Foundations of Statistical and Predictive Analytics

Introduction and Context

Business analytics is the disciplined use of data, statistical reasoning, computational methods and domain knowledge to support better decisions. Organisations collect data from transactions, websites, sensors, customer interactions, financial systems and operational processes. Data alone, however, does not create value. Value emerges when data is transformed into reliable evidence and then connected to an appropriate managerial action.

Statistical and predictive analytics provide two closely related foundations. Statistics helps analysts describe variation, estimate unknown quantities, test claims and assess uncertainty. Predictive analytics uses historical and current data to estimate future or unknown outcomes. Both require careful problem definition, appropriate data, valid assumptions and responsible interpretation.

An analyst must therefore do more than select a software package or construct a model. The analyst must understand the business context, identify the population being studied, distinguish a sample from the wider population, recognise the difference between a parameter and a statistic, and communicate what the evidence does—and does not—support.

This chapter introduces four forms of analytics—descriptive, inferential, predictive and prescriptive—and presents the CRISP-DM lifecycle as a practical framework for conducting analytics projects.

Learning Objectives

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

1. Explain the purposes and differences of descriptive, inferential, predictive and prescriptive analytics.
2. Distinguish populations, samples, parameters, statistics, variables and observations.
3. Calculate and interpret basic descriptive measures.
4. Frame a broad business concern as a precise analytical question.
5. Explain the stages, iterative nature and governance requirements of CRISP-DM.
6. Identify common sources of bias, uncertainty, leakage and misuse in analytics.
7. Relate analytical results to business decisions and professional responsibilities.

1. The Four Forms of Analytics

Analytics is often represented as a progression from understanding the past to recommending action. The categories overlap, but each has a distinct primary purpose.

Descriptive analytics: What happened?

Descriptive analytics summarises observed data. It answers questions such as:

• How many units were sold last month?
• What was the average delivery time?
• Which products generated the greatest revenue?
• How did customer complaints vary by region?

Typical outputs include tables, dashboards, frequency distributions, charts and summary statistics. Descriptive analytics does not, by itself, establish why an event occurred or what will occur next. A sales dashboard may show that revenue declined, but it cannot automatically establish whether the decline was caused by price, competition, stock shortages or seasonal demand.

Important descriptive measures include the arithmetic mean, median, range, variance and standard deviation. For observations x₁, x₂, …, xₙ, the sample mean is:

Here, x̄ is the sample mean, xᵢ is observation i, and n is the number of observations. The mean is appropriate when numerical values are meaningfully additive and extreme values do not distort interpretation excessively.

The sample variance is:

The sample standard deviation is:

The denominator n − 1 is used when s² estimates the variance of a wider population from a sample. If the analyst is describing the complete population rather than estimating from a sample, the population variance uses N in the denominator and the population mean μ:

The standard deviation is expressed in the same units as the original variable; variance is expressed in squared units. This distinction matters when communicating results.

Inferential analytics: What can be concluded beyond the observed data?

Inferential analytics uses sample data to draw conclusions about a population. It includes estimation, confidence intervals, hypothesis tests and experimental analysis. For example, a retailer may survey 500 customers and use the results to estimate satisfaction among all customers.

Inference requires uncertainty to be made explicit. A sample proportion p̂ estimates a population proportion p, but different samples may produce different values. A confidence interval gives a range of plausible values under specified assumptions. For a large random sample, an approximate confidence interval for a population proportion is:

Here, p̂ is the sample proportion, n is the sample size, and z* is the standard-normal critical value selected for the desired confidence level—for example, approximately 1.96 for a 95% interval. This approximation is most defensible when observations are independently sampled and the sample is sufficiently large for the normal approximation to be reasonable.

A 95% confidence interval should not be interpreted as saying that there is a 95% probability that a fixed population parameter lies inside the particular interval. In the frequentist interpretation, the procedure produces intervals that contain the true parameter in approximately 95% of repeated samples, assuming the model and sampling process are appropriate.

