Case study and guesstimate questions, answered the way you’d say them.

First ask which decisions they make, how often, and on what device. Then put 3 to 5 key numbers against target and last period at the top left, a trend beneath them, a breakdown by store or product to show where any gap is, and a drill-through to detail, all on one uncluttered screen. Document each metric’s definition and refresh time, and test the draft with the manager.

Every number needs context, such as a target or last year, or the manager can’t tell whether to act.

Microsoft Learn: Tips for designing a great Power BI dashboard

Practise it with Prism

Start with the question and the decision it supports, then collect the data, clean it, explore it, analyse or model it, and share findings in a form people can act on. Data cleaning makes the data fit for that question: removing duplicates, fixing types and formats, standardising values such as city names, handling missing values and checking totals against the source. Do it carefully and write down each change, because every later step inherits its mistakes.

If they want a named framework, CRISP-DM’s six phases run from business understanding through data understanding, preparation, modelling and evaluation to deployment.

IBM SPSS Modeler CRISP-DM Guide

Practise it with Prism

First confirm the drop is real: check tracking, the data pipeline and the metric’s definition, because a broken event or late data can look exactly like a fall in demand. Then find when it started, split orders into sessions × conversion, and break both down by platform, app version, city, channel, new versus returning users and payment mode to see where it sits. Only then test causes, internal or external, against the data.

1. Is it real?   tracking, pipeline, definition, time zone
2. When?         sudden or gradual; which day; the same week last year
3. Where?        orders = sessions × conversion; by platform, version, city, channel
4. Why?          internal: release, outage, price change; external: festival, competitor sale
5. So what?      size the impact, fix or monitor, tell the people affected

Ask clarifying questions first, such as how the 12% was measured and against which baseline.

Dmitriev et al. (2017), A dirty dozen: twelve common metric interpretation pitfalls in online controlled experiments, KDD

Practise it with Prism

Start from the goal, then choose signals and metrics for it, as in Google’s HEART framework (Happiness, Engagement, Adoption, Retention, Task success). Here: adoption, the share of eligible users who use reorder; task success, the time from opening the app to a paid order; and repeat orders per user as the primary metric. Add guardrails such as basket value and cancellations, and compare against a randomised holdout, not last month.

Define “eligible” first: only users with a past order can reorder, so divide by them, not by all users.

Rodden, Hutchinson and Fu (2010), Measuring the user experience on a large scale (the HEART framework), CHIDmitriev et al. (2017), A dirty dozen: twelve common metric interpretation pitfalls in online controlled experiments, KDD

Practise it with Prism

Give the structure first, then label every number as an assumption: likely buyers × share using a tiffin service × meals a month × price per meal × 12, where likely buyers are students and working people living away from family. Use round numbers and say each one aloud. Then sanity-check from the supply side, such as how many kitchens that volume implies, and end with a range and the assumption that moves it most.

Every number is an assumption to state, not a fact:
People in Bengaluru                            1.4 crore
Students and workers living away from family   20%   ->  28 lakh
Share using a tiffin service                   10%   ->  2.8 lakh
Meals a month per user                         40
Price per meal                                 ₹80
Yearly market = 2.8 lakh × 40 × ₹80 × 12       ≈ ₹1,075 crore

Show the sensitivity: if 5% use a tiffin service instead of 10%, the market halves.

Practise it with Prism

Work top-down and call each number an assumption: people × share who are adults × share of adults using UPI × payments per user a day. Then check bottom-up from the other side: merchants accepting UPI × payments each receives a day, plus person-to-person transfers. If the two disagree badly, find the assumption that’s off rather than averaging them, and give a range.

Every number is an assumption to state, not a fact:
Top-down:   40 lakh people × 75% adults × 50% use UPI × 2 payments a day  ≈ 30 lakh
Bottom-up:  1 lakh merchants × 25 payments a day = 25 lakh, + 5 lakh person-to-person
Both land near 30 lakh a day; give a range, such as 20 to 40 lakh

Say which real data you’d check it against, such as national UPI volumes scaled by the city’s share of the population.

Practise it with Prism

Profit is revenue minus costs, so with revenue flat, look at costs and at the mix behind the revenue. Split costs into ingredients, staff, rent, delivery-platform commissions, discounts and refunds, by outlet and month, to see which line grew and where. Then check the mix: a shift to delivery orders that pay commission, or to discounted orders, can keep revenue flat while shrinking the margin.

