Customer churn (synthetic)
One row per customer: tenure, spend, support tickets, late payments, product count, NPS, contract type — and whether they churned. The classic supervised-learning business problem.
19 datasets you can actually practise on
Every lab in this course runs on a real file, not a toy example. Each dataset comes with three levels of task and an answer key computed from the file itself, so you can check your work.
Loan risk (synthetic)
Loan applications with income, debt-to-income, credit history, previous defaults — plus approval and default outcomes. Credit risk, selection bias, and threshold economics.
Messy sales data (synthetic)
A deliberately dirty sales extract: inconsistent region/product spellings, missing reps, "missing"/"n/a" strings, malformed dates. Your data-cleaning and AI-assisted-analysis playground.
Support tickets (synthetic)
Support tickets with segment, free-text message, priority, resolution hours, and escalation. Text classification, triage design, and service-automation economics.
Dunder Mifflin — retail orders (case study)
A full US retail order book (Superstore-style): 9,994 line items across Furniture, Office Supplies & Technology, with sales, profit, discount, segment, region, and a returns flag. The definitive "AI-assisted analytics" playground — big enough that eyeballing fails and AI + verification wins.
Airbnb Paris — pricing (case study)
11,241 Paris Airbnb listings with price per guest, bedrooms, beds, bathrooms, review score, instant-booking, cancellation policy, host verification and district. A real prediction + decision problem: what drives nightly price, and should an investor convert a long-term rental to short-term?
Latvia job vacancies — real labor market (case study)
Every open vacancy registered with the Latvian State Employment Agency (NVA) on 31 Oct 2025: job title, employer, address, and a free-text salary range. Real, messy, text-heavy — the perfect GenAI extraction + labor-market analytics case.
HR & compensation (104 employees)
A tidy HR table — department, position, years of experience, gross salary, and an ordinal satisfaction rating. Ideal first dataset for data types (nominal/ordinal/interval/ratio), descriptive stats, and pay-equity analysis.
EU AI Act classification exercise (20 systems)
20 realistic AI systems to classify by EU AI Act risk tier — with the correct tier and a one-line rationale in the answer columns. Classify first, cover the answer columns, then check.
Marketing A/B test (2,400 users)
A landing-page experiment: variant A vs B, user type, channel, conversion, revenue. Contains a deliberate Simpson's-paradox-style trap — the naive read overstates B's lift because of traffic mix.
RAG evaluation set (15 questions)
A ready-made golden set for testing a knowledge assistant: 10 answerable questions with expected sources, plus traps — out-of-corpus, inappropriate, synthesis, and a prompt-injection test.
Latvia real estate — price prediction (case study)
Real scraped Latvian apartment listings (city24.lv/ss.lv): city, district, area, rooms, floor, building series — and price. The canonical machine-learning regression problem: predict price from features.
Used cars — depreciation & price (case study)
10,000 real ss.com used-car listings: brand, model, year, mileage, engine, body, transmission — and price. Feature-rich regression where representation and non-linearity (depreciation curves) make it a great deep-learning-vs-classical comparison.
Latvia Top-500 companies — analytics (case study)
Real Latvian company financials (saraksts.lv): by industry, the top firms turnover, profit, assets, employees, and average net salary. A decision-intelligence playground — benchmarking, productivity, and "which industry to bet on".
Predictive maintenance (operations case study)
Machine sensor readings — temperature, speed, torque, tool wear, vibration, power — with a failure flag. The classic Industry-4.0 classification problem for the Operations lesson.
Market signals — finance case study
Daily equity data by ticker/sector: price, return, volume, market cap, P/E, 20-day realised volatility, and a rare "breakout" signal label. Finance-lesson material on anomaly detection and the danger of rare-event classification.
Pricing & promotion — marketing case study
Weekly store-SKU data: regular/net price, discount, competitor price, shelf position, display flag, promo type, and units sold. The marketing-lesson problem: what actually drives sales — price, promotion, or placement?
Insurance risk scoring — fairness case study
Telematics driver profiles: age cluster, gender, licence age, annual km, persona, claim history — and a base risk score. Small but pointed: the perfect fairness-and-ethics case for the Responsible-AI lesson.
Course completion & dropout — adoption case study
An LMS enrollment table (meta — an AI course much like this one): student, course, modality, final score, certificate, and a predicted dropout risk. For the operating-model / adoption lesson: what makes learning actually stick?