I cannot hear your hypothesis. I can only hear the next thing you ask for. In a first round of eight candidates, my estimate is that five or six will say "my hypothesis is" inside two minutes, and two at most will follow it with a data request the hypothesis made necessary. The rest say the sentence, then walk their framework left to right as if it had never happened.
That is the problem with "hypothesis-driven" as it is usually taught. Candidates treat the hypothesis as a thing you say. Interviewers treat it as a thing you do, and the doing is visible in one place: the order and specificity of your questions. A hypothesis that does not change your next question is decoration. One that does is worth points even when it is wrong.
What follows: how to state a hypothesis with a price attached (the one number that would kill it), ask for that number first, and drop it in two sentences when it says no.
The word the firms never use
The three firms most associated with hypothesis-driven casing get through their interview pages without the word.
McKinsey's interviewing page calls the case the "Problem-Solving Interview" and says it is there "to evaluate your analytical thinking and approach to solving complex problems." BCG's case interview preparation page asks you to "identify the most important factors," and its six "Do" items run from "listen actively" to "show your thinking." Bain's case interview page says the firm is "looking for how you think, structure problems, and build on ideas as the conversation evolves." I read all three on 2026-09-24. "Hypothesis" does not appear on the BCG or Bain page, and it is absent from McKinsey's description of the interview.
So where does it score? On the sheets I have scored on, there was no line labeled hypothesis. There were lines for prioritization, use of data, and judgment, and a good hypothesis scores on all three at once: it is why you chose that branch first, why you asked for that number, and evidence you know which drivers matter in this kind of business. Firms do not publish their sheets, so treat that as one partner's experience, not a rule.
The word comes from engagement work. Ethan Rasiel's The McKinsey Way (1999) describes forming an "initial hypothesis" before the analysis begins, with a team and weeks of data. You have a paragraph of prompt and thirty seconds, so the interview version is cheaper: a bet about where the answer lives, not the answer.
Three things candidates call a hypothesis
When I ask for a hypothesis, I get one of three things. Only one is a hypothesis.
| What the candidate says | What it is | Can one exhibit kill it? | How the sheet reads it |
|---|---|---|---|
| "The client should enter the market." | A recommendation, stated early | No. It takes the whole case | Ignored, or marked down if defended against the data |
| "This is a cost problem." | A branch label | Not really. "Cost" is a third of the tree | Neutral. Structure, not a bet |
| "The premium line lost volume to a new entrant. If units are flat and price is down, I'm wrong." | A driver, a direction, a falsifier | Yes. One exhibit on units and price | Prioritization, data use, and judgment at once |
The first is the answer-first trap: decisive-sounding and untestable in the room, so you either defend it against the exhibits or quietly drop it and hope nobody noticed. Interviewers notice. The second feels safe. "It's probably costs" is not wrong, but it is not a bet, and it tells me nothing about whether you understand the business.
The third is the only one that changes what you ask for next. It is also the only one that can be wrong, which is exactly why it scores. A hypothesis you cannot kill with one exhibit is a topic.
Price the bet: driver, direction, and the number that settles it
The format that works has three parts, said in one breath: the driver (which lever moved: volume, price, mix, unit cost, fixed cost, capacity, churn), the direction, and the number that settles it.
"I think the profit decline is mostly mix, not volume. I'd like units and average price separately; if units are down more than a couple of points, I'll move to distribution instead."
Twelve seconds, four jobs: what you believe, why you want this number, what would change your mind, where you go next. Most candidates spend those twelve seconds on "I'd like to look at revenues and costs."
| Case type | A priced bet | The number that settles it | If it comes back the other way |
|---|---|---|---|
| Profit decline | "Mix shifted toward lower-margin products" | Revenue and gross margin by product line | Flat mix: move to unit cost or volume |
| Market entry | "Distribution, not demand, is the constraint" | Channel access held by incumbents | Open channels: move to willingness to pay |
| Pricing | "We are underpriced for the value delivered" | Price and churn versus the two closest rivals | Already the highest price: move to cost to serve |
None of those bets is guaranteed right. Each tells the interviewer which number you want first and why, which is the part that scores.
