Foundations of Predictive Analytics

Introduction

Organisations generate data whenever customers make purchases, employees complete tasks, machines operate, websites receive visitors, or financial transactions are processed. Data becomes valuable when it supports better decisions. Predictive analytics is the systematic use of historical and current data, statistical methods, and computational models to estimate what is likely to happen in the future or for previously unseen cases.

Predictive analytics is not a crystal ball. It produces estimates under uncertainty. A useful prediction may be expressed as a probability, a category, a numerical forecast, or a ranking. Its value depends not only on model accuracy but also on whether the organisation can act on the result, whether the data are trustworthy, and whether the decision process is fair and lawful.

This chapter establishes the conceptual foundation for later study of regression, classification, forecasting, machine learning, model evaluation, and data-driven decision making.

Learning objectives

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

• distinguish descriptive, diagnostic, predictive, and prescriptive analytics;
• explain the difference between prediction and causation;
• identify common business applications of predictive analytics;
• use foundational terminology accurately;
• identify the roles of important analytics stakeholders;
• explain how analytics contributes to digital transformation;
• recognise limitations, ethical risks, and governance requirements; and
• evaluate a predictive analytics initiative from a business perspective.

The analytics continuum

Analytics can be understood as a progression from understanding what has happened to deciding what should happen. The four commonly used categories are related, but they answer different questions.

Descriptive analytics

Descriptive analytics organises and summarises observed data. Common techniques include totals, averages, rates, frequency distributions, dashboards, and time-series charts. A retailer may calculate average transaction value by store, while a university may report enrolment by programme and semester.

Descriptive analysis is necessary but not sufficient for decision making. A dashboard can reveal that customer complaints increased, but it does not automatically identify the cause or recommend a response. Care is also required when selecting summaries. An average can conceal major differences between customer groups, and a percentage can be misleading if its denominator is not stated.

Diagnostic analytics

Diagnostic analytics investigates possible explanations for an observed outcome. It may involve segmentation, drill-down analysis, comparisons across time or groups, correlation analysis, or investigation of process events. For example, an organisation may discover that delivery delays are associated with a particular supplier and a specific distribution centre.

Diagnostic findings should be interpreted cautiously. An association can suggest where to investigate, but it does not by itself establish that one factor caused another. Diagnostic analytics often generates hypotheses that require controlled testing or deeper domain investigation.

Predictive analytics

Predictive analytics estimates an unknown or future outcome from available information. Examples include:

• forecasting product demand;
• estimating the probability that a customer will cancel a subscription;
• identifying transactions that may be fraudulent;
• predicting machine failure;
• estimating credit default risk; and
• ranking job applicants or sales leads for further review.

A predictive model learns a relationship between input variables and an outcome using historical examples. The outcome may be numerical, such as revenue, or categorical, such as “will churn” and “will not churn.” Predictive performance must be evaluated on data that were not used to fit the model, because performance on training data may be overly optimistic.

Prescriptive analytics

Prescriptive analytics links predictions to possible actions. It may use business rules, simulation, optimisation, or decision analysis. If a model predicts demand, a prescriptive system might recommend order quantities while considering budget, storage capacity, delivery times, and service-level targets.

Prescriptive analytics does not eliminate managerial judgement. Recommendations depend on the objectives, constraints, and assumptions supplied to the system. A mathematically optimal action may be unacceptable if it violates law, fairness requirements, safety standards, or organisational values.

Fundamental terminology

A clear vocabulary is essential when communicating with technical and non-technical stakeholders.

• Observation: One recorded case, such as one customer or one transaction.
• Feature or predictor: An input variable used by a model.
• Target, outcome, or response: The variable the model attempts to predict.
• Label: The known target value used in supervised learning.
• Record: A structured representation of one observation.
• Dataset: A collection of records and variables.
• Training data: Data used to estimate model parameters.
• Validation data: Data used to compare models or tune settings during development.
• Test data: Data reserved for final, relatively unbiased evaluation.
• Model: A structured representation of a relationship used for prediction or explanation.
• Algorithm: A procedure for processing data or fitting a model.
• Parameter: A value learned from data by a model.
• Hyperparameter: A setting selected before or during model training, such as tree depth.
• Prediction: A model-generated estimate for a case.
• Probability: A numerical expression of estimated likelihood, ranging from 0 to 1.
• Forecast: A prediction concerning a future time period.
• Scoring: Applying a trained model to new cases.
• Residual or error: The difference between an observed value and a predicted value.
• Drift: A change over time in data patterns, relationships, or outcome rates.
• Explainability: The extent to which a model’s operation or output can be understood.