Profit  = Revenue − Costs
Revenue = orders × average order value   (by channel: dine-in, delivery, takeaway)
Costs   = ingredients + staff + rent + commissions + discounts + refunds
Compare each line by outlet and month: which grew, where, and since when?

Size each driver’s share of the drop before recommending anything, so effort goes where most of the money went.

Practise it with Prism

Follow the customer’s path and the money: acquisition, visitors and cost per new customer by channel; conversion, sessions to orders with the funnel steps between; value, orders, average order value and GMV, then net sales after cancellations and returns; retention, repeat-purchase rate and customer lifetime value; and operations, delivery time and on-time rate. Then pick one north-star metric for the goal at hand and treat the rest as its drivers and guardrails.

GMV          = sessions × conversion rate × average order value
Net sales    = GMV − cancelled and returned orders
Repeat rate  = customers with 2 or more orders in the period ÷ all customers in the period

Define each metric before you quote it: GMV counts orders before cancellations and returns, so it can rise while net sales fall.

Practise it with Prism

Clarify the scope first, such as men’s barber shops, not beauty parlours. Then size demand and divide by what one shop can serve: men × visits a year, against chairs per shop × customers per chair a day × working days. Label every number as an assumption, use round figures, and finish with a range and the assumption that moves it most, here how often men visit.

Every number is an assumption to state, not a fact:
People in Delhi                                    2 crore
Men and boys                                       50%   ->  1 crore
Barber visits a year each (about monthly)          12    ->  12 crore visits
One shop: 2 chairs × 12 customers a day × 300 days       =   7,200 visits a year
Shops ≈ 12 crore ÷ 7,200 ≈ 16,700, so say 15,000 to 20,000

Sanity-check from the supply side: 16,700 shops for 2 crore people is about one for every 1,200 residents; does that match the streets you know?

Practise it with Prism

Start from the goal, such as profit within two years, and score candidate sites on demand (people in the catchment, their income, footfall and online orders already coming from there), competition (rivals nearby, and how much the store would take from your own branches), cost (rent, staff, fit-out) and access (roads, parking, visibility, delivery reach). Then forecast each site’s sales from your existing stores’ data, and check the model against your most recent openings before trusting it.

Name cannibalisation explicitly: a store that mostly takes sales from your own branch nearby adds little.

Practise it with Prism

The aim is to balance riders and drivers, so start with demand, requests by area and time, and supply, available drivers and pickup times. Then price elasticity: how many riders drop off as fares rise, and whether higher fares bring drivers to busy areas. Add guardrails: caps on the multiplier, fairness, trust, competitors’ fares and any regulatory limits. Test it with switchback experiments that alternate the rule over time periods, since riders and drivers in one market affect each other.

Say what you’d watch beyond revenue: completed rides, cancellations, wait times and repeat riders.

OpenStax, Principles of Economics 3e: 5.1 Price elasticity of demand and price elasticity of supplyBojinov, Simchi-Levi and Zhao, Design and analysis of switchback experiments (arXiv)

Practise it with Prism

Pick one specific problem you’ve seen, not a wish list, and say who it affects and how you know, such as reviews, support tickets or a drop in the funnel. Propose one change, then define success: a primary metric tied to that problem, plus guardrails such as cancellations and order value. Finish with how you’d test it, a randomised rollout to a share of users, and what result would make you drop the idea.

Keep the critique about the user’s problem rather than the team’s choices: you may be talking to the people who built it.

Rodden, Hutchinson and Fu (2010), Measuring the user experience on a large scale (the HEART framework), CHIKohavi, Henne and Sommerfield (2007), Practical guide to controlled experiments on the web, KDD

Practise it with Prism

Group them by what the business needs: users, such as monthly transacting users and how many new users make a second payment; usage, transactions and value per user, split into person-to-person and person-to-merchant; reliability, payment success rate and time to complete, by bank and app version; merchants, active merchants and their repeat use; and money, revenue per user from lines such as loans, insurance and merchant services. Then name one north-star metric and say why it’s the one.

Split every metric by bank, app version and city before drawing conclusions: an overall rate can hide one failing segment.