Inference is weakened by convenience sampling, non-response, measurement error, dependence between observations and poorly specified populations. A very large biased sample can produce a highly precise but systematically wrong estimate.

Predictive analytics: What is likely to happen?

Predictive analytics estimates an unknown or future outcome using historical patterns and explanatory variables. Examples include:

• predicting whether a customer will cancel a subscription;
• forecasting next month’s demand;
• estimating the probability that an invoice will be paid late;
• predicting equipment failure from sensor readings.

A predictive model may be statistical, machine-learning based or a combination of both. Common methods include linear regression, logistic regression, decision trees, random forests, gradient boosting and time-series models.

A simple linear prediction model can be written as:

Here, ŷ is the predicted value of the outcome, x is a predictor, β₀ is the intercept and β₁ is the estimated change in predicted y for a one-unit increase in x. The model describes an estimated relationship; it does not automatically prove that changes in x cause changes in y.

Predictive performance must be evaluated on data that were not used to fit the model. For a set of m test observations, mean absolute error is:

Here, yᵢ is the observed outcome and ŷᵢ is the prediction. MAE is in the same units as the outcome and gives equal weight to each absolute error. Other measures, such as root mean squared error, accuracy, precision, recall and area under the ROC curve, may be suitable depending on the problem and consequences of errors.

Prediction is not the same as explanation. A variable can improve prediction without representing a meaningful causal mechanism. Predictive performance may also deteriorate when economic conditions, customer behaviour, regulations or operating processes change.

Prescriptive analytics: What should be done?

Prescriptive analytics combines predictions, objectives, constraints and possible actions to recommend or evaluate decisions. Examples include:

• selecting delivery routes subject to vehicle capacity;
• determining inventory replenishment quantities;
• allocating a marketing budget across channels;
• deciding which customers should receive retention offers.

Prescriptive analysis often uses optimisation, simulation, decision analysis or policy evaluation. A simplified optimisation problem may be represented as:

Here, a represents a possible action or set of actions, U(a) is the organisation’s objective function—such as profit or service value—and A is the set of feasible actions satisfying operational, legal and ethical constraints. A recommendation is only as useful as its objective, assumptions and constraints. Maximising short-term revenue while ignoring customer welfare, fairness or regulatory requirements may produce an unacceptable decision.

Comparing the four forms

These forms are complementary. A demand-planning project may begin with descriptive analysis of historical sales, use inference to estimate seasonal effects, build a forecast, and finally prescribe an inventory policy.

2. Statistical Foundations

Population and sample

A population is the complete set of units relevant to a question. The units may be people, transactions, stores, machines, accounts or time periods. The target population is the population to which the analyst wishes to generalise. The accessible population is the portion that can actually be reached or observed.

A sample is a subset of the population. Sampling may be necessary because measuring every unit is expensive, slow or impossible. Probability sampling gives units a known, non-zero chance of selection and supports stronger statistical generalisation. Simple random sampling, stratified sampling and cluster sampling are examples. Non-probability methods such as convenience or voluntary-response sampling may be practical, but they can produce substantial selection bias.

The sampling frame is the operational list from which a sample is selected. If the frame excludes inactive customers, rural branches or customers without digital contact details, the resulting study may not represent the stated target population.

Parameters and statistics

A parameter is a numerical characteristic of a population, such as the population mean μ, population proportion p or population standard deviation σ. A statistic is a numerical quantity calculated from sample data, such as the sample mean x̄, sample proportion p̂ or sample standard deviation s.

Parameters are commonly unknown; statistics are observed or calculated. Statistical inference uses a statistic to estimate or test a claim about a parameter. Confusing these concepts leads to overconfident reporting—for example, presenting the average of surveyed customers as though it were automatically the average of every customer.

