Most business and data roles do not expect you to be a statistician, but they do expect you to reason clearly about numbers: what an average actually tells you, when a sample might be misleading, how confident you can really be in a conclusion. This deck focuses on the statistical concepts that come up in business contexts specifically, not abstract math theory, so the framing stays close to how you would actually use it in an interview or on the job. No statistics background required to start.
16 cardsConsultingFree, no signup
Card 1 of 16
You're advising Ferrow Retail Group on store performance. The regional director's slide says “average weekly revenue across our 40 stores is $84K” — but one flagship store does 5x the volume of any other location. What's the risk in leaning on that number, and how do mean, median, and mode differ?
Mean is the average, median is the middle, mode is the most common — skewed data favors the median.
The regional director's mean gets pulled upward by that one outlier store, making the “typical” store look stronger than it actually is. Mean adds up every value and divides by the count, median is the middle value once everything's sorted, and mode is whichever value shows up most often. Because the mean gets dragged around by outliers, the median holds steady and gives a better read on what a typical store is actually doing — the same logic applies to skewed data like income or home prices.
Card 2 of 16
Calder Foods' operating margin went from 4% to 5%. The associate wrote “margins up 1%” on the page and your engagement manager circled it. What's wrong with that, and what's a percentage point change?
Going from 4% to 5% is one percentage point but a 25% increase, so say which one you mean.
When the figure you're describing is itself a percentage, there are two honest ways to describe a move and they give very different numbers. Subtract them and you get the percentage point change, usually written pp: 5 minus 4 is one percentage point. Divide the change by where you started and you get the percentage change: 1 divided by 4 is a 25% increase. Writing “margins up 1%” blurs the two, and a client reading it may think almost nothing happened when the business is actually keeping a quarter more profit on every dollar of sales.
Card 3 of 16
What does standard deviation tell you?
Standard deviation measures how tightly values cluster around the average — small means consistent, large means volatile.
Standard deviation tells you how far values typically stray from the mean. A small standard deviation means the data clusters tightly around the average — consistent, predictable — while a large one means values swing widely in either direction. In a case, pairing it with the mean tells you not just what a client's typical store, deal, or customer looks like, but how reliable that “typical” figure actually is.
Card 4 of 16
A client tells you the average call-centre wait is 4 minutes, so service is clearly fine. You pull the raw data and the 90th percentile is 19 minutes. What's a percentile, and why does that number change the story?
A percentile says what share of cases fall below a point, and the tail is where customers get angry.
Sort every call from shortest wait to longest, and the 90th percentile is the value that 90% of callers come in under: nine in ten wait less than 19 minutes, one in ten waits longer. The median you already know is simply the 50th percentile, the middle of the queue. A 4-minute average can sit on top of a long tail of callers stuck for twenty minutes, and those are the people who hang up and complain, which is why service targets, salary bands, and delivery promises are usually written as medians and percentiles rather than averages.
Card 5 of 16
Northbay Logistics hands your team four million delivery records and a two-day deadline. A partner asks whether you really need all of them. What is a sample, and why can a few thousand records tell you anything about four million?
Random sampling works because every record had an equal chance of being picked, not because it is huge.
A sample is a subset of the full group, called the population, that you measure in place of measuring everything. It works when selection is random, meaning every record had an equal chance of being drawn, because a random draw tends to reproduce the population's mix of fast and slow deliveries in miniature. What matters most is how the sample was chosen rather than what fraction of the population it covers, so a clean 2,000-record sample beats a carelessly chosen 400,000-record one. Accuracy does improve as the sample grows, but slowly, which is why doubling the pull rarely doubles your confidence.
Card 6 of 16
A client's product team surveys only its loyalty-program members about a new subscription tier and reports 90% approval. Your case team should flag this before it goes in the client's board deck — why?
Sample bias means your data doesn't represent the population you're trying to understand, so conclusions don't generalize.
That 90% is sample bias in action: loyalty-program members aren't representative of the broader customer base the client is trying to understand, so the result won't generalize. It's the classic version of surveying only your existing customers about a new idea — you'd completely miss the opinions of everyone who looked at what you offer and decided not to buy, which is often exactly the group a new subscription tier needs to win over.
Card 7 of 16
Ferrow Retail's franchising head studies the 40 franchisees still trading after ten years, finds most run a large-format store, and calls large format the winning formula. Your manager calls this survivorship bias. What did the study miss?
Studying only the survivors hides everyone who failed doing exactly the same thing.
Survivorship bias is what happens when you analyse only the cases that made it and never see the ones that dropped out along the way. Plenty of large-format franchisees may have opened and closed over those ten years, and if they failed at the same rate as the small ones, format explains nothing at all. The fix is to rebuild the original cohort, every franchisee that ever opened, and compare survivors against closures on the same measures. It's worth spotting early, because it quietly infects fund performance tables, customer surveys, and any best-practice study built only on today's winners.
Card 8 of 16
Ferrow Retail Group's CEO points out that stores with an in-store café also have higher electronics sales, and concludes the café is driving people to buy more electronics. What's wrong with that conclusion, and how would you push back?
Two variables moving together doesn't mean one causes the other — a hidden third factor might explain both.
