days and identify customer segments for which a retention intervention is economically justified.” The second statement identifies a population, a time horizon, an outcome, explanatory factors, and a decision purpose.

A useful translation sequence is:

Business concern → decision to be made → analytical objective → analytical questions → hypotheses → data and methods → decision criteria.

The decision should be made explicit because the same data can support different analyses. If the decision is whether to close a branch, descriptive analysis of branch revenue may be insufficient; the organization may need a forecast, a cost analysis, and an assessment of alternatives. If the decision is whether to target customers with a promotion, predictive accuracy alone is not enough. The analysis should consider incremental response, campaign cost, margin, and the risk of contacting customers unnecessarily.

A well-framed analytical question is specific enough to guide analysis but open enough to permit investigation. “What is happening to sales?” is too broad. “How did weekly sales by region and product category change after the pricing revision, and which changes are statistically and commercially meaningful?” is more useful. Strong questions commonly specify:

• the unit of analysis, such as customer, transaction, store, employee, or month;
• the outcome or phenomenon of interest;
• relevant explanatory factors or comparisons;
• the population and time period;
• the intended decision or action.

Analysts should also distinguish descriptive, diagnostic, predictive, and prescriptive questions. Descriptive questions ask what happened. Diagnostic questions ask why it may have happened. Predictive questions ask what is likely to happen. Prescriptive questions ask what should be done. These categories are related but require different evidence. A dashboard may describe late deliveries, whereas an experiment may be needed to determine whether a new routing policy causes improvement.

Objectives, Hypotheses, Metrics, and Decision Criteria

Analytical objectives

An analytical objective states what the analysis will deliver and why it matters. It should be connected to an organizational decision. “Analyse customer data” is not an objective. “Assess whether a targeted retention programme can reduce preventable cancellations while maintaining positive contribution margin” is more actionable.

Objectives should be bounded by scope. The plan should specify what is included and excluded, the relevant time period, the population, and the level of detail. Scope discipline protects the project from uncontrolled expansion and clarifies expectations among stakeholders.

Hypotheses

A hypothesis is a testable proposition about a relationship, difference, or mechanism. It is not merely a prediction or an opinion. For example:

• Customers experiencing two or more failed deliveries are more likely to contact support.
• Reducing checkout steps decreases cart abandonment.
• A revised staffing schedule reduces average service waiting time during peak periods.

A hypothesis should identify the expected direction where appropriate and define how it will be tested. Analysts may state a null hypothesis, representing no difference or relationship, and an alternative hypothesis. However, statistical significance should not be confused with practical importance. A very small effect may be statistically detectable in a large dataset but irrelevant to managerial action.

Hypotheses can also be operational or causal. An operational hypothesis may ask whether a relationship exists. A causal hypothesis asks whether changing one factor produces a change in an outcome. Causal claims require stronger designs, such as randomized experiments, natural experiments, or carefully justified quasi-experimental methods. Observational association alone does not establish causality.

Metrics and measurement design

A metric is a defined measure used to monitor performance or evaluate an outcome. A key performance indicator is a metric given particular managerial importance. Metrics require precise definitions. “Customer retention” might mean the percentage of customers active at the end of a period, the percentage renewing at least once, or the proportion remaining active after a specified number of days. Each definition can produce a different result.

A metric specification should include its name, purpose, formula, numerator, denominator, unit, population, time window, data source, aggregation level, owner, and known limitations. It should also identify whether the measure is a count, rate, ratio, average, percentile, index, or monetary value.

Metrics should be balanced. Focusing only on speed may reduce quality; focusing only on revenue may encourage unprofitable sales; focusing only on model accuracy may produce decisions with poor business value. A balanced measurement system may include:

• outcome metrics, such as profit, retention, or service quality;
• driver metrics, such as response time, defect rate, or usage frequency;
• guardrail metrics, such as complaints, fairness indicators, or safety incidents;
• process metrics, such as data completeness and delivery timeliness.

Metric design also requires attention to denominators. The number of complaints means something different from complaints per 1,000 orders. Averages can conceal variation, so medians, percentiles, distributions, and segment-level measures may be needed.

Decision criteria

Decision criteria specify how results will influence action. They may be statistical, financial, operational, strategic, or ethical. For example, a product change may be adopted if it increases conversion by at least 3 percentage points, does not reduce gross margin, and produces no material deterioration in customer complaints. A fraud rule may be accepted only if expected loss reduction exceeds implementation and review costs while false-positive rates remain within an approved limit.