Variables, observations and measurement scales

A variable is a characteristic that can take different values across units. An observation is one recorded value of a variable for one unit. Variables may be:

• Categorical: labels such as region or product type.
• Binary: two categories, such as paid/not paid.
• Ordinal: ordered categories, such as low, medium and high.
• Discrete numerical: countable values, such as number of complaints.
• Continuous numerical: measurements on a continuum, such as delivery time.

Nominal categories have no inherent order. Ordinal categories have order but unequal distances cannot generally be assumed. Interval scales have meaningful differences but no true zero in the measurement system. Ratio scales have meaningful differences and a meaningful zero, allowing statements such as “twice as much.” The measurement scale affects which summaries and statistical methods are defensible.

Worked example: describing delivery times

Suppose five deliveries took 2, 3, 3, 4 and 8 days.

The mean is:

The median is 3 days because it is the middle ordered value. The range is 8 − 2 = 6 days. The deviations from the mean are −2, −1, −1, 0 and 4; their squared deviations sum to 22. If these five deliveries are treated as a sample, the sample variance is:

The sample standard deviation is approximately:

The mean is greater than the median because the eight-day delivery is an unusually high value. Reporting both mean and median therefore gives a more informative account than reporting either alone.

3. From Business Concern to Analytical Question

A business concern is usually broad: “Customer retention is poor,” “Operations are inefficient” or “Marketing is expensive.” An analytical question converts that concern into a question that can be answered with defined data and methods.

A useful framing process is:

1. State the decision. What decision will the analysis support?
2. Identify the decision-maker and time horizon. Who will act, and when?
3. Define the unit of analysis. Is it a customer, order, store, product or week?
4. Specify the outcome. What exactly will be measured or predicted?
5. Identify candidate drivers. Which variables may explain or predict the outcome?
6. Define the population and scope. Which customers, locations and dates are included?
7. Specify success criteria. What improvement or model performance would be useful?
8. Identify constraints. Consider cost, privacy, fairness, latency, staffing and regulation.

For example, “Reduce churn” is incomplete. A stronger formulation is: “Among active monthly subscribers in the domestic market, which customers are most likely to cancel within the next 30 days, and can the retention team contact high-risk customers within a budget of 2,000 offers per week?” This formulation identifies a population, prediction horizon, action and resource constraint.

The distinction between association and causation is essential. “Customers who use the mobile app churn less” is an association. It does not establish that forcing all customers to use the app will reduce churn. Customers who use the app may differ in tenure, income, service plan or engagement. Causal claims require appropriate designs, such as randomised experiments or credible quasi-experimental methods.

A well-formed analytical question also prevents metric substitution. If the goal is profitable retention, maximising prediction accuracy alone may be inadequate. The analysis should consider the value of correctly contacting a customer, the cost of an offer, false positives and possible adverse effects.

4. The CRISP-DM Analytics Lifecycle

CRISP-DM—Cross-Industry Standard Process for Data Mining—is a widely used framework for organising analytics work. It has six phases:

1. Business understanding
2. Data understanding
3. Data preparation
4. Modelling
5. Evaluation
6. Deployment

The phases are iterative rather than strictly linear. Evaluation may reveal that the business objective was misunderstood; data preparation may expose an unavailable variable; deployment may reveal that predictions arrive too late to support action.

Business understanding

This phase defines the business objective, analytical objective, scope, stakeholders, constraints, risks and success criteria. Deliverables may include a problem statement, decision map, project charter and initial risk assessment.

A business objective might be “reduce avoidable delivery failures.” The analytical objective could be “predict the probability of a delivery failure at the time an order is scheduled.” These are related but not identical. The latter must specify the prediction time because using information that becomes available only after delivery would create leakage.

Data understanding

The analyst identifies available sources, examines structure and quality, and assesses whether the data can answer the question. Activities include data inventory, metadata review, exploratory analysis, missingness assessment, duplicate checks and preliminary bias analysis.