The CEO is treating a correlation as if it were proven causation. Two variables moving together doesn't mean one causes the other — cafés and strong electronics sales might both simply show up in larger, higher-traffic stores, a hidden third factor driving both. Before recommending the client roll out cafés chain-wide, you'd want to isolate that store-size effect or run a controlled test, since proving causation usually takes an experiment or careful statistical work, not just an observed pattern.
Card 9 of 16
Ferrow Retail Group's e-commerce lead wants to know whether a redesigned checkout page increases conversion before rolling it out site-wide. How would you design a test to answer that fairly?
A/B testing randomly splits an audience into two groups to fairly compare which version performs better.
You'd run an A/B test: randomly assign incoming visitors to either the current checkout page or the redesign, then compare conversion rates between the two. Randomization is what makes the comparison fair — it ensures the two groups of visitors are alike in every other way, so any difference in conversion can be attributed to the redesign itself rather than, say, one group happening to skew toward more loyal customers.
Card 10 of 16
What is a p-value, in plain terms?
A p-value is the odds you'd see a result this extreme if nothing real were actually going on.
A p-value estimates the probability of getting a result at least as extreme as what you observed, assuming there's actually no real effect. The smaller the p-value, the more surprising your result would be in a world where nothing real is going on. It is not the probability that your finding is a fluke, which is the version interviewers like to probe. Below 0.05 is the common cutoff, which is why it's the usual bar for calling a client's test result “statistically significant” rather than noise.
Card 11 of 16
An A/B test across two million users shows a new banner lifting click-through from 3.00% to 3.04%, with a p-value of 0.001. The client's marketing lead wants to roll it out immediately. What would you tell her?
Statistical significance says an effect is probably real, not that it is big enough to be worth doing.
Statistical significance only means the difference is unlikely to be a fluke of sampling, and with two million users even a microscopic gap clears that bar comfortably. It says nothing about size. A lift of 0.04 percentage points is real but tiny, so whether it's worth anything depends on the revenue behind each extra click set against the cost of building and maintaining the change. The honest answer keeps the two questions apart, is the effect real and is the effect material, and lets only the second one drive the recommendation.
Card 12 of 16
What is a confidence interval?
A confidence interval is a range of plausible values, not a single guess — it expresses how much uncertainty surrounds an estimate.
A confidence interval gives a range of plausible values for something you don't know exactly — like a population average — built from your sample data, rather than one single point estimate. A 95% confidence interval means that if you repeated the same sampling process many times, about 95% of the intervals you'd construct would contain the true value. It's a way of expressing uncertainty around an estimate — useful shorthand when you need to tell a client how much to trust a number, not a guarantee about any one specific interval.
Card 13 of 16
Bramwell & Co. wants to know how much of NorthPeak Airlines' recent revenue uptick came from its new marketing campaign versus other factors, like added routes or seasonal demand. What tool would you reach for, and what would it tell the client?
Regression puts a number on how much an outcome changes when an input changes.
This is a regression problem: it estimates the relationship between an outcome (revenue) and the factors that might explain it (marketing spend, route count, seasonality), letting you put a number on how much revenue shifts as each input changes while holding the others constant. That's exactly the kind of question regression is built for in a business setting — forecasting sales, isolating what marketing spend or pricing actually does to revenue, and ranking which factors matter most.
Card 14 of 16
Calder Foods put its 20 worst-performing stores through a new training program. Six months later those stores improved more than the rest of the chain, and the client wants to roll training out everywhere. Why would you be sceptical?
Extreme results drift back toward normal on their own, making almost any intervention look like it worked.
Any store's result in a given period is part real performance and part luck, so the worst 20 are likely to be stores having a genuinely weak run plus a dose of bad luck on top. Bad luck doesn't repeat, so on average those stores would have improved next period with no training at all. That's regression to the mean, and it's why deliberately picking the extremes for an intervention almost guarantees a flattering result. The only clean read is a control group, a comparable set of equally weak stores left untrained, so you measure the training against what would have happened anyway.
Card 15 of 16
Northbay Logistics' new routing software has a worse on-time rate than the old one overall, yet it wins on short-haul routes and wins again on long-haul routes looked at separately. How is that possible, and which number goes in the deck?
An uneven mix of easy and hard cases can flip an aggregate, so check the segments before trusting a total.
This is Simpson's paradox: a pattern that holds inside every segment can reverse once the segments are pooled together. It happens when the groups differ in both size and difficulty. If the new software was mostly trialled on long-haul routes, where every carrier runs late more often, its overall rate gets dragged down by the harder mix rather than by worse performance. The segment numbers are the honest ones because they compare like with like, and you'd show the client both, explaining that the aggregate is really describing the route mix rather than the software.
Card 16 of 16
It's Thursday, the client readout is Monday, and your one real finding is that a segment's margin looks six points better — but it rests on nine months of data from a single region. How do you present it without overselling or burying it?
State the finding, then the evidence it rests on, then what would change your mind, in that order.
Lead with the finding in plain language and the number behind it, then state the basis in the same breath: nine months, one region, so treat it as a signal rather than a settled result. Say what would confirm it, another region or a full year of data, and be explicit about which decisions it can carry and which it can't. Clients rarely punish a caveat that arrives with a clear next step, and they never forgive a number that quietly falls apart in a board meeting, which is why senior people scrutinise how a finding is framed as hard as they scrutinise the analysis under it.
Reading is not the same as remembering. Study this deck with scheduled review so the cards come back before you forget them.