Each submission must use real data collected from your own workplace, professional practice, or a Nigerian/African organisation you have direct access to. Simulated or publicly downloaded datasets may only supplement primary data — they cannot replace it.

This assessment asks you to act as a practising data scientist tackling a real problem in your own organisation or sector. You will collect or extract genuine data, frame an analytical question, apply five techniques from the textbook, and communicate your findings through a reproducible Quarto document.

The defining feature of this assessment — and the source of the 30-mark practical component — is the use of real data that clearly maps to your business operations. These 30 marks are awarded for the quality, transparency, and operational relevance of your data, not for the analyses themselves. Each submission must include:

• A professional disclosure statement (Section 2 of your document) explaining your job role, your organisation's domain, and why the five techniques you chose are directly relevant to your day-to-day work.
• At least one primary dataset you collected yourself — from your organisation's systems, a survey you designed and administered, web scraping, or direct field observation. The dataset must contain a minimum of 100 observations and at least 5 variables.
• A data provenance section describing: how the data were collected, the sampling frame, the time period covered, ethical approvals or consent obtained, and any data-sharing restrictions that affect what you can publish.
• A technique justification paragraph for each of the five methods, explaining why it is the appropriate technique given your specific business context and data structure.

30-Mark Practical Component — Breakdown

Your submission is a single Quarto (.qmd) document rendered to HTML. It must:

• Open with a YAML header specifying title, author, date, format: html, toc: true, code-fold: true, self-contained: true.
• Use R and/or Python code chunks. A panel-tabset showing both languages is strongly encouraged and attracts higher marks within the 100-mark component.
• Include all data loading, cleaning, and analysis code — the document must be fully reproducible from top to bottom.
• Embed all visualisations inline; no external image files.
• Be published as a live HTML URL on RPubs (rpubs.com/publish from RStudio) or Posit Connect Cloud (connect.posit.cloud from Positron).
• Optionally: push the .qmd source and data files to a public GitHub repository and include the repo URL in your submission — this attracts up to +5 bonus marks.

A minimal YAML header:

---
title: "[Your Case Study Title]"
author: "[Your Full Name]"
date: today
format:
  html:
    theme: flatly
    toc: true
    code-fold: true
    self-contained: true
---

Section A — Real-Data & Business-Operations Component (30 marks)

See Section 1.1 breakdown above.

Section B — Submitted Analytical Work (80 marks)

Section C — Defence / Viva Voce (20 marks) — held approximately one week after submission

1.1   Overview

This case study focuses on the first and most important phase of any analysis: understanding what you have. Before fitting models, a rigorous analyst spends significant time on exploratory data analysis, visualisation, and formal statistical testing. You will apply these foundational techniques to data from your own professional context and demonstrate that you can move fluently between exploratory insight and inferential conclusion.

1.2   Required Techniques

1.3   Business Context Examples (illustrative only — use your own)

• Bank credit officer: Collect anonymised loan application data from your team's portfolio. EDA the distribution of loan amounts and repayment status. Test whether interest rates differ significantly across loan categories. Correlate credit scores with default rates. Regress repayment probability on applicant characteristics.
• HR manager: Survey your team (or extract from your HRIS) on attendance, performance scores, tenure, and training hours. EDA distributions. Test whether performance differs by department. Correlate training hours with ratings. Regress performance on tenure and training investment.
• FMCG sales representative: Record weekly sales across your territory by product and outlet type. EDA sales distributions and seasonal patterns. Test whether sales differ by outlet type. Correlate promotional spend with uplift. Regress sales on price and promotion.

1.4   Data Requirements

• Minimum 100 observations; 6 variables (at least 3 numeric, 2 categorical, 1 date or time variable).
• Variables must include at least one outcome/dependent variable and several predictors that make substantive business sense.
• If your primary data has fewer than 100 rows, supplement with a second collection period or related source — but document this clearly in Section 3.

1.5   Specific Deliverables

• An EDA section that identifies at least 2 data quality issues and explains how you handled them (missing values, outliers, skewness).
• A visualisation narrative: at least 5 plots in a cohesive layout that tells a single story about your dataset.
• Hypothesis testing: formulate at least 2 hypotheses, state H₀ and H₁, check assumptions, run the test, report p-value and effect size, and interpret in plain business language.
• A correlation matrix with heatmap; discuss the 2–3 strongest correlations and their business implications.
• A regression model with diagnostic plots; interpret each significant coefficient as a concrete business action.

