Skip to main content

CBE NEXUS · UNDERGRADUATE TEXTBOOK SERIES

Chapter 11: Business Intelligence, Analytics and Decision Support

Information Systems

Comprehensive undergraduate notes, analytical diagrams, applied case studies and university assessment.

Undergraduate level · Chapter 11 of 16 · From organisational data to evidence-informed decisions

Learning outcomes: Explain BI architecture; design dimensional data models; assess data quality and governance; distinguish analytical methods; communicate evidence through visualisation; and evaluate privacy, fairness and analytical model performance.

Business intelligence: meaning, scope and value

Business Intelligence (BI) comprises the processes, architectures, technologies and practices used to transform organisational data into information that supports decisions. It includes data integration, storage, modelling, reporting, dashboards, visual analytics, alerts and governance. BI connects operational activity with management questions about performance, customers, resources, risk and strategy.

Transaction systems record events such as orders, payments, registrations and inventory movements. BI organises these records to reveal patterns across time, products, locations and user groups. Operational systems are optimised for reliable transactions, whereas analytical systems are designed for querying, aggregation and comparison. The two environments must exchange information without compromising integrity or performance.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Data, information, knowledge and decisions

Data consists of recorded observations or values. Information is data organised and interpreted in context, while knowledge includes understanding that supports judgement and action. A value such as 14,500 becomes useful when identified as monthly customer orders, compared with a target and interpreted alongside seasonality, product mix and fulfilment capacity.

Decisions may be structured, semi-structured or unstructured. Structured decisions follow stable rules; semi-structured decisions combine procedures with judgement; and unstructured decisions involve novel strategic questions. BI can support all three, but the degree of human interpretation differs. A useful analytical question identifies the decision, population, time period, outcome and constraints.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

BI architecture and information flow

A BI architecture typically includes source systems, data ingestion, staging, transformation, storage, semantic modelling, analytics tools and user-facing reports. Sources may include ERP, CRM, HR, SCM, websites, sensors, spreadsheets and external datasets. Each source has its own structure, timing, quality and access constraints.

Data is extracted or captured, validated, transformed and loaded into analytical stores. A semantic layer defines business entities and measures consistently. Access controls should apply throughout the pipeline, not just to dashboards. Lineage should show where a metric came from, when it was refreshed and which transformations were applied.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

OLTP and OLAP

Online Transaction Processing (OLTP) supports day-to-day transactions. It commonly uses normalised structures, constraints and short transactions to preserve consistency. Examples include recording a payment, updating an enrolment or reserving warehouse stock. OLTP workloads need reliable writes, concurrency control and prompt transaction responses.

Online Analytical Processing (OLAP) supports complex queries, aggregation and multidimensional analysis. It may organise information by date, customer, product, region and channel. Analysts can drill down from annual revenue to monthly or product detail, or roll up detailed activity to a summary. Separating analytical workloads can protect operational performance.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Data warehousing concepts and architecture

A data warehouse is a managed repository integrating data from multiple sources for analysis and reporting. Traditional definitions emphasise subject orientation, integration, time variation and non-volatility. Modern warehouses may support incremental updates and streaming ingestion while retaining governed analytical structures.

Architecture can be centralised, hub-and-spoke or organised around governed data marts. A central warehouse promotes shared definitions; data marts provide focused views. Poorly governed marts reproduce silos, so definitions and sources should remain traceable to enterprise standards. Design should consider workload, latency, concurrency, retention, recovery, security, cost and growth.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Dimensional modelling: facts and dimensions

Dimensional modelling organises analytical data around measurable business events and descriptive dimensions. A fact table stores measurements such as sales amount, units, discount or processing duration, together with keys to dimensions. Dimensions describe context such as date, customer, product, location and sales channel.

The grain states exactly what one row represents. One row might represent one product line on one sales order. Grain must be defined before selecting measures and keys. Mixing daily totals with transaction lines can cause double counting. Dimensions may use surrogate keys and preserve historical changes where analysts need to know which attributes were valid when an event occurred.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Star schemas, snowflake schemas and data marts

A star schema places a central fact table around descriptive dimensions, making common queries straightforward. A snowflake schema further normalises dimensions into related tables. Star schemas are often easier for business users, while snowflake designs may reduce repeated descriptive data but increase joins and complexity.

A data mart is a subject-focused analytical store for a domain such as finance, sales, HR or student performance. Dependent marts derive from a shared warehouse; independent marts may load sources separately. Independent marts can be quick to establish but risk inconsistent definitions. Document grain, keys, measures, hierarchies and historical behaviour.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

ETL and ELT data integration

Extract, Transform and Load (ETL) extracts data, transforms it in a processing environment and loads it into a target store. Extract, Load and Transform (ELT) first loads source data and performs transformations in the target platform. Both patterns can work depending on platform capabilities, security, cost and workload.

Extraction may be full or incremental, using timestamps or change-data capture. Transformations standardise dates, currencies, units, identifiers and categories. Load strategies include append, merge, overwrite and historised loading. Pipelines should record batch identifiers, source watermarks, rejected rows and reconciliation totals; rerunning a batch should not create duplicates.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Data quality and profiling

Data quality describes whether data is fit for a particular purpose. Common dimensions include accuracy, completeness, validity, consistency, uniqueness, timeliness and integrity. A dataset may be complete but inaccurate, accurate but late, or valid according to a format while semantically incorrect.

Profiling examines distributions, patterns, missing values, duplicates, ranges, relationships and anomalies. A student dataset may contain valid-looking registration numbers but duplicate people, missing programme codes or impossible dates. Quality rules should identify the field or relationship, expected condition, severity, owner and remediation route.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Data governance, stewardship and metadata

Data governance defines decision rights, policies, standards and accountability for data throughout its lifecycle. It covers ownership, quality, access, classification, retention, sharing, lineage and acceptable use. Governance helps an organisation agree what important terms mean and who may approve changes to shared information.

Data owners are accountable for a domain; stewards coordinate definitions and quality; custodians manage technical storage and controls. Metadata describes definitions, structure, source, owner, sensitivity, lineage, refresh schedule and permitted use. A business glossary and technical catalogue help prevent reports from silently using different meanings.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Data lakes, lakehouses and modern platforms

A data lake stores varied formats, including tables, JSON, logs, images and text. It can preserve detailed source data for exploration and machine learning. Without catalogues, access controls, quality checks and lifecycle policies, however, a lake can become a data swamp in which users cannot identify trusted information.

A lakehouse combines flexible storage with management features such as transactions, schema enforcement, catalogues and query optimisation. Platform selection should reflect workloads, skill availability, integration, latency, security, portability and cost. Classify data, define retention, identify owners and implement access policies before expanding ingestion.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Business metrics and KPI design

A measure is a quantitative value; a metric is a defined way of assessing performance; and a Key Performance Indicator is selected because it signals progress towards an important objective. Organisations should not label every available number a KPI because excessive indicators weaken attention.

A KPI definition should include name, purpose, formula, unit, population, exclusions, period, source, owner, target and refresh frequency. On-time delivery might be eligible orders delivered by the promised date divided by all eligible delivered orders. Treatment of cancelled, rescheduled and partially delivered orders must be explicit.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Descriptive analytics and reporting

Descriptive analytics summarises what has happened or what is observed. It uses counts, totals, averages, distributions, frequency tables, trends and segmented reports. Routine management reports may compare actual performance with budgets, targets, previous periods or peer groups.

Aggregation can hide variation. An organisation-wide average response time may look acceptable even if one region experiences severe delays. Examine distributions, percentiles, outliers and meaningful segments. Reports should state period, refresh time, filters and definitions; avoid excessive precision and misleading scales.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Diagnostic analytics and root-cause investigation

Diagnostic analytics investigates why an observed outcome occurred. Techniques include drill-down, segmentation, variance analysis, correlation, cohort comparison, process analysis and root-cause methods such as five whys and fishbone diagrams. The objective is to test plausible explanations rather than select a convenient story.

A fall in sales may coincide with price changes, stock shortages, competitor entry or seasonality. Examine timing, affected segments, alternative explanations and data quality. Correlation alone does not prove causation. Root-cause analysis should distinguish symptoms from mechanisms and connect recommendations to evidence.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Predictive analytics and forecasting

Predictive analytics uses historical and current data to estimate future outcomes or unknown values. Methods include regression, classification, time-series forecasting, survival analysis and machine learning. Applications include demand forecasting, credit risk, churn prediction, equipment maintenance and staffing requirements.

