Interview guides · Haystack People

Haystack People Data Scientist Interview Questions & Guide 2026

A two-week prep plan for Haystack People data scientist candidates: write your answer before reading the solution, revisit misses three days later.

The 60-second brief

As a Data Scientist at Haystack People, you are at the intersection of product innovation and analytical rigor. Your primary objective is to transform complex datasets into actionable product strategies, directly influencing how our users interact with our platforms. You will work closely with cross-functional teams, including product managers and engineers, to define success metrics, build predictive models, and ensure our growth is backed by sound data.

Seniority shifts the scope more than the words in the title do. Earlier-career loops mostly check that you execute a well-posed analysis correctly; senior loops check that you can decide which question is worth answering and defend what you chose not to do.

Focus areas

Set fraud thresholds by expected cost
Decompose expected loss into PD, LGD, EAD
Reconcile amounts in minor units and currency

Overview

This role is critical to the success of Haystack People because we operate in a landscape where evidence-based decision-making is our competitive advantage. You will not just be reporting numbers; you will be the architect of our experimentation culture. Whether you are diagnosing a sudden drop in engagement metrics or designing an A/B test for a new feature, your insights will guide the direction of our product roadmap and ensure we are solving the right problems for our users.

The Haystack People interview process

Haystack People candidates report 5 rounds · ≈ 4-6 weeks. The stages below are what candidates describe, not a published process.

Initial Screening

Most candidates lose this call inside the first two minutes, during the walkthrough of their own background. The account runs chronologically, sits at the level of tools and titles, and never arrives at a decision anyone could have disagreed with. Anchor on a problem instead of a timeline: what the team could not answer, what you did about it, what happened next. Ninety seconds is enough, and stopping on time leaves room for the half of the call that belongs to you. What you ask about how work gets prioritised signals your level more reliably than the walkthrough does.

What to demonstrate

Whether your background summary has a shape (problem, decision, consequence) or is a chronological list of tools and employers
Whether you can account for gaps, short stints and the reason you are looking, unprompted and without hedging
The substance of the questions you ask back, which an experienced screener reads as a level signal

How to prepare

Time your opening walkthrough against a clock. If it runs past two minutes, compress the earliest role into a single clause and spend the recovered time on the most recent one
Write one honest sentence for every gap or short stint visible on your resume and offer it before being asked about it
Prepare questions about how work arrives and gets prioritised: who writes the request, how often priorities change, and what happens to an analysis after it is delivered

Technical Assessments

Before anything else, this round is a reading test. You are given a small schema and a question phrased in business language, and most of the difficulty sits in the gap between them. Who counts as an active user, does a refunded order still count as an order, is that date column an event time or a load time. Weak answers start typing immediately and compute something precise about the wrong population. Strong ones pin the definition in one sentence, name the column that encodes it, then write the query. On a timed assessment with nobody to tell, write the definition in a comment anyway.

What to demonstrate

Whether an ambiguous term becomes a specific column and filter before any computation happens
Whether you read the schema for keys and cardinality rather than only for column names
Whether the result answers the question at the grain it was asked at, per user or per session or per day

How to prepare

Take three metrics you already use and write down the exact filter and exact grain behind each, then practise stating one of them in a single sentence out loud
On a schema you have never seen, spend the first minute writing what one row of each table means and which key it is unique on, then predict which joins can duplicate rows
Rehearse a version where the definition changes halfway through, and edit the query you have instead of starting over

Product Case Study

This round runs as a working session, so part of what it decides is whether you are useful to think with. The interviewer will interrupt: a hint that the data you assumed does not exist, a challenge to your metric, a nudge toward a branch you skipped. Treating those as interference is the common failure. Reason out loud while your thinking is still provisional so there is something to react to, and when a redirect arrives, take it instead of defending the path you had already started down.

