Pricing Data Scientist Interview Questions: Elasticity, Confounding, and Experiment Design

Pricing data scientist interview questions require two answers at once: how customers respond to price, and how you know that response is caused by price. A regression can fit historical sales well while recommending a harmful price change if the historical pricing policy reacted to demand.

Define the causal effect before optimizing the price. Name the population, product, exposure, and time horizon; distinguish revenue from contribution; and choose an experiment that reflects how customers encounter prices. The same elasticity number can imply different decisions for one-time purchases and subscription renewals.

Evidence boundary: Official engineering material and primary research support the causal principles below. No verified candidate report establishes this exact exercise or employer rubric. The prices, quantities, costs, and suggested answers are original practice examples. The related Bayesian A/B testing guide covers uncertainty and decision costs from a different angle.

What elasticity are you trying to estimate?

Price elasticity of demand describes the proportional change in quantity associated with a proportional price change. For a differentiable demand curve, local elasticity is (dQ/dP) × (P/Q). A causal estimate asks how quantity would change under a price intervention for the specified population and conditions.

Ask whether quantity means units sold, conversion per eligible visitor, purchases per account, or retained subscriptions. If traffic differs between two arms, comparing raw unit counts can confuse exposure with demand response. Normalize to the appropriate eligible population or compare a properly randomized outcome per assigned unit.

Also specify the price: list price, paid price after discounts, delivery-inclusive price, or total subscription cost. A nominal price increase accompanied by a coupon may leave the effective price unchanged. Tax, shipping, bundles, and financing terms can change the actual choice a customer faces.

Short-run and long-run responses differ. Customers may buy early before an announced increase, delay purchases during a temporary promotion, or cancel only at renewal. A seven-day conversion experiment cannot by itself establish annual subscriber elasticity.

Editorial answer pattern: “I would estimate the effect of the offered all-in price on purchases per eligible account over a defined window, including nonbuyers. I would separately report revenue and contribution per assigned account, then check whether inventory or repeat exposure changes the interpretation.”

Why can historical price and demand rise together?

Original example: A platform charges $10 on quiet days and sells 100 units; it charges $12 on busy days and sells 150. A naive percentage calculation suggests demand rises when price rises. But the platform may have raised prices because it anticipated stronger demand.

Draw the causal diagram: a demand shock points to both price and quantity, while price itself also points to quantity. The positive association can be dominated by the common cause even when a higher price reduces demand within a comparable situation.

Official engineering explanation: Uber describes how observational comparisons can reverse a plausible causal story when groups differ systematically, and why choosing adjustment variables requires domain knowledge. Its example concerns delivery delays and engagement, not a reported pricing interview. We apply the same identification principle to this original price case. Causal inference at Uber.

Adding more predictors is not automatically a solution. A variable affected by price may be a mediator rather than a pre-treatment confounder. Controlling for it can remove part of the total effect you intended to estimate. Conditioning on purchase also excludes the people whose decision not to buy is central to elasticity.

For example, realized shipping service or basket size may change after seeing a price. Whether to adjust for either depends on the causal question and timing. Write the assumed graph before deciding that every available column belongs in the regression.

Promotions, competitor prices, seasonality, inventory, and customer mix are plausible pre-treatment factors, but measuring several of them does not prove all confounding is removed. State the remaining identification assumption rather than treating model complexity as evidence of causality.

Work through a randomized price calculation

Original exercise: Equal-sized, comparable randomized groups receive a $10 or $11 offer. The product, eligibility, stock availability, and outcome window are otherwise comparable. The lower-price group buys 1,000 units; the higher-price group buys 950. Assume sufficient exposure and complete observation for this point-estimate exercise; no sampling uncertainty has yet been calculated.

Using the original price and quantity as denominators:

Quantity change = (950 − 1,000) / 1,000 = −5%
Price change = (11 − 10) / 10 = 10%
Finite-change elasticity estimate = −5% / 10% = −0.50

This is a base-period percentage calculation. It is not exactly the derivative at $10. A log-ratio estimate is log(950/1000) / log(11/10), approximately −0.538. The midpoint, or arc, calculation is approximately −0.538 as well. These methods need not give identical answers for finite changes.

Say which convention you use. The interviewer may be checking whether you understand the denominator rather than whether you memorized a single formula. For small changes the conventions are close; for large changes the distinction matters more.

Now compute business outcomes with a constant $6 variable cost per unit:

Offer	Revenue	Contribution before fixed costs
$10, 1,000 units	$10,000	($10 − $6) × 1,000 = $4,000
$11, 950 units	$10,450	($11 − $6) × 950 = $4,750

Revenue rises 4.5%, while contribution rises 18.75%. Unit sales fall, but the larger unit margin more than compensates in this example. An answer focused only on conversion would miss the stipulated business objective.

The arithmetic is a point estimate, not a launch authorization. We still need uncertainty, customer retention, refunds, capacity effects, and the cost of implementation. A higher current contribution could be offset by lower future demand.

Can you calculate the break-even demand loss?

The contribution comparison gives a useful sensitivity check. At $10 and $6 variable cost, baseline contribution is $4,000. At $11, the margin is $5 per unit. Matching baseline contribution therefore requires 4000/5 = 800 units.

