Analyst case rounds rarely test whether you can compute something. They test whether you'd notice that the number you were handed is misleading, and whether you'd say so to the person who is pleased with it.
That second half is the harder one, and it's most of the job.
Below is a case with the full numbers, worked to the three sentences you'd actually send. It takes about twenty minutes in a real interview.
The brief
Our average order value went from £42 to £48 last quarter — a 14% increase. The CEO wants to put it in the board pack. Can you confirm it?
"Can you confirm it" is the trap and the gift. The request assumes the answer is yes-or-no, and the entire case is about what a good analyst does with a question that has been framed too narrowly.
First: is the number right?
Yes. Check it and say so plainly, because a candidate who starts hedging immediately looks evasive.
| Last quarter | This quarter | |
|---|---|---|
| Orders | 12,000 | 12,040 |
| Revenue | £504,000 | £576,000 |
| Mean order value | £42.00 | £47.84 |
"The arithmetic is right — 47.84, so 14% up. Before I confirm it for a board pack I'd want to know one thing: whether the average is describing what people think it's describing. An average moves for two different reasons, and only one of them is good news."
That sentence buys you the next five minutes, and it's the whole case in miniature.
Second: look at the distribution
The single most valuable habit an analyst has is refusing to summarise a distribution with one number.
| Last quarter | This quarter | |
|---|---|---|
| Mean | £42.00 | £47.84 |
| Median | £31.00 | £28.00 |
"The mean is up 14% and the median is down 10%. Those move in opposite directions when the top of the distribution pulls away from the middle — so my working theory is that a small number of very large orders appeared, and the typical customer is spending less than they were."
Third: find what's pulling it
| This quarter | Orders | Revenue | Mean |
|---|---|---|---|
| Orders over £500 | 40 | £90,000 | £2,250 |
| Everything else | 12,000 | £486,000 | £40.50 |
| All orders | 12,040 | £576,000 | £47.84 |
There it is, and it's worse than "the increase is partly explained by large orders."
"Forty orders — 0.3% of the total — carry £90,000. Strip them out and the remaining 12,000 orders average £40.50, against £42.00 last quarter. So the ordinary business didn't grow by 14%. It shrank by about 3.5% on the mean and 10% on the median.
The headline is true and the story it implies is the opposite of what happened."
The move being tested is checking the median before reporting a mean. It costs one line of SQL and it is the difference between an analyst who produces numbers and one who can be trusted with them. Every case round of this type is built on a distribution that a single summary statistic misrepresents, because that is the failure that does real damage in real companies — nobody audits a number that went in the right direction.
Fourth: the question nobody asked
Before writing anything, one more check — and this is what separates a strong case from a correct one.
"Forty orders over £500 is unusual enough that I'd want to know whether it's forty customers or one. If it's one new wholesale account, three things follow. The increase isn't a trend, it's an account. It'll repeat only if they reorder, so the board pack shouldn't imply a run rate. And we now have a revenue concentration we didn't have last quarter, which is a risk worth naming rather than a win."
Assume it comes back as one account. The case is now about something entirely different from average order value, and you got there from a distribution check.
The three sentences
Analyst rounds almost always end with some version of "so what would you tell the CEO?" The answer is not the analysis. It's the compression — and compressing without softening is the skill.
"The 14% is correct but it's one new wholesale customer: 40 orders carrying £90,000. Excluding them, the typical order fell — median £31 to £28 — so the underlying business got slightly worse this quarter, not better. Happy to put the 14% in the pack if we say what's behind it, but on its own it would tell the board the opposite of what's happening."
Three sentences: what's true, what's underneath, what you recommend. No hedging, no apology, and an offer rather than a refusal.
Answering the question asked
“Confirmed — AOV is £47.84, up 13.9% quarter on quarter. Let me know if you need it broken down by channel.”
Accurate, fast, and it lets a misleading number into a board pack. The offer to break it down at the end is the tell: the analyst sensed there was more and made it someone else's job to ask for it.
Answering the question meant
“The number's right. It's also one account, and underneath it the typical order fell 10%.”
Same data, thirty seconds longer, completely different consequence. And it's phrased so the CEO isn't wrong — the number they liked is still true, it just needs a sentence attached.
The follow-ups
"Isn't the wholesale order still real revenue?" — checking whether you'll over-correct. "Completely, and I'm not suggesting we discount it. £90,000 is £90,000. The issue is only what it implies about next quarter. If we report it as a 14% AOV improvement, the board reads a trend. If we report it as a new wholesale account plus a slight decline in retail, they read it correctly — and they'll ask the much better question, which is whether we can get more accounts like that one."
"What if the CEO doesn't want the caveat?" — they're checking whether you'd fold. "Then I'd ask what the number is being used for. If it's celebrating a good quarter internally, I care much less. If it's a board pack that feeds a forecast, I'd want the sentence in, and I'd rather have that argument now than when someone asks in three months why the 14% didn't repeat."
"How would you stop this happening again?" — the systems answer. "Report median alongside mean for anything with a long tail, and put a rule on the dashboard that flags when they diverge by more than a set amount. That's cheap and it catches this class of problem without anyone having to be suspicious."
What the panel scored
| Moment | What it demonstrated |
|---|---|
| Confirming the arithmetic first | You don't hedge before you've checked |
| Pulling the median unprompted | You don't trust a single summary statistic |
| Isolating the 40 orders | You quantify rather than characterise |
| Asking whether it's one customer | You think about what generated the data |
| Compressing to three sentences | You can be read by someone who won't ask a follow-up |
| Not over-correcting on the wholesale revenue | You're balanced, not contrarian |
Practising this
The exercise that transfers is small: take any metric you've reported as an average and pull its median. Do it for five metrics. On at least one of them the two will tell different stories, and that's the case round rehearsed on real data.
Then practise the compression, because it's the part that isn't intuitive. Write your finding, then cut it to three sentences where the first one concedes that the original number is correct. Leading with agreement is what makes the correction land instead of starting an argument — the same reason it works when explaining anything technical to a non-technical audience.
For the rest of the analyst loop — the SQL screen, the dashboard question, the "what should we do" round — the question set is here: Data Analyst Interview Questions. And for a longer case where the twist is composition rather than distribution, the data science version is here: Data Scientist Interview Case Study.
Practical target: say the three-sentence version out loud and listen for whether you apologised. Most people insert "sorry" or "I might be wrong but" in front of a correction to a senior person, and it converts a finding into a suggestion. The analyst who says "the number's right, and here's what's underneath it" without a cushion is the one who gets asked next time.