Ned's rule. If your hypothesis did not change your next question, you did not have one. Name the driver, name the number that would kill it, and ask for that number before anything else.
Worked example: Kestrel Tools
Invented company, invented figures. The point is the sequence.
Prompt. Kestrel Tools makes cordless power tools sold through big-box retailers. Revenue fell 8 percent last year while the category grew 3 percent. Why?
Opening bet. "An 11-point gap against the market in a retail-sold category is usually lost shelf space, a price cut, or customers moving to cheaper products. My first bet is shelf space, because a big-box range review can take a brand out overnight. I'd like units and average selling price, separately. If units are only slightly down, I'll move to price and mix."
Exhibit 1. Units: 1,000 thousand to 980 thousand. Blended average selling price: $168 to $158.
"Units down 2 percent, price down 6. 0.98 times 0.94 is 0.92, which is the 8 percent. Two points are volume, six are price or mix. Shelf space would show up in units, so my first bet is mostly dead. Is the $10 drop a discount on the same products, or a shift toward cheaper ones? Price and units by product line."
Exhibit 2. Pro line: $240 to $238; 400 thousand units to 320 thousand. Home line: $120 to $119; 600 thousand units to 660 thousand.
"Like-for-like prices are down about 1 percent on both lines, so it isn't discounting. The mix moved: Pro went from 40 percent of units to about 33. Pro lost 80 thousand units, a fifth of its volume, while Home gained 60 thousand. This is a Pro-line problem wearing a revenue costume."
The decomposition:
| Effect | Calculation | Impact on revenue |
|---|---|---|
| Last year's revenue | 400k × $240 + 600k × $120 | $168.0M |
| Volume | 20k fewer units × $168 blended | −$3.4M |
| Mix | 980k units at the new mix vs the old mix, at old prices ($156.0M vs $164.6M) | −$8.6M |
| Price | New mix at new prices vs old prices ($154.7M vs $156.0M) | −$1.3M |
| This year's revenue | 320k × $238 + 660k × $119 | $154.7M |
Mix is about two-thirds of the decline. The candidate got there in two requests because each was the number that could kill the current bet. Walking the tree from "revenues and costs" reaches the same exhibit four or five requests later.
Third bet, if the interviewer lets you continue: "Pro users are leaving Kestrel or trading down to Home. Did a competitor launch in the Pro segment last year?" A hypothesis about the world, and the right moment for one, because both internal explanations are dead.
Now the other candidate, same exhibits. Opens with "my hypothesis is that Kestrel is losing share to competitors." After Exhibit 1: "this confirms we're losing share." After Exhibit 2: "competitors must be taking Pro customers." Not wrong. But the hypothesis never cost anything, so it never earned anything: neither exhibit changed a word of it, and no number was attached. The first candidate's bet died inside ninety seconds and scored higher for it.
Why smart people defend dead hypotheses
A sixty-year-old experiment describes the case room. In Wason's 1960 study, 29 subjects were shown 2, 4, 6 and asked to find the rule behind them (three numbers in increasing order) by proposing triples and hearing yes or no. Only 6 of the 29 announced the correct rule at their first attempt; 13 announced one wrong rule first, and 9 announced two or more. Wason's verdict on the repeat offenders: they "were unable, or unwilling to test their hypotheses." They kept proposing triples that fit their guess instead of ones that could break it.
That is the case-room failure in miniature: you believe it is a cost problem, you ask for cost data, costs are up a little, that "confirms" it, and you never ask for the price and volume split that held the real driver. The fix is mechanical: ask for the number that would prove you wrong, not the number that would prove you right. In the Kestrel case, "units separately from price" was designed to kill the shelf-space bet.