Practise it with Prism

Work from households: people in NCR ÷ household size × the share of households with a car × cars per car-owning household gives the private car fleet. Multiply by the share that run on diesel, and add taxis and company fleets separately if the interviewer wants them. Label every number as an assumption, give a range, and say how you’d check it, such as against vehicle registration data.

Every number is an assumption to state, not a fact:
People in Delhi NCR                        4.5 crore
Households (4 people each)                 1.1 crore
With a car                                 25%   ->  28 lakh households
Cars per car-owning household              1.2   ->  34 lakh cars
Share that run on diesel                   25%   ->  about 8 lakh diesel cars

Name what’s special about NCR as something to check rather than guess, such as how rules on older diesel vehicles shape the fleet.

Practise it with Prism

First define better: faster, cheaper or more reliable, measured by delivery time, cost per shipment, on-time rate, first-attempt success and the share of orders returned undelivered. Then break the journey into steps, from order to packing, line haul and last mile, and find where time and money go, by city, pin code and courier partner. Fix the biggest gap first, with levers such as stock placed nearer demand, better address data, delivery slots and checks on risky cash-on-delivery orders.

Say what you’d trade: promising next-day delivery everywhere can cost more than it earns, so target it where demand justifies it.

Practise it with Prism

Agree what value means first: spend, profit or expected future spend. A quick, explainable start is RFM: score each customer on recency, frequency and monetary value over a fixed window, such as 12 months, and look at the top segments. Then refine with margin instead of spend and estimate customer lifetime value, the present value of future cash flows; finally, check the segment predicts next period’s spend, not just last period’s.

SELECT customer_id,
       CURRENT_DATE - MAX(txn_date)          AS recency_days,
       COUNT(*)                              AS frequency,
       SUM(amount)                           AS monetary,        -- total spend
       NTILE(5) OVER (ORDER BY SUM(amount))  AS monetary_score   -- 5 = top fifth
FROM transactions
WHERE txn_date >= CURRENT_DATE - INTERVAL '12 months'
GROUP BY customer_id;

Explain why the window matters: someone who spent heavily two years ago and nothing since is a win-back target, not a high-value customer.

Fader, Hardie and Lee (2005), RFM and CLV: using iso-value curves for customer base analysisPostgreSQL 18 docs: Window Functions

Practise it with Prism

First define churn for the business. A cancelled subscription is clear, but where there’s no contract, as in most shops and apps, you choose an inactivity window, such as no order in 90 days, from how often customers normally buy. Then measure churn by signup cohort and segment to see where it concentrates, look for drivers such as falling usage, complaints, failed payments or price changes before people left, and test retention offers against a control group.

Choose the inactivity window from data: plot the gaps between each customer’s orders and pick a point that few active customers exceed.

Fader, Hardie and Lee (2005), RFM and CLV: using iso-value curves for customer base analysisKohavi, Henne and Sommerfield (2007), Practical guide to controlled experiments on the web, KDD

Practise it with Prism

Count from the supply side, which you can picture: departures a day × seats per flight × the share of seats filled, then add arrivals, roughly equal. Cross-check from demand, with the catchment’s population × flights a year per person. Label every number as an assumption, and say where you’d verify it, such as the Airports Authority of India’s monthly traffic reports, before analysing trends by month and route.

Every number is an assumption to state, not a fact:
Departures a day                           60
Seats per flight                           180
Seats filled                               80%   ->  60 × 180 × 0.8 ≈ 8,600 departing
Arrivals, about the same                         ->  about 17,000 passengers a day
Check: 17,000 × 365 ≈ 62 lakh a year; compare it with AAI's figures for the airport

Airports Authority of India: Traffic News

Practise it with Prism

Go top-down: people in Bengaluru × the share who order food online × orders per user a month ÷ 30 gives the city’s daily orders, then multiply by the app’s share. Check it from the supply side: restaurants on the app × orders each handles a day. Label every number as an assumption, and finish with a range and the assumption the answer is most sensitive to, usually how often people order.