In supervised learning, a model uses examples for which the outcome is known. In unsupervised learning, the outcome is not supplied; the objective may be to discover groups, reduce dimensions, or identify unusual observations. Predictive analytics commonly uses supervised learning, although unsupervised methods can support feature creation, segmentation, and anomaly detection.

Prediction is not causation

A prediction concerns whether information helps estimate an outcome. Causation concerns whether changing one factor would change an outcome, all else relevant being equal. These are different questions.

Suppose customers who receive more promotional emails purchase more frequently. A predictive model may use email frequency successfully to estimate purchase probability. However, this does not prove that sending more emails causes more purchases. Customers who are already highly interested may both open more emails and purchase more. Interest is a possible confounding factor.

A causal claim generally requires stronger evidence, such as randomised experimentation, a credible natural experiment, or a well-justified causal design. In a randomised controlled trial, eligible units are randomly assigned to treatment and control groups. Randomisation helps balance observed and unobserved characteristics on average, allowing the effect of the intervention to be estimated under appropriate assumptions.

Predictive models can still be valuable without causal interpretation. A bank may predict default using account behaviour even if the variables are not causes of default. The model’s operational purpose is risk estimation. Problems arise when managers treat predictive associations as guaranteed effects—for example, assuming that changing a highly predictive variable will necessarily change the outcome.

A useful distinction is:

• Predictive question: Who is likely to leave?
• Causal question: What intervention would reduce leaving, and for whom?

The first supports targeting or prioritisation. The second requires evidence about intervention effects. A model designed for prediction should not automatically be used to select treatment without assessing causal and fairness implications.

Common business applications

Marketing and customer management

Organisations predict customer acquisition, response to campaigns, churn, lifetime value, and product recommendations. A churn model may prioritise customers for retention activity. Its practical value depends on whether the organisation can offer an effective intervention and whether contacting selected customers creates unwanted discrimination or poor experiences.

Finance and insurance

Financial institutions estimate credit risk, detect suspicious transactions, forecast cash flows, and assess claims. Because these uses affect access to financial services, documentation, validation, fairness monitoring, and regulatory compliance are especially important.

Operations and supply chains

Demand forecasting supports purchasing, production, staffing, and inventory allocation. Predictive maintenance estimates the likelihood of equipment failure from sensor readings, operating conditions, and maintenance history. Forecast uncertainty should be incorporated into safety stock and contingency planning rather than ignored.

Human resources

Analytics may forecast workforce demand, identify training needs, or estimate attrition. Employment applications are sensitive. Models should not reproduce historical discrimination or use inappropriate proxies for protected characteristics. Human review and transparent governance are essential.

Healthcare and public services

Predictive tools can support patient risk stratification, appointment management, resource allocation, and early intervention. Errors may have serious consequences, and predictive outputs should support—not replace—qualified professional judgement. Privacy, consent, security, and equitable performance across groups must be addressed.

A basic predictive modelling workflow

A professional analytics project is a decision process, not merely a modelling exercise.

1. Define the decision and outcome. Specify what action may follow the prediction, who will use it, and what success means.
2. Understand the data-generating process. Identify sources, collection methods, ownership, timing, missingness, and possible bias.
3. Prepare the data. Correct errors, manage missing values, define variables, combine sources carefully, and prevent information leakage.
4. Create development datasets. Separate training, validation, and test information appropriately. For time-dependent problems, later observations should generally not be used to predict earlier ones.
5. Establish a baseline. Compare the proposed model with a simple rule or existing process.
6. Fit and evaluate models. Select methods appropriate to the outcome and decision context. Use metrics that reflect the cost of errors.
7. Interpret and challenge results. Examine errors, subgroup performance, stability, and whether relationships make domain sense.
8. Deploy with controls. Define access, monitoring, escalation, documentation, and human review.
9. Monitor and improve. Track data drift, outcome drift, model performance, operational impact, and unintended consequences.