Questions include:

• Who created the data and for what original purpose?
• What does each row represent?
• Are timestamps reliable and consistently defined?
• Which values are missing, impossible or duplicated?
• Does the data cover the target population?
• Are labels measured consistently over time?

A data dictionary should define fields, units, permitted values, provenance and update frequency.

Data preparation

Data preparation may include joining sources, correcting formats, handling missing values, encoding categories, constructing features, aggregating records and separating training, validation and test data. Every transformation should be documented and reproducible.

Time-dependent problems require special care. Randomly splitting transactions may allow future information to enter the training set. A time-based split—training on earlier periods and testing on later periods—often provides a more realistic estimate of future performance.

Modelling

The analyst selects methods appropriate to the outcome, data structure, decision context and interpretability requirements. A baseline model should usually be established before complex methods. For example, a demand forecast may first be compared with a naive forecast based on the previous period.

Model selection should consider more than predictive accuracy. Interpretability, computational cost, stability, fairness, maintainability and compatibility with operational systems can determine whether a model is useful.

Evaluation

Evaluation asks two different questions:

1. Does the model perform adequately on unseen data?
2. Does the result support the original business decision?

A classification model may have high accuracy when the event is rare but fail to identify most positive cases. Confusion-matrix measures should therefore be selected according to consequences. Precision is the proportion of predicted positives that are actually positive; recall is the proportion of actual positives that are identified:

TP denotes true positives, FP false positives and FN false negatives. These measures assume the denominators are non-zero and refer to a defined classification threshold.

Evaluation should include sensitivity analysis, subgroup checks, calibration where probabilities are used, and comparison with operational benchmarks. A model can be statistically strong but operationally useless if staff cannot act on its output.

Deployment

Deployment means putting analytical outputs into use. It may involve a dashboard, scheduled report, application programming interface, workflow alert or policy. Deployment planning should define ownership, monitoring, retraining conditions, access controls, documentation and retirement procedures.

A deployed model can degrade through data drift, concept drift, process changes or strategic responses by users. Monitoring should therefore track data quality, prediction distributions, performance where labels become available, subgroup outcomes and business impact.

5. Data Quality, Bias and Responsible Analytics

Data quality is multidimensional. Accuracy concerns whether values represent reality; completeness concerns missing values; consistency concerns agreement across systems; timeliness concerns availability when needed; validity concerns adherence to permitted formats and meanings; and uniqueness concerns duplicate records.

Bias may arise at any stage. Sampling bias occurs when observed units differ systematically from the target population. Measurement bias occurs when a variable is recorded inaccurately or differently across groups. Label bias occurs when the outcome reflects prior human decisions rather than the underlying phenomenon. Historical data may encode institutional disadvantage. Removing a sensitive attribute does not necessarily remove bias because other variables may act as proxies.

Privacy and security are also central. Analysts should collect and use only data necessary for a legitimate purpose, control access, protect sensitive information, document retention periods and comply with applicable law and organisational policy. De-identification reduces risk but may not guarantee anonymity when datasets can be linked.

Responsible analytics requires transparency about data sources, assumptions, uncertainty and limitations. Automated decisions should have appropriate human oversight, especially in high-impact areas such as employment, credit, insurance, healthcare and education. A model should not be deployed merely because it is technically feasible.

6. Illustrative Case: Reducing Subscription Churn

Background

A fictional digital-learning company, LearnPlus, offers monthly subscriptions. Management believes churn has increased and wants to improve retention without sending costly discounts to every customer.

Problem

The initial request—“build a churn model”—is too vague. The analytics team reframes it as: “Using information available at the end of each billing period, estimate each active subscriber’s probability of cancelling within the next 30 days, and identify a contact policy that improves expected contribution margin within the retention budget.”

The unit of analysis is a subscriber-month. The target population is active domestic subscribers. The outcome is cancellation within 30 days. The team excludes information recorded after the prediction date to prevent leakage.