Decision criteria should be agreed before results are examined whenever possible. Pre-specification reduces the risk of changing the standard after seeing favourable or unfavourable findings. It also distinguishes evidence from preference: stakeholders can debate the threshold openly rather than allowing the analysis to conceal a value judgement.

Stakeholders, Governance, and the Decision Environment

Stakeholder analysis is not a ceremonial exercise. Different stakeholders possess different forms of knowledge, authority, incentives, and risk exposure. A senior sponsor may own the decision and budget; operational managers understand processes; subject-matter experts clarify definitions; information technology teams explain system limitations; legal and compliance teams assess obligations; and affected customers or employees may experience consequences that are absent from formal project documents.

A stakeholder map can classify participants according to influence and interest. More importantly, an analytical plan should identify the decision owner, the analysis sponsor, the data owner, the technical delivery team, and the groups affected by the decision. The decision owner is accountable for acting on the evidence; the analyst is accountable for producing a credible analysis, not for making an unauthorized managerial decision.

Stakeholder interviews should explore:

• What decision is pending?
• What would count as useful evidence?
• What actions are feasible?
• Which constraints cannot be violated?
• How are relevant terms currently defined?
• What previous analyses or interventions exist?
• Which groups may benefit or be harmed?

Stakeholders may disagree. Marketing may value customer growth, finance may prioritize margin, and operations may prioritize capacity stability. Such disagreement is not a reason to avoid analysis. It is a reason to document objectives, trade-offs, and decision rights explicitly.

CRISP-DM and the Scientific Method

CRISP-DM—Cross-Industry Standard Process for Data Mining—offers a practical iterative framework with six phases: business understanding, data understanding, data preparation, modelling, evaluation, and deployment. Although developed for data-mining projects, its central value is the connection between business purpose and technical work.

Business understanding defines objectives, success criteria, resources, risks, and an initial plan. This phase is where problem framing occurs.

Data understanding collects initial data, describes structures and variables, explores distributions, and identifies quality issues. It may reveal that the desired outcome is poorly recorded or that important groups are absent.

Data preparation includes selection, cleaning, integration, transformation, feature construction, and documentation. It often requires the greatest practical effort and should not be treated as an invisible preliminary task.

Modelling applies statistical, machine-learning, simulation, optimization, or other methods suitable for the question. The method should follow the decision need, not the popularity of a technique.

Evaluation assesses technical performance, validity, robustness, business usefulness, and unintended consequences. A model can perform well on a test set and still fail the business problem because its predictions arrive too late or cannot be acted upon.

Deployment converts results into a report, dashboard, process change, model service, policy, or experiment. Deployment also includes monitoring, maintenance, retraining, and review.

CRISP-DM is iterative rather than strictly linear. Data discoveries can change the business question; evaluation can reveal that the objective was unrealistic; deployment can expose new data or adoption problems. Iteration should be controlled through documented decisions and versioned artefacts.

The scientific method provides a complementary logic: formulate a question, develop a hypothesis, design an investigation, collect evidence, analyse results, assess uncertainty, and communicate conclusions. In business analytics, the method must also address feasibility, costs, organizational action, and stakeholder values. Together, CRISP-DM supplies project structure while the scientific method supplies disciplined reasoning about evidence.

Designing an Analysis Plan

An analysis plan is a written bridge between the problem statement and the final output. It should be sufficiently detailed that another competent analyst can understand the intended investigation and identify its assumptions.

A strong plan normally contains:

1. Problem statement and decision context: the situation, decision owner, and consequences of inaction.

2. Objective and scope: the analytical purpose, population, time period, exclusions, and deliverables.

3. Questions and hypotheses: the propositions to investigate and the evidence required.

4. Outcome and explanatory variables: precise definitions and expected relationships.

5. Data inventory: sources, access rights, granularity, refresh frequency, quality concerns, and linkage keys.

6. Methodology: descriptive analysis, statistical tests, forecasting, experimentation, segmentation, modelling, or optimization, with reasons for selection.

7. Validation strategy: data splitting, cross-validation, sensitivity analysis, comparison groups, error analysis, or triangulation.

8. Success criteria: technical, commercial, operational, and ethical thresholds.

9. Risks and assumptions: possible biases, missing data, confounding, timing issues, and implementation barriers.

10. Outputs and decision process: reports, dashboards, recommendations, owners, deadlines, and follow-up actions.

A data inventory should distinguish data availability from data suitability. A dataset may be accessible but fail to represent the target population. Historical records may contain survivorship bias, inconsistent definitions, duplicate entities, or changes in data collection practices. Analysts should record data lineage: where a field originated, how it was transformed, and who is responsible for it.