McKinsey's own writing says the same thing in gentler language. In a 2020 McKinsey Quarterly article, Charles Conn and Robert McLean describe problem solving as trial and error: "We form hypotheses, porpoise into the data, and then surface and refine (or throw out) our initial guess at the answer." Their way of challenging any answer that implies certainty is to ask "What would we have to believe for this to be true?" Ask it before you say a hypothesis aloud and you will find the number that settles it.
When the number says no, the pivot is two sentences: what died, then where you are going and why. "Units are only down two points, so shelf space is not the main story. I'm moving to price and mix, because that is where the other six points must be." Do not pretend you predicted it; the interviewer has notes. Do not apologize either; a dead hypothesis is a completed test. Pushback on a live hypothesis gets the same shape of answer: which number would settle this, and can we look at it. Hold with data, fold with data.
Where the bet goes in each format
Not every case gives you a soapbox. Where you say the hypothesis changes.
- Interviewer-led (McKinsey-style): the interviewer controls the sequence, so the hypothesis lives inside each question. "Before I look at this, I expect the decline to sit in the Pro line," said before the chart turns over, shows you knew where to look.
- Candidate-led (BCG- and Bain-style): it lives in the order of your branches and your first data request, right after the structure.
- Written or chatbot cases (BCG Casey-style): the one line you write before the analysis.
- Final-round partner cases: how you handle pushback. Hold or fold on numbers.
Practice this today
One assignment, twenty minutes.
- Take three case prompts you have not seen. For each, write one line: driver, direction, the number that settles it. Sixty seconds per prompt. Then write the pivot you would say if it comes back the other way.
- Open the five-minute first rep, three typed turns, no account needed. Write your bet on paper before your first request, then check whether the request matched the bet. If not, you have found the gap this article is about.
- Run a live voice case. The seven-score debrief scores Hypothesis separately from Structuring, the quickest way to see the gap between a good tree and a good bet. CoachNed's hypothesis drills and 64-case library are the repetition.
Everything is open for seven days, no card; then $120 for a recruiting season or $49 a month.
CoachNed is independent and not affiliated with McKinsey, BCG, or Bain.
Frequently asked questions
Do you have to state a hypothesis in a case interview?
No firm says so publicly, and none of the three pages cited above uses the word. A testable hypothesis is the fastest way to show prioritization, use of data, and reasoning, but a candidate who asks for the right numbers without ever saying "hypothesis" scores the same.
What is the difference between a hypothesis and a framework in a case interview?
The framework is the map of where the answer could be. The hypothesis is your bet about where it is, plus the number that would prove you wrong. "Revenue and costs" is a framework. "Mix shifted toward the cheaper line; units by product line will show it" is a hypothesis.
What if my hypothesis is wrong?
Then you learned something in one data request, which is the point. What costs points is defending it after the data has contradicted it, or stating one so vague that no data could.
Is hypothesis-driven the same as answer-first?
No. Answer-first is a communication rule: conclusion before supporting points, at the end of the case. A hypothesis is a prioritization rule for the start: which number to ask for first. Confusing the two produces a recommendation in minute two that you spend twenty minutes defending.
Sources
- Interviewing at McKinsey — McKinsey's public description of the Problem-Solving Interview. Checked 2026-09-24.
- Case Interview Preparation, BCG Careers — BCG's "What to expect" text and "Do" list. Checked 2026-09-24.
- Case interview preparation, Bain & Company — Bain's description of what it looks for. Checked 2026-09-24.
- Wason, P. C. (1960), "On the failure to eliminate hypotheses in a conceptual task," Quarterly Journal of Experimental Psychology, 12(3), 129–140 (PDF, University of Florida mirror) — the 2-4-6 experiment; 6 of 29 subjects found the rule at their first announcement. Checked 2026-09-24.
- Six problem-solving mindsets for very uncertain times, McKinsey Quarterly, September 15, 2020 — Conn and McLean on refining and throwing out hypotheses. Checked 2026-09-24.