Analysis and approach

During data understanding, the team finds subscription history, login activity, support contacts, payment status and plan type. It discovers that cancellation codes are missing for some historical records and that one region entered support data only after a new system was introduced. These limitations are documented rather than hidden.

The team prepares a time-based dataset: earlier months are used for training, a later month for validation and the most recent month for testing. A simple baseline is compared with logistic regression and a tree-based model. Evaluation includes recall among high-risk subscribers, precision at the available contact capacity, calibration of predicted probabilities and performance by region and plan type.

The model is not treated as a causal explanation. High support-contact frequency predicts churn, but the team does not conclude that support contacts cause cancellation. Instead, the result may indicate unresolved problems or dissatisfaction.

Solution and outcome

The company deploys a weekly list of high-risk subscribers, subject to an eligibility rule that excludes customers who have already received a recent offer. Staff record the action and outcome. The retention programme is evaluated using a controlled pilot rather than assuming that all retained customers were saved by the model.

After deployment, the team monitors missingness, score distributions, contact rates, redemption, incremental retention and profitability. It also reviews whether particular groups receive substantially more contacts or poorer outcomes. The model is scheduled for review if data definitions change or performance falls below an agreed threshold.

Learning points

The case illustrates that successful analytics depends on more than model accuracy. It requires a precise decision question, a correctly defined prediction time, leakage control, realistic validation, operational capacity, causal caution, ethical review and post-deployment monitoring.

7. Professional Best Practices and Contemporary Developments

Effective analysts maintain a clear chain from decision to data, method, result and action. They establish a baseline, document assumptions, use reproducible workflows, separate development from evaluation data, and communicate uncertainty in plain language. They avoid unnecessary complexity and explain why a selected method is appropriate.

Modern analytics increasingly combines cloud data platforms, automated pipelines, interactive dashboards and machine-learning systems. Generative AI may assist with code drafts, documentation or exploratory queries, but generated outputs require verification. It should not be trusted to invent data definitions, make unsupported causal claims or expose confidential information. Human accountability remains essential.

The movement toward responsible and trustworthy AI has increased attention to model cards, data sheets, explainability, robustness, fairness assessment and model-risk management. These practices are not substitutes for statistical reasoning; they operationalise it. A transparent but inaccurate model is not acceptable, and an accurate model with unacceptable discriminatory impact is not acceptable either.

Chapter Summary

Statistical and predictive analytics transform data into evidence for decision-making. Descriptive analytics summarises observed events; inferential analytics generalises from samples to populations while representing uncertainty; predictive analytics estimates future or unknown outcomes; and prescriptive analytics evaluates or recommends actions under objectives and constraints.

A population is the complete set relevant to a question, whereas a sample is a subset. Parameters describe populations, while statistics describe samples and may be used to estimate parameters. Sound analysis depends on understanding variables, measurement scales, sampling processes and data quality.

Business problems must be translated into precise analytical questions that specify the decision, unit of analysis, outcome, population, time horizon, constraints and success criteria. CRISP-DM provides an iterative lifecycle consisting of business understanding, data understanding, data preparation, modelling, evaluation and deployment. Bias, leakage, uncertainty, privacy, fairness and monitoring must be addressed throughout the lifecycle.

Key Terms

Analytics: The systematic use of data, methods and reasoning to support decisions.

Descriptive analytics: Analysis that summarises observed or historical data.

Inferential analytics: Analysis that uses sample evidence to draw conclusions about a population.

Predictive analytics: Analysis that estimates future or unknown outcomes.

Prescriptive analytics: Analysis that recommends or evaluates actions under objectives and constraints.

Population: The complete set of units relevant to an analytical question.

Sample: A subset of a population used for analysis.

Parameter: A numerical characteristic of a population.

Statistic: A numerical quantity calculated from sample data.

Variable: A characteristic that can take different values across units.

Observation: A recorded value for a unit and variable.

Sampling frame: The operational list or source from which a sample is selected.