For a binary classifier, a confusion matrix records true positives, true negatives, false positives, and false negatives. Two common metrics are precision and recall:

Here, TP is the number of correctly identified positive cases, FP is the number of negative cases incorrectly classified as positive, and FN is the number of positive cases missed. Precision answers, “Of the cases predicted positive, how many were actually positive?” Recall answers, “Of the actual positive cases, how many were found?” Both formulas assume a defined positive class and a specified decision threshold. They are not interchangeable.

For numerical prediction, a commonly used measure is mean absolute error:

Here, n is the number of evaluated observations, yᵢ is the observed value, and ŷᵢ is the predicted value. MAE has the same units as the target and gives equal weight to the absolute size of each error. It is meaningful when the target is measured on a scale for which absolute differences are decision-relevant. No single metric is universally best.

Worked example: evaluating a retention model

A subscription company evaluates a model that flags customers likely to cancel. In a test sample, the model identifies 80 customers as high risk. Of these, 60 actually cancel. There are 100 customers who cancel in total.

The precision is:

The recall is:

Thus, three quarters of flagged customers cancel, but the model identifies only 60% of all cancelling customers. Whether this is acceptable depends on intervention cost, the value of retaining a customer, and the consequences of contacting customers who would not have cancelled. Management should not approve the model using accuracy alone, because a highly imbalanced customer population can make accuracy appear high even when a model detects few cancellations.

Stakeholders and responsibilities

Predictive analytics involves multiple stakeholders whose objectives may differ.

Accountability should remain with the organisation, not be transferred to an algorithm. Clear ownership is needed for data quality, model approval, decisions, complaints, and retirement.

Analytics and digital transformation

Digital transformation is broader than purchasing software or converting paper records into digital form. It involves redesigning processes, services, capabilities, and organisational behaviour around digital technologies and information.

Analytics contributes in several ways. First, it creates visibility through shared measures and near-real-time monitoring. Second, it supports process automation and early warning. Third, it enables personalisation and more responsive services. Fourth, it helps organisations experiment, learn, and allocate resources more effectively.

However, analytics can also expose weaknesses. Poorly defined processes produce inconsistent data; fragmented systems prevent a reliable customer view; and a culture that does not trust evidence may ignore useful models. Digital transformation therefore requires data governance, interoperability, skilled people, leadership, change management, and a willingness to redesign decisions. Artificial intelligence and machine learning can increase scale and complexity, but they do not remove the need for sound problem definition and responsible management.

Illustrative case: predictive maintenance in a logistics company

Background

A fictional logistics company operates 600 delivery vehicles. Unexpected breakdowns cause missed deliveries, emergency repair costs, and customer dissatisfaction. The company records mileage, engine alerts, maintenance history, vehicle age, route characteristics, and breakdown events.

Problem

The maintenance department follows a fixed schedule. This can replace parts that still have useful life, while some vehicles fail between scheduled services. The company wants to predict whether a vehicle is likely to experience a breakdown within the next 30 days.

Analysis

The analytics team defines one observation as a vehicle-day. The target is a binary label: breakdown within 30 days, yes or no. Features are constructed only from information available on the prediction date. This timing rule is essential. For example, a repair invoice created after a breakdown cannot be used as a predictor for that same breakdown, because doing so would create information leakage.

The team divides historical observations by time, using earlier periods for development and a later period for testing. It compares a simple mileage-based rule with a predictive model. Evaluation considers recall because missed failures are costly, but also precision because unnecessary inspections consume workshop capacity.

Approach and solution

Vehicles with high predicted risk are scheduled for inspection, while medium-risk vehicles receive additional monitoring. The model output is not treated as a guaranteed failure. Maintenance managers review risk scores alongside safety rules, driver reports, and current workshop capacity. The company records which intervention was applied and monitors outcomes after deployment.

Outcome

In this illustrative case, the pilot reduces unplanned breakdowns among vehicles in the monitored fleet. It also reveals that sensor readings are less reliable for one vehicle model, so the organisation improves sensor maintenance before expanding the system. The result demonstrates that predictive analytics is a combined data, process, and governance initiative rather than a standalone algorithm.

Learning points

• The target and prediction horizon must be clearly defined.
• Features must be available at the time of prediction.
• Evaluation metrics should reflect operational costs.
• Human expertise remains important in safety-related decisions.
• Monitoring is required because equipment, routes, and data quality change.