Every number is an assumption to state, not a fact:
People in Bengaluru                        1.4 crore
Order food online                          30%   ->  42 lakh
Orders per user a month                    4     ->  1.68 crore a month ≈ 5.6 lakh a day
This app's share                           50%   ->  about 2.8 lakh orders a day
Supply check: 25,000 restaurants × 11 orders a day ≈ 2.75 lakh

Practise it with Prism

Start with data: searches that find no good match, category growth, and competitors’ prices and reviews show whether there’s demand and at what price. Then prepare a strong listing, with clear images, specifications and a price set against comparable products, and plan the launch with deals, sponsored placements or a sale event. Track the funnel from impressions to product-page views, add-to-cart and orders, then ratings, returns and repeat purchases, and adjust price and promotion from what you see.

Practise it with Prism

Bound it first. The floor is cost plus the margin the business needs; the ceiling is what customers will pay, judged from competitors’ prices and the product’s value to them. Within that range, price elasticity matters, the percentage change in demand for a percentage change in price, so estimate it from past price changes or, better, test a few price points across comparable items or regions. Then check the whole picture: discounts, returns and what the price signals about quality.

OpenStax, Principles of Economics 3e: 5.1 Price elasticity of demand and price elasticity of supply

Practise it with Prism

Go from customers to orders: Indians who shop online × the share who buy fashion on this app in a year × orders each makes a year ÷ 365. State each figure as an assumption and round it. Then say whether you mean an average day or a sale day, since sale events can multiply a normal day’s orders, and give a range.

Every number is an assumption to state, not a fact:
Indians who shop online                    25 crore
Buy fashion on this app in a year          15%   ->  3.75 crore customers
Orders each a year                         5     ->  18.75 crore a year
On an average day                          ÷ 365 ->  about 5 lakh orders

Practise it with Prism

Start with the decision, not the data: what would change if you knew the answer, and how precise does it need to be? Then size the gap: which records are missing, and whether the gaps are random or tied to something that could bias the result. Fill what you can from proxies or another source, state your assumptions, test how far the conclusion moves if they’re wrong, and recommend with a range and the data that would settle it.

Practise it with Prism

Estimate how many bats are in use, then divide by how long one lasts. Split players into casual ones who share a cheap bat and regular ones who own a bat, give each group a replacement time, then add schools and clubs, which buy in bulk. Label each figure as an assumption, give a range, and name the biggest swing factor, here how many casual players share a bat.

Every number is an assumption to state, not a fact:
Casual players     10 crore, 1 bat per 5 players, lasts 2 years   ->  1 crore a year
Regular players    1 crore, own bat, lasts 1 year                  ->  1 crore a year
Schools and clubs  5 lakh, 4 bats a year each                      ->  20 lakh a year
Total                                                              ->  about 2.2 crore bats a year

Practise it with Prism

Not proven: a before-and-after comparison mixes the redesign with everything else that changed, such as a festival, a sale, a marketing push, a different traffic mix or a tracking change. Check those, compare with the same weeks last year, and look at segments that didn’t get the redesign. The clean answer is a randomised test, because random assignment spreads other factors evenly across the groups.

If an experiment isn’t possible, say which comparison group you’d use and what assumption makes it fair.

OpenStax, Introductory Statistics 2e: 1.4 Experimental design and ethicsKohavi, Henne and Sommerfield (2007), Practical guide to controlled experiments on the web, KDD

Practise it with Prism

Separate the extra sales from those that would have happened anyway. For targeted offers, the best tool is a randomised holdout: keep a random group out of the offers and compare their spending with everyone else’s, during the sale and for weeks after. Without one, compare with a forecast or similar past periods, and count the dip as people wait for the sale or stock up during it, cannibalised full-price sales and the discount’s cost to margin.

Report incremental margin, not only incremental revenue: a sale can lift revenue and still lose money.

Kohavi, Henne and Sommerfield (2007), Practical guide to controlled experiments on the web, KDD

Practise it with Prism

Weigh what it should gain against what it costs. Gains: conversion, especially from first-time buyers and in categories where fit is a worry, and repeat purchases. Costs: the return rate, reverse-logistics cost, the share of returns that can be resold, and abuse, such as heavy returners or items worn once and sent back. Judge it on net revenue after returns and on margin, ideally comparing randomly chosen regions or categories with and without the policy, and add limits for abusers if needed.

Kohavi, Henne and Sommerfield (2007), Practical guide to controlled experiments on the web, KDD

Practise it with Prism