Bias: Systematic deviation that causes results to differ from the truth or target population.

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

CRISP-DM: A six-phase framework for planning and conducting data-mining and analytics projects.

Baseline model: A simple reference method against which more complex models are compared.

Precision: Among predicted positive cases, the proportion that is actually positive.

Recall: Among actual positive cases, the proportion correctly identified.

Model drift: Deterioration or change in model behaviour as data, relationships or operating conditions change.

Data dictionary: Documentation defining data fields, meanings, formats, units and provenance.

Review Questions

Short Answer Questions

1. What is the principal difference between descriptive and predictive analytics?
2. Distinguish a population parameter from a sample statistic.
3. Why can a large sample still produce biased results?
4. What is data leakage, and why is it dangerous in predictive modelling?
5. Name the six phases of CRISP-DM.

Long Answer Questions

1. Explain the four forms of analytics and show how they could be combined in a retail demand-management project.
2. Discuss how sampling frames, measurement scales and missing data affect the validity of statistical analysis.
3. Explain why predictive accuracy is not sufficient for selecting a model for operational deployment.
4. Describe each phase of CRISP-DM and explain why the lifecycle is iterative rather than strictly linear.
5. Discuss the ethical, privacy and governance issues that should be considered when deploying a customer-churn model.

Analytical / Application Questions

1. A bank wants to reduce loan defaults. Frame a precise analytical question by specifying the decision, unit of analysis, outcome, prediction time, population and success criteria.
2. A survey reports that 72% of 250 sampled customers are satisfied. Identify the statistic, the likely population parameter and two conditions that should be considered before generalising the result.
3. A classifier identifies 80 customers as likely to default. Of these, 50 actually default. Among all customers who actually defaulted, 100 were observed. Calculate precision and recall, and explain which measure would matter more if missing a high-risk customer were especially costly.

Multiple Choice Questions with Answers

Question 1

Which form of analytics primarily answers the question “What happened?”

A. Descriptive analytics

B. Inferential analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: A. Descriptive analytics

Question 2

A numerical characteristic of an entire population is called a:

A. Sample

B. Parameter

C. Statistic

D. Variable

Correct Answer: B. Parameter

Question 3

For the observations 2, 4 and 6, what is the sample mean?

A. 2

B. 3

C. 4

D. 6

Correct Answer: C. 4

Question 4

Which CRISP-DM phase focuses on defining the business objective and project success criteria?

A. Data preparation

B. Business understanding

C. Modelling

D. Deployment

Correct Answer: B. Business understanding

Question 5

Why is a time-based split often appropriate for forecasting data?

A. It guarantees perfect predictions

B. It prevents all missing values

C. It better represents predicting later periods from earlier information

D. It removes the need for model evaluation

Correct Answer: C. It better represents predicting later periods from earlier information

Question 6

A statistic is best described as:

A. A value calculated from sample data

B. The complete set of relevant units

C. A legal constraint on data use

D. A future business decision

Correct Answer: A. A value calculated from sample data

Question 7

Which issue occurs when a model uses information that would only become available after the prediction time?

A. Stratification

B. Data leakage

C. Calibration

D. Randomisation

Correct Answer: B. Data leakage

Question 8

If 50 of 80 predicted positive cases are actually positive, precision is:

A. 0.40

B. 0.50

C. 0.625

D. 0.80

Correct Answer: C. 0.625

Question 9

Which statement about prediction and causation is correct?

A. A strong predictor always causes the outcome

B. Prediction and causation are identical

C. A variable may improve prediction without being a causal driver

D. Causal claims require no research design

Correct Answer: C. A variable may improve prediction without being a causal driver

Question 10

Which is a key responsibility after deploying a predictive model?

A. Stop measuring performance

B. Monitor data quality, performance and business impact

C. Delete the training documentation

D. Assume relationships never change

Correct Answer: B. Monitor data quality, performance and business impact