Business Analytics — From Data to Defensible Action
Business Analytics — From Data to Defensible Action
Business analytics is not the production of dashboards or model scores. It is a disciplined chain:
This course asks one question at every step: what decision becomes better, for whom, and how will we know?
The running case: HarborMart
HarborMart is a fictional UK omnichannel grocer. Customers browse online, order for delivery or collection, redeem promotions and sometimes stop buying. Stores replenish perishable and non-perishable products, schedule pickers and drivers, and promise delivery windows.
| Business concern | Analytic object | Possible action |
|---|---|---|
| late deliveries | probability an accepted order misses its window | change capacity or promise rules |
| stockouts | demand distribution by product–store–day | change replenishment quantity |
| promotion waste | incremental contribution caused by an offer | target, redesign or stop the offer |
| customer loss | calibrated probability of inactivity | intervene only when treatment is worthwhile |
| queue congestion | waiting-time distribution by hour | alter staffing or appointment capacity |
All HarborMart records and amounts are synthetic. Real organisations and studies are explicitly labelled.
Learning outcomes
By the end, you should be able to:
- turn an ambiguous request into a decision, unit of analysis, target, horizon, constraints and loss function;
- audit data grain, timestamps, missingness, joins, denominators and lineage before calculation;
- describe distributions, cohorts, funnels and KPIs without hiding uncertainty or subgroup differences;
- validate regression, forecasting and classification models against decision-relevant baselines without leakage;
- convert calibrated predictions into thresholds, allocations, inventory, staffing or pricing decisions;
- distinguish association, prediction, causal effect and optimisation;
- evaluate fairness, privacy, security, human review and failure recovery across an analytic system;
- present a recommendation that another analyst can reproduce and a decision owner can contest.
These outcomes suit advanced undergraduates and taught postgraduates. Graduate extensions add formal decision theory, causal identification, robust optimisation and deployment critique.
Preparation
You need percentages, algebra, probability, confidence intervals and basic Python. Prior machine learning or optimisation is not required.
Ten-minute readiness check
- Sales rise from £100,000 to £120,000. Is the growth rate 20 percentage points?
- A churn model is 95% accurate when only 3% of customers churn. Is it useful?
- A forecast is made on Monday using a field updated on Wednesday. What failed?
- A promotion is associated with larger baskets. Does that prove the promotion caused the increase?
- A model predicts demand perfectly. Can inventory be ordered without knowing holding and shortage costs?
Answers: no, it is 20%; not necessarily—a no-churn rule is already 97% accurate; temporal leakage; no, targeting and selection may confound the comparison; no, prediction does not specify the action or objective.
Course route
| Module | Central question | Main deliverable |
|---|---|---|
| 0. Decision and data foundations | What decision, population and evidence are we analysing? | analysis contract and data audit |
| 1. Descriptive analytics | What happened, to whom and with what uncertainty? | diagnostic brief and decision display |
| 2. Predictive analytics | What can be forecast at the decision time? | validated score or forecast |
| 3. Prescriptive analytics | Which feasible action has the best expected consequence? | policy, allocation or schedule |
| 4. Deployment and governance | Will the decision system remain useful, safe and accountable? | monitoring and control plan |
| Capstone | Can the entire chain survive an audit? | executive decision memo plus technical appendix |
| Appendix and revision tools | Which formula, pitfall or source supports the decision? | revision map and evidence trail |
The analytics contract
Complete this before opening a modelling library:
| Field | HarborMart delivery example |
|---|---|
| decision owner | regional fulfilment manager |
| decision | accept, re-promise or reroute an order |
| decision time | immediately before checkout confirmation |
| unit | one proposed order–delivery window |
| target | delivered after the promised end time |
| horizon | same-day operational outcome |
| available information | basket, store, slot, distance and capacity known at checkout |
| objective | contribution after delivery and service-failure costs |
| constraints | picker hours, driver routes, food safety and customer commitments |
| baseline | existing capacity rule |
| review rule | human override for disruptions; log reason and outcome |
A model without this contract may be statistically correct and operationally unusable.
Four claims that must not be merged
| Claim | Defensible example | Evidence required |
|---|---|---|
| descriptive | “Late-delivery rate was 8.2% last month.” | defined denominator and complete period |
| predictive | “This order has a 0.31 late-delivery risk.” | time-valid out-of-sample calibration |
| causal | “Adding one picker reduces lateness by 2 points.” | credible counterfactual design |
| prescriptive | “Add a picker from 17:00 to 20:00.” | effects, costs, constraints and alternatives |
Prediction can rank risks without identifying what intervention will work. Causal evidence can estimate an effect without selecting a feasible schedule. Optimisation can return a precise answer to a badly specified objective.
A 12-week teaching route
| Week | Preparation | Workshop output |
|---|---|---|
| 1 | decision framing | one-page analytics contract |
| 2 | data grain, time and lineage | join and leakage audit |
| 3 | distributions and robust summaries | diagnostic data profile |
| 4 | KPIs, cohorts and funnels | metric tree with denominator checks |
| 5 | visual evidence and uncertainty | one decision display and critique |
| 6 | validation and baselines | time-respecting evaluation plan |
| 7 | regression and forecasting | benchmark comparison |
| 8 | classification and calibration | probability audit and threshold table |
| 9 | trees, ensembles and drift | model card and monitoring triggers |
| 10 | optimisation, inventory and queues | feasible policy with sensitivity analysis |
| 11 | experiments and causal targeting | experiment or identification design |
| 12 | deployment and capstone defence | decision memo, rollback rule and oral defence |
Suggested workload is 90 minutes of preparation, two hours of workshop activity and 90 minutes of follow-up per week. Every interactive cell has a static worked example, so students can participate without running code. Text alternatives accompany diagrams and colour is never the only carrier of meaning.
Assessment alignment
The following is a teaching template, not an institutional grading rule:
| Task | Suggested share | Evidence assessed |
|---|---|---|
| data and metric audit | 20% | grain, timestamps, joins, denominator, uncertainty |
| predictive notebook | 25% | baseline, split, calibration, subgroup error, reproducibility |
| prescriptive case | 25% | objective, constraints, sensitivity and implementation |
| integrated capstone | 30% | complete decision chain, communication and governance |
For postgraduate work, require an explicit identification argument, robustness analysis and comparison between statistical loss and decision regret.
Current evidence window
The reading map was checked on 1 August 2026. Three current boundaries matter:
- NIST’s voluntary AI Risk Management Framework remains a useful govern–map–measure–manage structure; NIST also notes that version 1.0 is being revised.
- The EU framework changed days before this course check: Regulation (EU) 2026/1744 amended the AI Act, including implementation timelines. Legal conclusions must therefore be dated and checked, not copied from an old slide.
- Current organisational evidence shows analytics as a system problem: Walmart’s fiscal-2026 10-K links AI, automation and supply-chain investment, while Airbnb’s Chronon feature platform emphasises lineage, temporal correctness and training–serving consistency.
The reading and evidence map separates standards, scholarly results, organisational cases and emerging work. It is selective teaching evidence, not a systematic review.
Study discipline
For every chapter, produce four lines:
- claim: what is being asserted;
- calculation: how the number was produced;
- assumption: what must hold;
- decision consequence: what changes if the claim is wrong.
Start with decision framing, or use the formula and decision map when revising.