Define the target, forecast horizon, frequency, decision context and cost of error. Training data must represent deployment conditions. Data leakage occurs when information unavailable at prediction time enters training. Evaluate against simple baselines, use time-aware validation for time series, examine segment errors and communicate uncertainty.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Prescriptive analytics, optimisation and simulation

Prescriptive analytics evaluates actions that may improve an objective under constraints. Optimisation models may allocate staff, schedule production, route vehicles or balance inventory. They require an objective function, decision variables, constraints and assumptions about the environment.

Simulation explores system behaviour under scenarios. Discrete-event simulation represents queues; Monte Carlo simulation uses repeated random sampling; system dynamics models feedback and accumulation. Recommendations should explain trade-offs and feasibility. An optimisation model may minimise cost while worsening service equity or resilience if those objectives are omitted.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Statistical foundations for BI

Statistics provides methods for describing data, estimating uncertainty and evaluating relationships. Descriptive statistics include mean, median, mode, range, variance, standard deviation and quantiles. The mean is sensitive to extreme values, while the median is more robust for skewed distributions.

Probability describes uncertainty. Confidence intervals quantify uncertainty in estimates, while hypothesis tests assess compatibility of evidence with a stated null model. A p-value is not the probability that a hypothesis is true, and statistical significance does not automatically imply practical importance. Report effect sizes and uncertainty.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Data visualisation and chart selection

Visualisation communicates patterns through position, length, area, colour and shape. Line charts suit ordered time trends; bars compare categories; scatter plots show relationships between numerical variables; histograms show distributions; and maps display spatial patterns when geography matters.

Choose charts based on the analytical task. Pie charts become difficult to compare with many categories; dual axes can imply misleading relationships; and 3D effects distort perception. Use clear labels, units, scales and accessible colours. Provide text alternatives or accessible tables where needed.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Dashboard design and information storytelling

A dashboard presents related measures for a particular audience and purpose. Operational dashboards emphasise current queues and exceptions; tactical dashboards track departmental performance; strategic dashboards show outcomes and trends against long-term goals.

Put decision-relevant indicators first, show target and actual values, provide trend context and allow appropriate drill-down. Data storytelling explains what happened, why it matters, what is known and unknown, and what decision is required. Avoid exaggeration and cherry-picking.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Self-service BI and data literacy

Self-service BI allows authorised users to explore governed data and create analyses without submitting every request to a central team. It improves responsiveness but can create inconsistent metrics, uncontrolled extracts and security risks if users work outside shared definitions.

A governed model provides certified datasets, reusable semantic models, approved workspaces, training and role-based access. Data literacy includes formulating questions, understanding provenance, interpreting visualisations, reasoning statistically and communicating limitations. Training should use realistic organisational examples.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Real-time analytics and streaming data

Real-time analytics processes data as events occur or shortly afterwards. Sources include payment events, web activity, sensors, logistics updates and service logs. Applications include fraud detection, equipment monitoring, live inventory visibility and operational alerts.

Streaming architectures use event producers, message brokers, processors, state stores and consumers. Event time and processing time may differ because messages are delayed or out of order. Define acceptable latency, duplicate handling, replay, retention and recovery. Not every decision requires real-time data.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Big data and distributed processing

Big data is commonly described through volume, velocity and variety, with veracity and value also important. Large datasets may exceed traditional single-machine capacity, while fast and varied data creates ingestion and governance challenges.

Distributed processing partitions data and runs tasks in parallel. It introduces network failures, data skew, coordination, consistency and debugging challenges. Sampling, indexing, partitioning, columnar storage, caching and query optimisation may solve performance problems without unnecessary complexity.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Machine learning lifecycle and evaluation

A machine learning project begins with a defined decision problem and target. Its lifecycle includes data understanding, preparation, feature design, model selection, training, validation, deployment, monitoring and retirement. Business, data, technical and governance stakeholders should agree what constitutes a useful model.

Classification measures include precision, recall, specificity, F1 and ROC-AUC; regression measures include MAE, MSE and RMSE. The appropriate metric depends on the cost of false positives, false negatives and prediction errors. Governance includes versioning, reproducibility, documentation, fairness assessment and drift monitoring.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Artificial intelligence and generative BI

AI can support BI through natural-language querying, anomaly detection, forecasting, summarisation, semantic search and narrative generation. Generative AI may help users explore information conversationally, but it can produce plausible statements unsupported by the data or metric definitions.

A trustworthy design grounds responses in approved datasets, enforces user permissions, exposes query logic and links supporting evidence. Validate calculations independently, test for data leakage, log relevant interactions and monitor hallucinations. Human oversight remains essential for consequential decisions.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Privacy, security and responsible analytics

Analytical repositories combine information from many systems and may create detailed profiles of people. Privacy controls should establish lawful purpose, transparency, minimisation, retention, correction and permitted sharing. Data collected for one purpose should not automatically be reused for unrelated analysis.

Security measures include encryption, authentication, least privilege, row- and column-level controls, masking, audit logs and protected exports. De-identification does not guarantee anonymity when datasets can be linked. Responsible analytics also assesses fairness, representativeness, explainability and the impact of automated decisions.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

BI implementation and organisational adoption

A BI programme should begin with prioritised business questions, sponsors, users, data domains and measurable outcomes. Assess existing reports, source systems, quality, skills, infrastructure and governance. A roadmap can deliver a small trusted analytical product first, then expand based on adoption and value.

Implementation includes requirements, architecture, source access, modelling, pipeline development, metric definitions, dashboard design, security, testing, training and support. Involve users early. Monitor active usage alongside decision outcomes, data quality and feedback. Publishing dashboards alone does not complete a BI programme.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

BI risks, costs and vendor evaluation

BI project risks include inconsistent definitions, inaccessible sources, poor quality, privacy exposure, vendor lock-in, uncontrolled compute costs, skills shortages and low adoption. Assess likelihood, impact and sensitivity, assign owners and define mitigation.

Evaluate connectivity, modelling, visualisation, governance, scalability, accessibility, security, interoperability, administration, licensing and exit options. Compare total cost including storage, compute, refresh, licences, integration, training, support and internal administration. Preserve data portability and documented interfaces.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

BI in finance, marketing, operations and education

Finance BI supports revenue and expense analysis, cash-flow forecasting, budget variance, profitability, receivables, payables and control monitoring. Analysts should distinguish cash, recognised revenue, commitments and forecasts because they answer different questions.

Marketing analytics examines acquisition, conversion, retention, lifetime value and experience; operations analytics monitors throughput, quality, capacity and inventory; education analytics may examine enrolment, attendance and course completion. Educational analytics should support learners rather than stigmatise them or replace professional judgement.

Applied example: consider a university, retailer or manufacturer using this capability to improve a specific decision. The team should map the process that generates the data, identify where errors or delays enter, agree the business definition and test the output against an independent source. Users should be able to drill from a summary to an authorised explanation of the underlying records. If the data is incomplete, the report should display that limitation instead of implying certainty. In a real organisation, the analyst should record the question, decision owner, population, period, data sources, metric definitions, assumptions and acceptable uncertainty. Validate source totals against trusted records and document transformations. Explain what evidence supports and what it cannot establish. A recommendation should identify an action, a measure of its effect and a review date. Governance must assign responsibility for quality, access, maintenance and user support.

Design and evaluation considerations: choose a method proportionate to the decision's importance, volume and time sensitivity. Define quality thresholds, refresh expectations, access rules, failure alerts and an owner for exceptions. Test normal cases, unusual values, missing information and changes in source structure. Review performance after deployment, because data definitions, user behaviour and operating conditions can change. Document lessons and update the model, glossary and training material when evidence justifies a revision.

Case study: Retail sales and inventory intelligence

A retailer operates stores and an online channel. Managers receive weekly spreadsheets with different product codes and inconsistent definitions of sales. Some reports count orders when placed, others when dispatched, and inventory figures do not consistently account for returns or reserved stock.

Define transaction grain and metric rules, establish product and location master data, and build a governed sales and inventory model. Reconcile sources and represent returns, cancellations, reservations and channel transfers explicitly. Dashboards can show sales, margin, stock cover, sell-through, returns and availability by product and location.

Measure fewer reconciliation disputes, improved inventory accuracy, reduced stockouts, lower excess stock and faster planning. Pilot with representative stores and review definitions as products, channels and promotions change.

For assessment, prepare a source-to-report architecture, stakeholder and responsibility matrix, data-quality rules, privacy review and a benefits scorecard. Identify assumptions that must be verified before implementation. Pilot the solution with representative users and records, compare outputs with trusted evidence and define the conditions under which the organisation should pause, revise or scale the intervention.

Case study: University student success analytics

A university wants to identify students who may need academic support. Data is held in admissions, registration, attendance, learning platforms, assessment and student support systems. Leaders propose a single risk score, but staff are concerned about privacy, bias and labelling students unfairly.

Define a supportive purpose and involve academic staff, students and privacy specialists. Use only justified data, assess representativeness and document limitations. Distinguish correlation from causes and avoid treating a score as a diagnosis or final decision. Provide human review and correction routes.

Evaluate whether interventions improve timely support and student outcomes without increasing stigma or surveillance. Restrict access, validate across cohorts, test group error rates and retire models that cannot demonstrate net benefit.

For assessment, prepare a source-to-report architecture, stakeholder and responsibility matrix, data-quality rules, privacy review and a benefits scorecard. Identify assumptions that must be verified before implementation. Pilot the solution with representative users and records, compare outputs with trusted evidence and define the conditions under which the organisation should pause, revise or scale the intervention.

Case study: Manufacturing predictive maintenance

A manufacturer experiences unexpected equipment downtime. Maintenance records are inconsistent, sensor readings arrive at different intervals and production teams disagree about whether maintenance is too frequent or too late.

Improve asset identifiers, maintenance event coding, sensor calibration and time synchronisation. Define the objective, such as reducing unplanned downtime while controlling maintenance cost. Establish a baseline before comparing predictive models, and use time-aware validation to avoid leakage.

Track downtime, mean time between failures, maintenance cost, alert precision and missed failures. Keep expert review and safe operating procedures in place. Reassess models when equipment, operating conditions or sensor behaviour changes.

For assessment, prepare a source-to-report architecture, stakeholder and responsibility matrix, data-quality rules, privacy review and a benefits scorecard. Identify assumptions that must be verified before implementation. Pilot the solution with representative users and records, compare outputs with trusted evidence and define the conditions under which the organisation should pause, revise or scale the intervention.

Case study: Executive performance dashboard

A growing organisation's executive team receives a large monthly report with hundreds of metrics. Departments report different revenue and customer counts. Leaders spend meetings debating figures instead of deciding how to respond.

Identify strategic objectives and accountable metric owners. Agree revenue recognition, customer identity, reporting periods and exclusions. Publish a glossary and certified semantic model, reconcile initial results to authoritative records and design a concise dashboard with targets, trends and drill-down.

Review whether meetings become more decision-oriented, actions have owners and follow-up measures improve. Retire redundant reports carefully and maintain an auditable record of metric changes and decisions.

For assessment, prepare a source-to-report architecture, stakeholder and responsibility matrix, data-quality rules, privacy review and a benefits scorecard. Identify assumptions that must be verified before implementation. Pilot the solution with representative users and records, compare outputs with trusted evidence and define the conditions under which the organisation should pause, revise or scale the intervention.

Business intelligence and analytics diagrams

BI information pipeline
ERP, CRM, HR, web and sensor sources
↓
Secure ingestion and staging
↓
Quality checks, transformation and integration
↓
Warehouse, lake or lakehouse
↓
Semantic models and governed metrics
↓
Dashboards, reports and decisions
Dimensional warehouse model
Date, customer and product dimensions
↓
Location, channel and promotion dimensions
↓
Central sales fact table
↓
Measures: units, revenue, discount and cost
Analytics maturity spectrum
Descriptive: what happened?
↓
Diagnostic: why did it happen?
↓
Predictive: what is likely to happen?
↓
Prescriptive: what action should be taken?
Machine learning lifecycle
Define decision problem and target
↓
Prepare representative data
↓
Train and validate against baseline
↓
Deploy with controls and human oversight
↓
Monitor drift, fairness and outcomes
↓
Retrain, revise or retire model

University examination and revision questions

Use clear definitions, show calculations and explain the limitations of analytical evidence. Model answers provide concise guidance for core concepts.

Question 1. Define Business Intelligence.

Model answer: The processes, architectures, technologies and practices that transform organisational data into information for decision-making.

Question 2. Differentiate OLTP and OLAP.

Model answer: OLTP supports routine transactional processing; OLAP supports analytical queries, aggregation and multidimensional exploration.

Question 3. What is a data warehouse?

Model answer: A managed repository integrating data from multiple sources for analysis and reporting.

Question 4. Define the grain of a fact table.

Model answer: The precise meaning of one row, such as one product line on one sales order.

Question 5. Differentiate a fact and a dimension.

Model answer: A fact records measurable business events; a dimension describes their context.

Question 6. What is a star schema?

Model answer: A dimensional model with a central fact table connected to descriptive dimension tables.

Question 7. Differentiate ETL and ELT.

Model answer: ETL transforms before loading; ELT loads first and transforms within the target platform.

Question 8. State six dimensions of data quality.

Model answer: Accuracy, completeness, validity, consistency, uniqueness, timeliness and integrity.

Question 9. Explain data lineage.

Model answer: The traceable history of data sources, transformations, movement and use.

Question 10. What is a semantic layer?

Model answer: A governed business-facing model defining reusable entities, relationships, calculations and metrics.

Question 11. Differentiate descriptive and diagnostic analytics.

Model answer: Descriptive analytics summarises what happened; diagnostic analytics investigates why.

Question 12. Differentiate predictive and prescriptive analytics.

Model answer: Predictive analytics estimates likely outcomes; prescriptive analytics evaluates actions under objectives and constraints.

Question 13. What is a KPI?

Model answer: A measure selected to indicate progress towards an important organisational objective.

Question 14. Why define grain before designing a fact table?

Model answer: It prevents mixed levels of detail, incorrect joins and double counting.

Question 15. What is a slowly changing dimension?

Model answer: A method for managing changes to dimension attributes while controlling historical retention.

Question 16. Define data governance.

Model answer: Policies, decision rights, standards and accountability governing data throughout its lifecycle.

Question 17. What is a data mart?

Model answer: A subject-focused analytical store serving a domain such as finance, sales or HR.

Question 18. Explain a confidence interval.

Model answer: A range produced by a statistical procedure with stated long-run coverage under its assumptions.

Question 19. Why does correlation not prove causation?

Model answer: Confounding, reverse causality, selection effects or coincidence may explain an association.

Question 20. State four chart-selection principles.

Model answer: Match the chart to the question, label units, show context and use honest scales.

Question 21. What is self-service BI?

Model answer: Governed access that lets authorised users explore and analyse approved data.

Question 22. Define data drift.

Model answer: A change over time in the distribution or characteristics of model input data.

Question 23. Differentiate data drift and concept drift.

Model answer: Data drift changes input distributions; concept drift changes the relationship between inputs and outcomes.

Question 24. What is streaming analytics?

Model answer: Analysis of events as they arrive or shortly afterwards.

Question 25. Name four privacy controls.

Model answer: Purpose limitation, minimisation, access control, encryption, retention limits, masking and audit logging.

Question 26. Why should BI metrics have owners?

Model answer: Owners maintain definitions, approve changes, resolve quality issues and align measures with business purpose.

Question 27. A dashboard reports 240 late orders out of 1,200. Calculate the late-order rate.

Model answer: 240 ÷ 1,200 × 100 = 20%.

Question 28. A model correctly identifies 72 of 90 actual positive cases. Calculate recall.

Model answer: Recall = true positives ÷ all actual positives = 72/90 = 80%.

Question 29. What is a data lakehouse?

Model answer: An architecture combining flexible lake storage with management and analytical capabilities associated with warehouses.

Question 30. Give four measures for BI programme success.

Model answer: Adoption, decision-cycle time, data-quality improvement, reconciliation reduction, business outcomes and user trust.

Official Education References

Use these official organisations for curriculum guidance, examinations and further learning information.