A fall from 1,000 to 800 units is a 20% quantity decline. Under these exact assumptions, selling more than 800 units at $11 improves contribution; selling fewer reduces it. This threshold concerns contribution, not revenue, and assumes the cost per unit remains $6.

For revenue alone, the threshold is 10000/11, approximately 909.1 units. A result of 880 units would reduce revenue but increase contribution: $9,680 revenue and $4,400 contribution. That counterexample is useful when an interviewer asks which metric should decide the test.

Do not extrapolate a two-price result into a global optimum. Demand may curve, customers may switch plans, and costs may change with volume. Estimating a local slope near $10 does not validate a $20 recommendation.

If the company wants an optimal price, discuss a safe range of price points, precision, model checking, and the cost of exploration. Optimize the expected objective subject to capacity and customer constraints, then test the proposed policy. An impressive elasticity estimate does not remove the need to validate the decision it supports.

How would you design the price experiment?

Start with the business question and eligible population. New-customer acquisition, existing-customer renewal, and marketplace surge pricing are different experiments. Existing contracts and customer expectations may constrain which prices can be offered and when.

Original design contract: For a subscription acquisition case, assign eligible organizations consistently to one offer, analyze revenue or contribution per assigned organization, and follow a prespecified conversion and retention window. Keep assignment stable across sessions and devices where identity allows. State how anonymous visits and later account creation are reconciled.

Measure exposure separately from assignment. A user may be assigned but never see the pricing page. The main intention-to-treat estimate can preserve the randomized comparison; conditioning only on post-assignment visitors may introduce selection if treatment changes who reaches that page.

Predefine the primary outcome, minimum useful effect, sample-size assumptions, monitoring rule, and guardrails. Refunds, cancellation, support burden, and perceived inconsistency may matter. Do not stop at the first favorable daily p-value unless the analysis was designed for that monitoring.

In marketplaces, prices can change shared supply and competitors' opportunities. Primary research: Holtz and coauthors show why interference matters in marketplace pricing experiments and study clustered assignment. The appropriate unit depends on the connections the policy changes, not on which unit is easiest to randomize. Marketplace pricing experiments.

A market-time switchback may be suitable for a reversible policy with manageable carryover. It is less convincing when customers stockpile, sellers reposition for days, or the treatment changes persistent expectations. For those cases, explain a longer assignment or cluster design and its power limitations.

What if randomization is unavailable?

Begin by asking why. Sometimes the constraint affects only a subset of customers or a narrow price range, leaving a useful experiment possible. If the decision truly relies on observational data, make the assumptions visible.

Primary methods example: The Pricing Engine paper applies double machine learning to pricing effects in panel settings. Flexible nuisance models can help estimate relationships while protecting the target estimate from certain modeling errors. They do not establish that unobserved confounders are absent or that every instrument is valid. Pricing Engine.

With adjustment methods, defend conditional exchangeability and overlap: comparable units must receive different prices after conditioning on the chosen pre-treatment variables. If all premium customers always receive one price and all others another, the data may offer little support for the comparison you want.

With an instrumental variable, explain why the proposed instrument changes price and why it affects demand only through the relevant treatment pathway, alongside the other required assumptions. A cost shock may also change availability or delivery quality, violating a casual exclusion argument.

With a discontinuity or difference-in-differences design, explain the actual assignment mechanism, continuity or trend assumptions, and potential concurrent changes. Naming a causal method without defending its identification is weaker than stating that the available data cannot answer the question reliably.

Use sensitivity analysis and a limited follow-up experiment where possible. A pricing model should help expose where the recommendation is fragile, not conceal that fragility behind one precise coefficient.

Practice the causal and economic follow-ups

These related exercises test pricing calculations, marketplace responses, and experiment design. Their inclusion does not establish a company's current interview format.

Practice question	What to rehearse
Develop Dynamic-Pricing Algorithm for Lyft Balancing Key Factors	Separate causal demand response from supply reallocation.
Calculate Profit-Maximizing Price and Validate with Additional Data	Connect the price decision to marginal contribution.
Visualize Price Impact on Demand and Profit Trends	Explain why revenue and contribution can favor different prices.
Design A/B Test for Subscription Price Increase Effectiveness	Define assignment, retention horizon and price exposure.
Estimate price–ETA trade-offs causally	Defend identification when price and service quality co-vary.

Explain the $10-to-$11 experiment aloud, then change the outcome to 880 units. Your answer should distinguish the elasticity convention, revenue, contribution, uncertainty, and what remains unidentified before recommending a price.

Sources and Further Reading

• Uber: Using Causal Inference to Improve the Uber User Experience — official engineering discussion of confounding and adjustment.
• Pricing Engine: Estimating Causal Impacts in Real World Business Settings — primary methods paper on pricing-effect estimation.
• Holtz and coauthors: Reducing Interference Bias in Online Marketplace Pricing Experiments — primary evidence on marketplace experiment design.

Research checked September 21, 2026. The examples are simplified analytical exercises, with stated assumptions rather than claims about any company's actual prices.