Limitations, ethics, and governance

Data quality and representativeness

Historical data may contain missing values, measurement errors, inconsistent definitions, and biased sampling. A model trained on one region or customer segment may perform poorly elsewhere. Data quality should be assessed in relation to the decision, not merely by counting blank cells.

Bias and fairness

A model can produce unequal error rates across groups even when protected attributes are excluded. Proxy variables, historical decisions, and unequal access to services can reproduce disadvantage. Fairness is context-dependent; possible considerations include equal opportunity, access, error distribution, and impact. These goals can conflict, so organisations must make and document normative choices rather than claiming that a model is simply “objective.”

Privacy and security

Personal data should be collected and used for legitimate, clearly communicated purposes. Organisations should apply data minimisation, access control, retention limits, security safeguards, and appropriate consent or legal bases. Sensitive predictions can themselves create harm even when the input data appear ordinary.

Explainability and contestability

People affected by high-impact decisions may need understandable reasons, routes for appeal, and access to human review. A technically accurate model is not automatically acceptable if decision makers cannot identify when it fails or affected people cannot challenge an outcome.

Overfitting and drift

Overfitting occurs when a model captures random peculiarities in development data rather than general patterns. Testing on unseen data helps assess generalisation, but future conditions may still differ. Concept drift occurs when the relationship between predictors and outcomes changes; data drift occurs when the predictor distribution changes. Monitoring and periodic reassessment are therefore part of deployment, not optional extras.

Automation bias

Users may accept a model’s output uncritically, especially when it appears precise. Training should explain uncertainty, appropriate use, known failure conditions, and escalation procedures. A score should inform a decision, not conceal responsibility for it.

Best practices for professionals

1. Start with a consequential decision, not with an attractive dataset.
2. Define the target, time horizon, population, and action before modelling.
3. Establish a simple baseline and compare the model with current practice.
4. Use time-aware and leakage-resistant validation.
5. Select metrics based on error costs and stakeholder needs.
6. Examine performance across relevant groups and operating conditions.
7. Document data sources, assumptions, transformations, model versions, and limitations.
8. Separate development, approval, deployment, and monitoring responsibilities where feasible.
9. Provide human oversight for high-impact or safety-critical decisions.
10. Plan for retraining, rollback, complaints, and model retirement.

Contemporary developments

Cloud platforms and scalable data architectures have made predictive tools more accessible, while automated machine-learning systems can speed up model comparison. Generative AI can assist with coding, documentation, and interaction with data, but generated outputs require verification and should not be treated as evidence without checking. Responsible AI practice increasingly emphasises transparency, robustness, privacy, security, fairness, human oversight, and accountability. Organisations should align their controls with applicable privacy, consumer-protection, employment, sectoral, and AI regulation rather than assuming that a general-purpose model is suitable for every use.

Chapter Summary

Descriptive analytics summarises what happened, diagnostic analytics investigates why it happened, predictive analytics estimates what is likely to happen, and prescriptive analytics recommends actions under stated objectives and constraints. Predictive analytics supports marketing, finance, operations, human resources, healthcare, and public services, but its value depends on decision relevance and implementation quality.

Prediction is not causation. A variable can improve prediction without being a cause, so interventions require causal evidence or appropriate experimentation. A professional workflow defines the decision and target, prepares and validates data, compares models using relevant metrics, deploys with governance, and monitors performance over time. Stakeholders include business sponsors, analysts, engineers, subject experts, users, legal teams, auditors, and affected individuals.

Digital transformation uses analytics to redesign processes and improve responsiveness, but it also requires leadership, skills, integrated data, and responsible governance. Bias, privacy risks, leakage, overfitting, drift, weak explanations, and automation bias must be actively managed. The central professional principle is that analytics should improve accountable human decision making rather than replace it without scrutiny.

Key Terms

Predictive analytics: Use of data and models to estimate unknown or future outcomes.

Descriptive analytics: Analysis that summarises observed events or conditions.

Diagnostic analytics: Analysis that investigates possible reasons for observed outcomes.

Prescriptive analytics: Analysis that recommends actions subject to objectives and constraints.

Feature: An input variable used by a predictive model.

Target: The outcome a model attempts to predict.

Training data: Data used to fit or estimate a model.

Validation data: Data used during development to compare or tune models.

Test data: Data reserved for final performance assessment.

Prediction: An estimate produced for a particular case or time period.

Forecast: A prediction about a future time period.

