Choose a dataset and a question worth answering

Everything until now has used Bisi's data, because it was built to teach you specific things: the duplicates were placed, the dates were broken four ways, the price rise was waiting in April.

Real data isn't arranged for your benefit. Nobody has checked it's answerable. Nobody has confirmed a finding exists.

This module is your capstone: one analysis of your own, start to finish, good enough to show an employer or a client. It begins with the decision that determines everything after it — what to analyse, and what to ask of it.

Get this choice wrong and no amount of skill rescues it. A perfect analysis of a boring question is a boring analysis.

Words in this lesson

What makes a good capstone question

Four tests. A question that fails any one of them will waste your time.

That third test is the one people miss. "Is our revenue growing?" is weak, because you already know it is. A question worth asking is one where you'd genuinely accept either answer — which is exactly why "did the price rise cost us sales?" was a better question than "did revenue grow?"

Sharpening a question

Use the five-part frame from lesson 1.3 — decision, metric, scope, period, comparison:

Notice the sharp versions all name a comparison. A single number rarely says anything; a number next to another number usually does.

Where to find data

The best source is your own life

Your workplace. Your family's business. Your church or mosque's records. Your cooperative. Your own bank statements or airtime spending.

This is genuinely the strongest option for a portfolio, for three reasons:

• You understand the context, which is where analysis actually goes wrong
• You can ask questions of whoever collected it
• Nobody else has analysed it, so your work is unique

The catch is Module 8. If it contains personal data, you must handle it properly, anonymise before sharing, and get permission from whoever owns it before publishing anything. Get that in writing, even informally.

Public sources worth knowing

(Checked September 2026.)

Two cautions on public data:

Check the licence. Kaggle datasets carry different licences; some restrict commercial use or require attribution. Read it before building anything you intend to publish.

Beware the over-analysed dataset. If a Kaggle dataset has 900 existing notebooks, your analysis adds nothing to your portfolio. A small dataset nobody has touched is worth more than a famous one everybody has.

Judging a dataset before you commit

Spend thirty minutes here and you may save a week.

That last one sinks more projects than anything else. A column called status with values 1, 2, 4, 9 and no documentation is unanalysable, and no cleverness fixes it.

Run the Module 2 diagnosis on any candidate before committing: row count, uniques per column, blanks, date range. Thirty minutes tells you whether the next twenty hours are possible.

Scope it small enough to finish

The commonest capstone failure is ambition. "Analyse Nigeria's economy" produces nothing; "compare food inflation with headline inflation over three years" produces a finished piece.

A finished small analysis beats an abandoned large one, every time, in a portfolio and in a job.

Write your scope in three sentences before you start:

"I'm analysing my uncle's poultry business, January to December 2025, from his sales notebook and feed receipts. I'll answer which months lose money and why. I'm not covering labour costs, because he doesn't record them."

The third sentence matters most. Naming what you're not covering is what separates a scoped project from an unfinished one, and it's the coverage statement from lesson 4.4 written in advance.

What if there's no finding?

Sometimes you do the work and everything looks normal. That's a real result, and it happens to professionals constantly.

Three things to do before concluding there's nothing:

1. Cut the data a different way. By time, by group, by size band. Bisi's branch totals said one thing; per trading day said another.
2. Check what you assumed. The Surulere finding came from noticing a zero, not from a calculation.
3. Compare like with like. Unfair comparisons hide real findings as often as they invent fake ones.

And if there's still nothing: say so. "I expected X, tested it three ways, and found no evidence for it" is a legitimate, employable piece of work. It demonstrates method, and method is what you're actually selling.

What you must not do is keep slicing until something looks interesting. That's the noise-mining from lesson 4.5, and a finding produced that way won't survive a question.

Decide what you'll produce

Before starting, name your deliverables. For this capstone:

Knowing the destination changes what you do on day one — you'll keep the change log as you go, instead of reconstructing it from memory at the end.

Common mistakes

The short version

• A good question passes four tests: someone cares, the data exists, it could come out either way, it fits your time.
• Sharpen it with the lesson 1.3 frame, and make sure it contains a comparison.
• Your own life is the best data source — you have the context and nobody else has analysed it. Handle personal data per Module 8, and get permission in writing.
• Public sources: NBS (nigerianstat.gov.ng), World Bank Open Data, Kaggle. Check the licence, and avoid datasets with hundreds of existing analyses.
• Judge a dataset in thirty minutes — rows, period, blanks, granularity, join keys, documentation — before committing twenty hours.
• Scope it in three sentences, including what you're not covering.
• "No finding" is a legitimate result. Say it honestly rather than slicing until something appears.

Try it yourself

1. List five questions you could answer with data you can actually get. Score each against the four tests.
2. For your top two, write the five-part sharpened version from lesson 1.3.
3. Find the data. Run the Module 2 diagnosis on each candidate: rows, date range, blanks per column, uniques.
4. Check the licence and permissions. Write down what you're allowed to publish.
5. Write your three-sentence scope, including what you're not covering.
6. Write down what you expect to find, before you look. Keep it — you'll compare at the end.
7. List your deliverables and the deadline for each.
8. Show your question to someone who'd care about the answer. If they shrug, pick another.

You'll know you've got it when you can state your question, your scope and what you're not covering in under a minute — and someone else can tell you why the answer would matter.