1.6   Guiding Questions

• What does the distribution of your key outcome variable tell you about the business process that generated it?
• Which visualisation type best communicates the most important pattern in your data, and why did you choose it over alternatives?
• What would a statistically significant result in your hypothesis test mean for a decision your organisation faces right now?
• Which correlation in your data is most plausibly causal, and how would you design a test to confirm or refute that causality?
• How would you translate your regression coefficient into a recommendation for a non-technical manager?

2.1   Overview

Machine learning transforms data into decisions. In this case study you will move beyond description and inference to build predictive and segmentation models on your own data. You will demonstrate that you understand not just how to run a model, but how to evaluate it, explain it, and connect its output to a concrete business action. You will also show that unsupervised techniques can reveal structure that supervised methods assume away.

2.2   Required Techniques

2.3   Business Context Examples (illustrative only)

• Retail bank: Collect customer transaction data. Build a classifier to predict whether a customer will default or churn in the next 90 days. Explain the model to a credit committee with SHAP. Cluster customers into risk/value segments. Use PCA to visualise the segment landscape. Forecast monthly transaction volumes with ARIMA.
• Supply chain manager: Collect SKU-level demand and delivery data. Classify deliveries as on-time or late. Explain which supplier variables drive delays. Cluster SKUs by demand pattern. Reduce the feature space with PCA for visualisation. Forecast demand for the top-10 SKUs for the next quarter.
• Hospital administrator: Collect patient admission records. Predict 30-day readmission. Explain the model with SHAP for clinical staff. Cluster patients by risk profile. Use PCA on clinical indicators. Forecast weekly bed demand.

2.4   Data Requirements

• Classification: minimum 200 observations, a binary or multi-class outcome variable, and at least 6 predictor variables.
• Time series: minimum 24 time periods (weekly or monthly). If your data is daily, aggregate to weekly for this exercise.
• Clustering: the same dataset as classification is acceptable — segment observations independently of the outcome variable.

2.5   Specific Deliverables

• A classification pipeline: train/test split (or cross-validation), at least two model types compared, ROC curve, confusion matrix, and a deployment recommendation (which model and why).
• A SHAP summary plot and waterfall plot for one representative prediction; explain each top-5 feature in plain language.
• A cluster analysis with optimal-k justification (elbow + silhouette), cluster profile table, and a naming/labelling exercise for each cluster.
• A PCA biplot showing where clusters sit in the reduced feature space.
• A time series decomposition plot, ACF/PACF analysis, and a 3-period forecast with prediction intervals.

2.6   Guiding Questions

• Which model architecture performed best, and does the performance difference justify the added complexity?
• If presenting to a non-technical board, which SHAP output would you show and how would you explain it?
• What do your clusters reveal about heterogeneity that aggregate statistics would hide?
• How would you use cluster membership as a feature in your classification model (target encoding, dummy variables)?
• Is your time series stationary? What transformation — if any — was required and why does stationarity matter for ARIMA?

3.1   Overview

The frontier of business analytics extends well beyond structured tables of numbers. Organisations generate text — customer complaints, employee surveys, board minutes, social media mentions. Decisions involve risk and uncertainty that can be quantified through simulation. Operational systems need demand forecasts and optimal allocation of constrained resources. In this case study you apply five such advanced methods to your own organisational data.

3.2   Required Techniques

3.3   Business Context Examples (illustrative only)

• Marketing manager: Collect customer reviews or NPS open-text responses. Analyse sentiment and extract topics with LDA. Simulate the uncertainty in your campaign ROI with Monte Carlo. Forecast next quarter's revenue with Prophet. Compute CLV for your customer base. Apply market basket analysis to identify cross-sell opportunities.
• Operations / logistics manager: Collect delivery records, incident logs, and demand data. Analyse the text of incident reports. Simulate procurement cost uncertainty. Forecast demand hierarchically. Model driver retention with survival analysis. Solve a vehicle routing or inventory-allocation LP.
• HR / talent manager: Collect exit interview text, performance reviews, and headcount data. Sentiment-analyse exit text. Simulate annual attrition under different retention-intervention cost scenarios. Forecast headcount demand. Model attrition with Cox proportional-hazards. Optimise training-budget allocation with LP.