Model: A representation of relationships used to generate predictions or support analysis.

Algorithm: A defined procedure for processing data or fitting a model.

Overfitting: Fitting random or sample-specific patterns that do not generalise.

Data leakage: Use of information unavailable at prediction time or improperly shared between development and evaluation.

Drift: Change in data patterns, outcome rates, or predictor-outcome relationships over time.

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

Recall: The proportion of actual positive cases that are correctly identified.

Causation: A relationship in which changing one factor produces a change in an outcome under specified conditions.

Data governance: Policies, roles, processes, and controls for responsible data management.

Review Questions

Short Answer Questions

1. What question does predictive analytics answer, and how does it differ from descriptive analytics?
2. Define a feature, target, training dataset, and test dataset.
3. Why does correlation or predictive usefulness not establish causation?
4. Distinguish precision from recall.
5. Name three stakeholder groups involved in a predictive analytics project and state one concern of each.

Long Answer Questions

1. Explain the four categories of analytics and show how they can support one business decision from reporting through action.
2. Describe a complete predictive modelling workflow, including data preparation, evaluation, deployment, and monitoring.
3. Discuss how predictive analytics contributes to digital transformation and explain why technology alone is insufficient.
4. Analyse the ethical and governance issues that arise when predictive models are used in employment or financial decisions.
5. Explain the difference between prediction and causation using a business example, and describe how causal evidence could be obtained.

Analytical / Application Questions

1. A retailer’s churn model has high accuracy but identifies very few customers who actually leave. Which additional metrics and business information should managers examine before deployment?
2. A hospital wants to predict readmission using patient data collected after discharge. Identify a possible leakage problem and redesign the prediction setup so that it supports a valid decision.
3. A delivery company’s model performs well during testing but poorly six months after implementation. Develop a structured investigation covering data drift, concept drift, operational changes, and governance responses.

Multiple Choice Questions with Answers

Question 1

Which type of analytics primarily answers the question “What is likely to happen?”

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: C. Predictive analytics

Question 2

A model flags 80 customers as likely to cancel, and 60 of those customers actually cancel. What is the model’s precision?

A. 25%

B. 60%

C. 75%

D. 80%

Correct Answer: C. 75%

Question 3

Which statement best distinguishes prediction from causation?

A. Prediction always requires an experiment, while causation never does

B. A variable can improve prediction without being a cause of the outcome

C. Causation is concerned only with numerical outcomes

D. Prediction and causation are identical when a model is accurate

Correct Answer: B. A variable can improve prediction without being a cause of the outcome

Question 4

What is data leakage?

A. Deleting duplicate records from a dataset

B. Using information that would not be available at the time of prediction

C. Combining data from two authorised systems

D. Encrypting sensitive data during transfer

Correct Answer: B. Using information that would not be available at the time of prediction

Question 5

Which metric answers the question “Of the actual positive cases, how many were identified?”

A. Precision

B. Recall

C. Mean absolute error

D. Specificity only

Correct Answer: B. Recall

Question 6

Which dataset should generally be reserved for final, relatively unbiased evaluation?

A. Training data

B. Source data only

C. Test data

D. Synthetic data only

Correct Answer: C. Test data

Question 7

A change over time in the relationship between predictors and the target is commonly called:

A. Concept drift

B. Data encryption

C. Descriptive aggregation

D. Randomisation

Correct Answer: A. Concept drift

Question 8

Which is the most appropriate role of prescriptive analytics?

A. Summarising last year’s sales

B. Explaining why a complaint increased

C. Estimating next month’s demand only

D. Recommending an inventory action under capacity constraints

Correct Answer: D. Recommending an inventory action under capacity constraints

Question 9

Why can a highly accurate model still be unsuitable for detecting rare fraud cases?

A. Accuracy never applies to classification

B. A model can appear accurate while missing most positive cases in an imbalanced dataset

C. Fraud models cannot use historical data

D. Precision and recall are always equal

Correct Answer: B. A model can appear accurate while missing most positive cases in an imbalanced dataset

Question 10

Which principle is most important when using predictive analytics for high-impact decisions?

A. Remove all human review to improve consistency

B. Use the most complex model available regardless of context

C. Provide governance, monitoring, and appropriate human accountability

D. Optimise only for training-data performance

Correct Answer: C. Provide governance, monitoring, and appropriate human accountability