The plan should specify how uncertainty will be communicated. Point estimates without intervals can create false precision. Forecasts should include prediction intervals where appropriate, and model outputs should be accompanied by performance measures relevant to the intended use. Assumptions should be stated in plain language, especially when they affect decisions.

An Illustrative Case: Reducing Missed Appointments

The following is a fictional illustrative case designed to demonstrate the framing process.

Background

A network of outpatient clinics reports that a growing proportion of booked appointments are not attended. Managers believe that missed appointments reduce capacity and increase waiting times for other patients. The network wants to decide whether to introduce reminder messages and a limited overbooking policy.

Situation and initial problem

The initial request to the analytics team is: “Find patients who will not attend.” This request is incomplete. It does not state whether the purpose is to send reminders, overbook appointments, redesign scheduling, or understand barriers to attendance. Each purpose involves different risks and analysis.

Framing the decision

The decision is whether to introduce a reminder intervention for selected appointments and whether any overbooking is safe for particular clinic types. The analytical objective becomes: estimate the baseline probability of non-attendance, assess whether reminders reduce missed appointments, and evaluate the operational and ethical consequences of targeted intervention.

The primary outcome is attendance status for a scheduled appointment. Secondary outcomes include cancellation timing, waiting time for other patients, staff utilization, patient complaints, and any evidence that intervention effects differ across patient groups.

Possible hypotheses are that late-booked appointments have higher non-attendance rates, that previous non-attendance predicts future non-attendance, and that reminders reduce missed appointments. The last hypothesis is causal and should ideally be tested through a randomized or carefully controlled pilot rather than inferred from historical associations.

Analysis and approach

The team first examines appointment records, patient communication preferences, clinic type, booking lead time, appointment time, prior attendance, cancellation history, and relevant operational variables. It checks whether demographic or socioeconomic variables are complete and whether their use is justified. Data quality checks identify duplicate appointments, changes in attendance coding, and clinics that began using a new booking system midway through the period.

Descriptive analysis establishes attendance rates by clinic, appointment type, booking lead time, and time of day. A predictive model may rank appointments by estimated risk, but the ranking is not automatically a recommendation to overbook. The team evaluates calibration, subgroup performance, false positives, and whether high-risk predictions can be acted upon.

For reminders, a pilot randomly assigns eligible appointments to standard practice or a reminder intervention, subject to consent, communication preferences, and operational safeguards. The analysis compares attendance rates and estimates the incremental effect of reminders. It also measures message costs, staff workload, complaints, and any unequal impact.

Decision and outcome

Suppose the pilot shows that reminders improve attendance for some appointment types but have little effect for others. The network may adopt reminders where the expected capacity benefit exceeds cost and where communication is appropriate. It may reject blanket overbooking because prediction uncertainty remains high and the cost of simultaneous attendance is substantial. Instead, it may revise cancellation procedures and offer easier rescheduling.

Learning points

The case demonstrates that prediction is only one part of analytical planning. The organization must define the decision, distinguish association from intervention effect, establish operational and ethical guardrails, and evaluate value rather than accuracy alone. It also shows why a model should not be deployed merely because it is technically feasible.

Challenges, Risks, and Common Mistakes

A frequent mistake is solution-first framing: stakeholders request a dashboard, machine-learning model, or survey before defining the decision. Technology may then determine the question rather than serve it. Another mistake is confusing correlation with causation, especially when historical operational data reflect many simultaneous changes.

Ambiguous definitions create hidden disagreement. Different departments may calculate “active customer,” “on-time delivery,” or “profit” differently. A metric dictionary and data-owner review can reduce this problem. Metric gaming is another risk: when a measure becomes a target, employees may optimize the metric while damaging the underlying objective. Balanced measures and periodic review are safeguards.

Sampling bias, missing data, measurement error, data leakage, and changing definitions can undermine an analysis. Data leakage occurs when information unavailable at the time of decision is used to train or evaluate a model. Analysts should reproduce the information environment that will exist at deployment.

Overly broad scope, unrealistic deadlines, inadequate stakeholder involvement, and failure to plan deployment are organizational risks. A technically excellent analysis may have no impact if no person owns the recommended action. Analysts should therefore identify the smallest useful analysis that can inform an immediate decision while preserving a path for later refinement.

Ethical, Legal, and Professional Considerations

Problem framing determines what data are collected, whose interests are represented, and which outcomes are optimized. Ethical review should begin before modelling. Analysts should ask whether the proposed use is necessary, proportionate, lawful, and fair; whether individuals understand relevant data use; and whether the decision may produce discriminatory effects.

Sensitive personal data require appropriate access controls, minimization, retention rules, and documented purpose. De-identification reduces risk but may not eliminate re-identification possibilities. Data should not be repurposed casually merely because they are available.

Fairness cannot be reduced to a single universal metric. Different fairness criteria may conflict, and the appropriate assessment depends on the decision, legal context, and harms involved. Analysts should examine performance and error patterns across relevant groups, consult affected stakeholders, and provide routes for review or appeal where decisions materially affect people.

Professional integrity requires reporting inconvenient findings, limitations, uncertainty, and failed hypotheses. Charts and statistical tests should not be selected solely to support a preferred narrative. Reproducible code, versioned data transformations, clear documentation, and independent review strengthen accountability.

Digital Tools, Automation, and Emerging Practice

Modern analytical planning is supported by data catalogues, collaborative notebooks, workflow orchestration, semantic layers, experiment platforms, and model-monitoring systems. These tools can improve traceability and repeatability, but they do not replace conceptual framing. Automated pipelines may efficiently produce a precisely calculated answer to an ill-defined question.

Generative AI and automated machine-learning tools can assist with code drafts, variable exploration, documentation, and alternative hypotheses. Their outputs require human verification because they may introduce errors, obscure assumptions, reproduce bias, or expose confidential information. Organizations should define acceptable use, maintain audit trails, and prohibit unreviewed automated decisions in high-impact settings.

Emerging practice increasingly emphasizes decision intelligence: combining descriptive evidence, prediction, optimization, simulation, and human judgement within a decision process. There is also greater attention to causal inference, responsible experimentation, data contracts, metric governance, and continuous monitoring. These developments reinforce a central principle: analytical value is created when evidence is connected reliably and ethically to action.

Best Practices for Analytical Planning

Effective analysts develop the following habits:

• Begin with the decision, not the dataset or preferred tool.
• Define terms, populations, time windows, and denominators explicitly.
• Separate descriptive, predictive, and causal claims.
• Involve domain experts and affected stakeholders early.
• Record assumptions, exclusions, risks, and data limitations.
• Predefine success criteria and decision thresholds where feasible.
• Use balanced outcome, driver, and guardrail metrics.
• Prefer the simplest method that adequately answers the question.
• Validate technical performance and practical usefulness separately.
• Design deployment, monitoring, ownership, and review before modelling.
• Communicate uncertainty and avoid overstating conclusions.
• Treat ethical and governance requirements as design conditions, not final checks.

Chapter Summary

Business problem framing transforms a broad organizational concern into a decision-oriented analytical problem. The process requires explicit objectives, scope, questions, hypotheses, metrics, and decision criteria. Stakeholder analysis clarifies authority, expertise, interests, and potential impacts, while data and governance review establish whether the proposed analysis is feasible and responsible.

CRISP-DM provides an iterative project structure from business understanding through deployment. The scientific method strengthens the logic of questioning, hypothesis development, evidence collection, and evaluation. An analysis plan connects these ideas by documenting data, methods, validation, assumptions, risks, outputs, and actions.

Effective planning distinguishes association from causation, accuracy from usefulness, and data availability from data suitability. It also recognises ethical, legal, operational, and organizational constraints. The most valuable analysis is not necessarily the most technically complex; it is the analysis that produces credible evidence for a clearly owned decision and supports responsible action.

Key Terms

Business problem: An organizational situation, opportunity, or decision requiring attention.

Analytical problem: A precise statement of the evidence needed to support a business decision.

Analytical objective: A bounded statement of what the analysis will deliver and why.

Hypothesis: A testable proposition about a relationship, difference, or mechanism.

Metric: A defined quantitative measure used to monitor or evaluate an outcome.

Key performance indicator: A metric given formal importance in managing performance.

Decision criterion: A rule or threshold used to determine an action based on evidence.

Unit of analysis: The entity about which observations and conclusions are made.

CRISP-DM: An iterative framework comprising business understanding, data understanding, preparation, modelling, evaluation, and deployment.

Scientific method: A disciplined process of questioning, hypothesizing, investigating, analysing, and evaluating evidence.

Data lineage: A record of a data element’s origin, transformations, and ownership.

Data leakage: The use of information in analysis that would not be available at the time of the intended decision.

Confounding: A situation in which another factor influences both an explanatory variable and an outcome, distorting their apparent relationship.

Guardrail metric: A measure used to ensure that improvement in one objective does not cause unacceptable harm elsewhere.

Predictive analysis: Analysis estimating what is likely to happen in the future or for an unobserved case.

Causal analysis: Analysis designed to determine whether changing one factor produces a change in an outcome.

Stakeholder: A person or group that influences, uses, owns, or is affected by an analytical project.

Deployment: The process of embedding analytical outputs into a decision, workflow, product, or operational system.

Metric governance: The policies and responsibilities used to define, maintain, and control organizational measures.

Decision intelligence: The coordinated use of data, models, optimization, simulation, and judgement to improve decisions.

Review Questions

Short Answer Questions

1. What is the difference between a business problem and an analytical problem?

2. Why should decision criteria be defined before results are examined where possible?

3. What information should be included in a metric specification?

4. Name the six phases of CRISP-DM.

5. Why is a predictive model not automatically a prescriptive recommendation?

Long Answer Questions

1. Explain how a managerial concern can be translated into a decision, analytical objective, questions, hypotheses, and metrics.

2. Discuss the role of stakeholders in analytical planning and explain the consequences of ignoring affected groups.

3. Compare the contributions of CRISP-DM and the scientific method to an analytics project.

4. Describe the main components of a professional analysis plan and explain why each is important.

5. Discuss the ethical and governance issues that should be considered before using personal data for predictive decision-making.

Analytical / Application Questions

1. A retailer reports declining profit despite stable sales volume. Frame the problem by proposing one decision, two analytical questions, three hypotheses, and suitable metrics.

2. A bank wants to deploy a loan-default model with high overall accuracy but uneven error rates across customer groups. Design an evaluation and governance plan.

3. A hospital wants to reduce appointment non-attendance. Explain whether it should begin with prediction, experimentation, process redesign, or a combination, and justify your proposed analysis plan.

22. Multiple Choice Questions with Answers

Question 1

Which statement best converts a broad business concern into an analytical problem?

A. Sales are declining and the company should use artificial intelligence.

B. Determine which regions experienced a meaningful decline in margin after the pricing change and identify feasible corrective actions.

C. Create a dashboard containing every available sales variable.

D. Train the most accurate model possible on historical sales data.

Correct Answer: B. Determine which regions experienced a meaningful decline in margin after the pricing change and identify feasible corrective actions.

Question 2

Which question is primarily causal rather than descriptive or predictive?

A. What was the average delivery time last quarter?

B. Which customers are likely to cancel next month?

C. Did the reminder intervention reduce missed appointments?

D. Which product categories had the highest revenue?

Correct Answer: C. Did the reminder intervention reduce missed appointments?

Question 3

Why should a metric include a clearly specified denominator?

A. To make the metric more complex

B. To ensure that rates and comparisons have a meaningful reference population

C. To eliminate the need for data validation

D. To guarantee that the metric is causal

Correct Answer: B. To ensure that rates and comparisons have a meaningful reference population

Question 4

In CRISP-DM, which phase focuses on examining data structure, distributions, and initial quality problems?

A. Business understanding

B. Data understanding

C. Modelling

D. Deployment

Correct Answer: B. Data understanding

Question 5

A model uses information about a customer’s cancellation request to predict whether that customer will cancel. What major problem does this illustrate?

A. Under-sampling

B. Data leakage

C. Metric balancing

D. Randomization

Correct Answer: B. Data leakage

Question 6

Which is the best example of a guardrail metric?

A. Total number of records in a database

B. Model training time only

C. Customer complaints monitored while reducing service waiting time

D. Number of variables used in a model

Correct Answer: C. Customer complaints monitored while reducing service waiting time

Question 7

Why is statistical significance alone insufficient for a business decision?

A. Statistical tests cannot use numerical data

B. A statistically detectable effect may be too small to matter commercially or operationally

C. Statistical significance always indicates biased data

D. Business decisions never require quantitative evidence

Correct Answer: B. A statistically detectable effect may be too small to matter commercially or operationally

Question 8

Who is primarily accountable for deciding whether an organization acts on analytical findings?

A. The decision owner

B. The database administrator in every case

C. The software vendor

D. The visualization tool

Correct Answer: A. The decision owner

Question 9

In the illustrative appointment case, why would historical association between prior non-attendance and future non-attendance not prove that reminders work?

A. Historical data cannot contain appointment records

B. Association does not establish the causal effect of an intervention

C. Reminder messages cannot be measured

D. Predictive models cannot be evaluated

Correct Answer: B. Association does not establish the causal effect of an intervention

Question 10

Which practice most directly improves the credibility and reproducibility of an analysis?

A. Changing success criteria after seeing favourable results

B. Keeping assumptions and data transformations undocumented

C. Versioning analytical artefacts and documenting data lineage and limitations

D. Selecting the most complex available algorithm

Correct Answer: C. Versioning analytical artefacts and documenting data lineage and limitations