3.4   Data Requirements

• Text data: minimum 50 text documents (survey responses, complaint tickets, emails, meeting notes, social media posts).
• Simulation: identify at least 3 uncertain inputs relevant to a real financial decision. Fit probability distributions to historical data — do not assume normality without testing.
• Forecasting: minimum 36 time periods for Prophet; minimum 52 weekly periods for LightGBM features. Justify your forecast horizon.
• Customer/People analytics: minimum 100 individuals with at least one event variable (churn date, attrition date, last purchase).
• Optimisation: formulate the LP or association-rule problem from actual operational constraints — your warehouse capacities, budget limits, or product catalogue.

3.5   Specific Deliverables

• A text preprocessing pipeline (lowercasing, stop words, stemming/lemmatisation) and a TF-IDF visualisation showing the 20 most distinctive terms.
• A sentiment trend chart (over time or by category) with business interpretation.
• An LDA topic model with coherence-based K selection; name each topic and explain what organisational behaviour it reflects.
• A Monte Carlo simulation with 10,000 runs: histogram of outcomes, P10/P50/P90 table, and a tornado sensitivity chart.
• A forecast with walk-forward cross-validation RMSE and a 3–6 month forward projection with confidence bands.
• A survival curve (Kaplan-Meier) or CLV table segmented by a meaningful business dimension (product, channel, or department).
• Either: association-rules output (top rules by lift with business interpretation) or an LP solution with shadow prices interpreted as managerial insight.

3.6   Guiding Questions

• What does the sentiment trend reveal about the trajectory of customer or employee experience in your organisation?
• Which topics emerged from LDA that were not visible in quantitative surveys — and what does that suggest about your measurement instruments?
• In your Monte Carlo simulation, which uncertain input has the greatest impact on the outcome? How should that drive your risk management priorities?
• What is the probability that the project you simulated generates a negative return? At what input value does it break even?
• Which customer segment has the highest CLV but also the highest churn risk — and what retention action does your analytics support?

4.1   Submission Instructions

• Render your Quarto document to HTML with self-contained: true.
• Publish to RPubs: in RStudio click Publish → RPubs. Copy the resulting URL.
• Or publish to Posit Connect Cloud from Positron: quarto publish connect.
• Submit the live URL (not a file) via the LMS by the stated deadline.
• For the GitHub bonus: create a public repo, commit your .qmd and data files (anonymise if needed), and include the repo URL in your submission.
• Each student must submit independently. Shared datasets, identical code, or near-identical documents will be flagged as collusion.

4.2   Defence / Viva Voce (Section C — 20 marks)

Approximately one week after the submission deadline, each student will attend a short individual defence session with Prof Bongo Adi or a designated examiner. The session lasts approximately 10–15 minutes.

• Bring your published HTML document open in a browser — no additional slides required.
• Be prepared to: walk through your data collection process, explain why you chose each technique, interpret any output the examiner points to, and discuss what you would do differently.
• The defence schedule will be communicated via the LMS within 48 hours of the submission deadline.
• Failure to attend without a documented medical or compassionate reason will result in zero marks for Section C (20 marks).
• The defence also serves as verification of independent work. Students who cannot credibly explain their own submission may have their overall grade reviewed.

4.3   Data Privacy & Ethics

If your data contains personally identifiable information (PII) — names, employee IDs, customer account numbers — you must anonymise before submission. Replace identifiers with codes (Customer_001, Employee_A, etc.). Do not publish raw financial data that your organisation treats as confidential. If in doubt, obtain written permission from your organisation before submitting and include a copy in an appendix.

4.4   Academic Integrity & AI Usage

You may use AI coding assistants (GitHub Copilot, Claude, ChatGPT) to help write code, but the analytical decisions — which technique, which model, how to interpret the output, what to recommend — must be yours. Include a brief AI usage statement at the end of your document (one paragraph) describing what you used AI for and where you made independent judgements. Presenting AI-generated interpretation as your own without disclosure constitutes academic misconduct.

4.5   Useful Resources

All submitted work must cite sources in APA 7th edition format. The minimum required citations are listed below. Add further references as your analysis demands.

Course Textbook (required citation in every submission)

Software & Package Citations

R and Python packages must be cited. Use the commands below to retrieve the correct citation for each package you use, then format in APA style.

How to Cite Your Data Source

Every dataset you use must be cited. Use the appropriate template below: