Your First 90 Days in Data, and the Word Nobody Defined

Theo’s first real request arrived in week three, by chat, with no capital letters: “what share of our electricity is clean? need it for the minister’s speech”

(Theo is a composite, a first-job analyst invented for this article. His employer is the energy office of the Verdant Union, the fictional union of five countries whose data we use in Green on Paper. The data is the real practice data: invented, built to behave like real environmental statistics, and messy on purpose.)

It looked like the easiest question he’d been asked since the interview. The electricity table, one row per country, year and source, even had a column called Renewable, true or false. So he did the obvious thing.

SELECT ROUND(100 * SUM(Production_GWh) FILTER (WHERE Renewable)
             / SUM(Production_GWh), 1) AS clean_pct
FROM energy_production
WHERE Year = 2022;

42.8%. Five minutes, start to finish. He sent it with a friendly emoji and went to make a coffee, feeling, for the first time, like an analyst.

By the time he got back there was a reply from the press office. The draft speech already said “more than two-thirds of our electricity is clean”, and could he tell them which of the two numbers was wrong? Only one of them had been there three weeks.

Two numbers, one table

So Theo did what he should have done first: he asked what “clean” means. Not in the dictionary. Here. The press office meant low-carbon: everything except coal, which in this table means nuclear as well as wind, solar and hydro. Theo’s column meant renewable: wind, sun and water. Nuclear power is low-carbon, but it isn’t renewable, and in the Verdant Union it’s a big number.

SELECT Source,
       SUM(Production_GWh) AS gwh,
       ROUND(100 * SUM(Production_GWh)
             / SUM(SUM(Production_GWh)) OVER (), 1) AS pct
FROM energy_production
WHERE Year = 2022
GROUP BY Source
ORDER BY gwh DESC;

Nuclear made 26.3% of the Union’s electricity in 2022. Count it, and “clean” is 69.1%. Leave it out, and it’s 42.8%. One query gives both, and neither is a mistake:

SELECT ROUND(100 * SUM(Production_GWh) FILTER (
         WHERE Source IN ('Wind', 'Solar', 'Hydro'))
       / SUM(Production_GWh), 1) AS renewable_pct,
       ROUND(100 * SUM(Production_GWh) FILTER (
         WHERE Source IN ('Wind', 'Solar', 'Hydro', 'Nuclear'))
       / SUM(Production_GWh), 1) AS low_carbon_pct
FROM energy_production
WHERE Year = 2022;

Both answers are correct, and both are defensible. They just answer different questions, and the request didn’t say which. Then Theo split it by country, and found out why the press office cared so much.

Brisland makes most of the Union’s nuclear power. Call clean “renewable” and Brisland comes fourth of five, at 33.0%. Call it “low-carbon” and Brisland is second, at 92.2%. Coralis and Estmark have no nuclear at all, so for them the word makes no difference. A speech that praised Brisland’s clean electricity would be true under one definition and quietly false under the other.

Nobody had lied. Nobody had written a bad query. In 2022 the Renewable flag matches the Source column row for row (not every year’s does, which is a story for another day), so for this question the column meant exactly what it said; the speech said clean and meant low-carbon. Theo’s SQL was fine. The request was missing a definition, and the definition turned out to be the whole answer.

Which is the lesson of week three, and the one that should shape the whole of your first 90 days. The most useful thing a new analyst produces early isn’t a dashboard, and it isn’t a model.

The first thing you ship
is a definition.

Not a clever window function. A definition, written down, with the query beside it and the people who agreed it. Sometimes, as here, it turns one vague word into two precise ones and bans the vague one. It’s the least glamorous thing you’ll produce all year, and it’s what stops the next new starter sending a minister a number without saying what it counts.

So here’s the 90-day plan we’d hand Theo, with hindsight. Three months, three jobs, and one thing to ship at the end of each.

Days 1 to 30: map before you build

The first month is for learning three things, in this order: the business, the tables and the words.

• The business. What the organisation does, what goes wrong, and who relies on its numbers. Ask to sit with those people. You’ll learn more about what “clean” means from an hour with the press office than from a week with the schema.
• The tables. For each one you touch: its grain (what one row is), its key, and where it comes from. Write it down as you go. A data dictionary started on day five beats a perfect one started in month six.
• The words. Every term people use as if it were obvious: clean, active, revenue, customer. Ask two people what each one means. If you get two answers, you’ve found your first piece of real work.

By day 30, ship a short glossary: your team’s words, each with its definition, its query and the person who agreed it.

Days 31 to 60: prove one number

Month two is for small requests, done carefully. Before you open the data, send a short brief back to the person asking: the decision it’s for, the question as asked and as agreed, the definitions, the period and what done looks like. It takes minutes, and it would have saved Theo his afternoon.

And when your number disagrees with someone else’s (it will), don’t quietly pick one. Put the two definitions side by side, the way Theo did, and show exactly where they part. A number with its definition attached is worth more to a new team than a fast one, because it’s the first thing they can trust you with.

By day 60, ship one number that matches somebody else’s, or a side-by-side that shows exactly why it doesn’t.

Days 61 to 90: own one decision

By month three you know the words and you’ve earned a little trust. Spend it on one memo that someone acts on: the question, the answer in one sentence, one table or chart of evidence, what it doesn’t show, and a recommendation. One page. If it needs two, it was two questions.

The word that has to go

Here’s the part nobody warns you about. Theo’s definition didn’t make anyone’s number smaller. “More than two-thirds” is perfectly true, as long as it’s called low-carbon. What the definition took away was the word. “Clean” sounds like wind turbines and sunshine; “low-carbon” can make people picture a reactor. The speech can have the flattering word with the smaller number, or the bigger number with the accurate word. It can’t have the flattering word and the bigger number together any more, and somebody has to choose out loud.

That isn’t a technical problem. It’s a conversation, and Theo is three weeks into the job. How hard should he push for the accurate word when the person choosing writes speeches for a minister?