What to demonstrate

Whether your reasoning is audible while it is still unsettled, or only after you have privately decided
What you do with a hint: absorb it and adjust, or argue for the original route
Whether your clarifying questions have answers that would change your approach, as opposed to filling silence
Whether you can be wrong about something in the middle of the case and keep moving without restarting

How to prepare

Run practice cases with a partner instructed to interrupt twice: once to remove a data source you assumed existed, once to reject the metric you chose. Practise absorbing both without going back to the start.
Before each practice case, write down the clarifying questions you plan to ask, then check afterwards whether any answer actually changed what you did. Drop the ones that did not.
Explain an analysis you already know well to someone outside the field and have them stop you at every point where the reasoning jumped a step.

Behavioral Interviews

Most of the weight in this round sits on the disagreement questions. Data work routinely produces an answer someone senior did not want, and the interviewer is trying to learn what you do in that hour. Both failure modes are common: folding as soon as a director pushes back, and treating the pushback as ignorance to be corrected with a better chart. A strong answer usually contains a specific thing the other person knew that you did not, and describes how you found out whether it changed the conclusion.

What to demonstrate

Whether you can state the other side's argument accurately before you explain why you disagreed
What you treated as evidence during the disagreement, such as a rerun under their assumption or a holdout check, rather than persuasion technique
Whether you distinguish being overruled from being wrong, and can give an example of each

How to prepare

Write out one disagreement where you turned out to be wrong, and say what in the data misled you. Candidates prepare the story where they were right, and the follow-up asks for the other one.
For your main disagreement story, be ready to say what result would have made you drop your position. If no such result exists, you were not arguing from the data.
Practise stating the opposing position out loud in one sentence the stakeholder would accept, then continue the story.

Team Interaction

An added round often puts you in front of someone outside the core hiring team: a partner engineer, a product owner, a domain expert, sometimes a more senior manager. The question they are really asking is not whether you can do the work but whether they would trust a number that came from you. That changes what a good answer looks like. Lead with what the decision cost and what it changed, keep the method available but not central, and be plain about the limits of your evidence. Overstating a result is the fastest way to lose this round.

What to demonstrate

Whether you can explain a technical choice to someone who will never read your code, without either flattening it into nothing or hiding inside jargon
Honesty about evidence strength: what the analysis establishes, what it only suggests, and what it cannot say at all
How you take disagreement, specifically whether you update on a good objection, hold your position with reasons, or fold on contact

How to prepare

Write the two-sentence version of your most technical project for a non-specialist, then check that neither sentence needs a method name to make sense.
For one result you are proud of, write the strongest objection someone could raise and a response that concedes the part of it that is correct.
Prepare one decision that turned out to be wrong: how you found out, what it cost, and what you changed afterwards. A senior cross-functional interviewer asks for this more often than a technical one does.

Where candidates lose points

Assuming a model is fair because protected attributes are not among its inputs

Postcode, device, tenure, income proxies and even transaction patterns correlate with protected characteristics, so a model can produce a disparate outcome without ever reading the attribute. Credit decisions additionally carry an explainability obligation in many jurisdictions, since a denial has to be accompanied by its principal reasons, which constrains model form and feature engineering rather than being a reporting afterthought. Treating fairness testing and reason-code generation as design constraints from the first model version is far cheaper than retrofitting them to a deployed one.

Averaging delinquency across a book that is growing

A loan three months old cannot be 90 days past due, so a portfolio with many recent originations reports a low blended 90+ rate purely from age mix. The blended rate falls fastest exactly when originations grow fastest, which is precisely when credit quality most needs watching, so the metric moves in the reassuring direction during the riskiest period. Only comparisons at equal months on book are valid, which is what a vintage or roll-rate view enforces.

Writing SQL without stating NULL and tie-breaking behaviour

Before calling a query finished, say what it does with NULLs, ties and empty groups. NOT IN against a subquery containing a single NULL returns no rows at all, and RANK, DENSE_RANK and ROW_NUMBER differ precisely on ties, so name which one the question requires.

Comparing periods without accounting for seasonality or day-of-week

Compare whole weeks against whole weeks and check whether the same swing appeared in prior cycles or prior years before attributing it to anything you changed. Weekday and weekend populations often differ enough that a Tuesday-to-Saturday comparison is meaningless.

Questions & practice

Bootstrap a fraud loss rate that clusters within merchant

Medium · bootstrap · clustered resampling · heavy tails

You have a per-transaction frame with auth_id, merchant_id, settled_amount_reporting and net_loss_reporting, both already in one reporting currency. Most rows carry zero loss, a few carry large ones, and losses cluster within merchant. Using only numpy's random generator and no resampling helper from any library, write a bootstrap that returns a 95 percent interval for net fraud loss in basis points of settled volume, resampling merchants with replacement and taking all rows belonging to each drawn merchant. Also produce the naive row-level interval and state which you would report.

Approach

State the estimator before writing it: total net loss divided by total settled volume, times 10,000. It is a ratio of sums, so each replicate recomputes both sums. Averaging per-transaction loss rates instead would weight a five-unit transaction like a five-thousand-unit one.
Pre-aggregate loss and volume to merchant level once. For a ratio of sums, drawing merchants and taking all their rows is arithmetically identical to drawing merchant-level (loss_sum, volume_sum) pairs, so a replicate becomes one integer draw plus two vectorised sums rather than a groupby inside the loop.
Draw B replicates of M merchant indices with replacement, where M is the observed merchant count, compute the ratio per replicate, and take the 2.5th and 97.5th percentiles. Say explicitly that this is a percentile interval and that BCa would correct the skew-induced bias if the decision is close.
Repeat with independent row draws for the naive interval and compare widths on the same replicate count.
Report the clustered interval. Rows within a merchant share an acceptance profile, a category code and a fraud exposure, so they are not independent, and the row-level interval understates variance by roughly the design effect.

Follow-up

Your clustered interval is three times wider. How do you explain that to someone who wanted a tighter number?
One merchant accounts for 40 percent of losses. What does that do to the interval, and what would you do about it?
How does this change if the question is whether two months differ rather than what this month's rate is?

Build a vintage delinquency table without pivot or unstack

Easy · vintage analysis · cohorts · groupby · cumulative max

fct_loan_performance_monthly gives loan_id, origination_month, months_on_book, days_past_due, charge_off_flag and restructured_flag. Produce a DataFrame with one row per origination_month and columns for months_on_book 0 through 12, each cell holding the share of that vintage's funded loans that had ever reached 90 or more days past due, or charge-off, by that age. You may not use pivot, pivot_table, crosstab or unstack. Cells for ages a cohort has not yet reached must be NaN rather than zero.

Approach

Define the per-row indicator as days_past_due >= 90 or charge_off_flag, then take a cumulative maximum of it per loan ordered by months_on_book, because the metric is reached-by-age-m, not in-that-state-at-age-m.
Deal with restructuring before the cumulative max. Restructuring resets days_past_due, so a restructured loan re-enters at current and, without the cumulative maximum carrying its pre-restructure worst state, reads as a cure.
Fix the denominator once as the count of distinct loan_id per origination_month across the whole cohort. Prepaid and charged-off loans stop producing rows, so a denominator recomputed at each age silently shrinks exactly where losses land.
Aggregate with groupby(['origination_month','months_on_book'])['ever_90'].sum(), then pre-build the output frame indexed by sorted origination months with integer columns 0 to 12 and assign from the grouped Series by .loc on its index.
Mask cells beyond each cohort's maximum observed months_on_book so an immature cell reads NaN instead of an artificially low rate.

Worked solution (30 min)

Sort by loan_id and months_on_book, build the ever_90 indicator, then apply groupby('loan_id')['ever_90'].cummax().
Compute cohort_size as the distinct loan_id count per origination_month, before any filtering on age.
Group the cumulative indicator by origination_month and months_on_book and sum it to get the numerator per cell.
Create the output frame with the sorted origination months as index and range(0, 13) as columns, fill it from the grouped Series, and divide each row by its cohort_size.
Compute each cohort's maximum observed months_on_book and set every cell to the right of it to NaN.

Expected result: A DataFrame with one row per origination_month and columns 0 to 12, values non-decreasing left to right within every row, near zero in column 0, and NaN forming an upper-right triangle where cohorts have not yet aged.

Verify it yourself

Every row is non-decreasing across columns; a decrease means the cumulative maximum was not applied or a restructure reset leaked through.
The denominator per row equals the cohort's distinct loan count and does not change across columns.
Column 0 is at or near zero for every vintage, since a loan cannot be 90 days past due at its first month end.
origination_month plus months_on_book is constant along an anti-diagonal, so one anti-diagonal should reproduce a direct calendar-month filter.

Follow-up

Two adjacent vintages diverge at months_on_book 6. How would you separate seasoning, mix shift and a genuine credit-quality change?
The three most recent vintages look best on this table. What do you check before saying so?
How does the table change if charge-off policy moved from 180 to 120 days past due partway through the series?

Implement accident-quarter loss ratio at twelve months development

Medium · loss ratio · development age · earned premium

fct_policy_period_monthly arrives as a stack of month-end snapshots: each row carries valuation_month alongside as_of_month, policy_id, product_line, written_premium_minor, earned_premium_minor, paid_loss_minor, case_reserve_minor, ibnr_reserve_minor and loss_adjustment_expense_minor. Compute the accident-quarter loss ratio at exactly 12 months of development: incurred losses over earned premium, both taken from rows whose as_of_month falls in the accident quarter, read from the snapshot 12 months after that quarter closes. Report quarters that cannot reach that age as incomplete rather than dropping them.

Approach

Derive accident_quarter from as_of_month, then define the evaluation snapshot per quarter as valuation_month equal to the quarter's final month plus twelve months. Every figure in the ratio comes from that one snapshot, not from whichever snapshot happens to be newest.
Numerator is paid_loss_minor plus case_reserve_minor plus ibnr_reserve_minor over the accident quarter's rows in that snapshot. Loss adjustment expense may be included or not, but the choice applies to every quarter and is named in an output column.
Denominator is earned_premium_minor over the same rows. Written premium is booked in full at inception, so in a growing book it runs ahead of earned premium and drags the ratio down, with the error reversing when the book shrinks.
Left-join the full quarter list against available valuation months so a quarter with no 12-month snapshot yields status incomplete and a null ratio, instead of disappearing and shortening the series without saying so.
Split by product_line, since both the loss ratio level and the speed of development differ by line, and a blended series moves with mix as much as with experience.

Follow-up

The most recent complete quarter came in four points better than the one before. What do you check before calling it an improvement?
How would you estimate the 12-month figure for a quarter that is only 6 months developed, and how would you label the estimate?
Why can an expense ratio legitimately use a different denominator from the loss ratio in the same presentation?

Using SQL window functions, how would you calculate the running total of user signups over the last 30 days?

Medium · sql and data manipulation

Approach

Check whether any join is one-to-many before aggregating, or the sums inflate.
Compute rates by summing numerator and denominator separately, never by averaging rates.
State the window function and its partition and ordering out loud before writing it.

Follow-up

What breaks if events arrive late or out of order?
How does the query change if the join becomes one-to-many?

Write a query to identify top-performing users based on specific engagement thresholds.

Medium · sql and data manipulation

Approach

State the window function and its partition and ordering out loud before writing it.
Say which table is the grain you start from, and join outward from it.
Compute rates by summing numerator and denominator separately, never by averaging rates.

Follow-up

How does the query change if the join becomes one-to-many?
What breaks if events arrive late or out of order?

Accident-quarter loss ratio on earned rather than written premium

Hard · loss ratio · earned vs written · development age

From fct_policy_period_monthly, compute the accident-quarter loss ratio by product_line: incurred losses, being paid_loss_minor plus case_reserve_minor plus ibnr_reserve_minor, over earned_premium_minor for the same accident quarter. State explicitly whether loss_adjustment_expense_minor is included and apply that choice consistently. Also output the same ratio computed on written_premium_minor so the two can be compared. The table holds current values with no valuation-date snapshot. Say in one line which comparison this schema cannot support and what you would need to support it.

Approach

Derive the accident quarter from as_of_month with date_trunc, and note that the table already attributes losses to the month of the loss event while earning premium pro rata into the same month, which is what makes the two sides comparable at all.
Aggregate earned_premium_minor, written_premium_minor and the three loss components to product_line and accident quarter in one pass, keeping loss adjustment expense as its own column so the inclusion choice is a final-select decision rather than something buried in a CTE.
Compute both ratios side by side and a third column for their difference, because the size and sign of that difference is a direct read on whether the book grew or shrank in the quarter.
State the limitation plainly: every row carries today's reserve estimate, so each accident quarter is observed at a different development age and a cross-quarter comparison mixes development with underwriting. A fixed development age needs a valuation-date dimension, that is one row per accident period per valuation, which this table does not have.
Guard against the mirror-image error on the numerator by confirming ibnr_reserve_minor is non-zero on recent quarters; if it is null or zero there, the recent periods are understated twice over and the series is not usable.

Worked solution (40 min)

CTE quarterly: group fct_policy_period_monthly by product_line and date_trunc('quarter', as_of_month), summing earned_premium_minor, written_premium_minor, paid_loss_minor, case_reserve_minor, ibnr_reserve_minor and loss_adjustment_expense_minor.
Final SELECT: build incurred_minor as the three loss components plus the LAE column, with the LAE inclusion written as a named expression so the choice is visible on the page.
Emit loss_ratio_earned and loss_ratio_written, both cast to numeric, plus their difference and the written-to-earned premium ratio.
Order by product_line and accident quarter, and append the one-line note about the missing valuation dimension to the query as a comment.

Expected result: One row per product_line and accident quarter with earned and written premium, incurred losses, both ratios and their difference. In quarters where written premium exceeds earned, loss_ratio_written is strictly below loss_ratio_earned, and the gap widens with the growth rate.

Verify it yourself

For a mature product line with flat volume, summed earned premium over a full policy term is close to summed written premium, so the two ratios converge there.
The sign of loss_ratio_earned minus loss_ratio_written matches the sign of written premium minus earned premium in every quarter, which is an arithmetic identity and not a data finding.
Recent accident quarters have non-zero ibnr_reserve_minor; a zero there means the numerator is missing its tail and the ratio is understated.
Within a quarter, loss_ratio_earned divided by loss_ratio_written equals written_premium_minor divided by earned_premium_minor exactly, because both ratios share one numerator. Any deviation means the LAE term was built into one ratio and not the other, which is the real test that the inclusion choice was applied consistently.
Recomputing with LAE excluded scales each quarter's ratio by incurred_minor / (incurred_minor + lae_minor). That factor is not constant across quarters unless LAE happens to be booked as a fixed fraction of incurred losses, so a uniform shift is evidence about how LAE is loaded, not evidence that the query is self-consistent — use the previous check for that.

Follow-up

Written premium exceeds earned premium by 18 percent this quarter and by 3 percent two years ago. What happened to the book, and what does it do to each ratio?
How would you build a development triangle from a valuation-dated version of this table, and what would you use the chain-ladder factors for?
Statutory presentation conventionally takes the expense ratio on written premium while the loss ratio uses earned. How do you avoid a combined ratio that quietly mixes the two bases?

We noticed a 10% drop in daily active users over the last week; how would you investigate the cause?

Medium · product sense and metrics

Approach

Decompose the metric into the rates that drive it, and say which one you would check first.
Fix the population and the time window before naming any metric.
Name one primary metric, then the guardrail that stops it being gamed.

Follow-up

Which segment would you cut first, and what would that rule out?
What would you do if the primary metric and the guardrail moved in opposite directions?

How would you design a metric to track the quality of user recommendations?

Medium · product sense and metrics

Approach

Decompose the metric into the rates that drive it, and say which one you would check first.
State what result would change your recommendation, so the answer is falsifiable.
Name one primary metric, then the guardrail that stops it being gamed.

Follow-up

What would you do if the primary metric and the guardrail moved in opposite directions?
How would you detect that the metric is being gamed rather than genuinely improving?

How do you prioritize your work when you have multiple competing requests from different teams?

Medium · product sense and metrics

Approach

Decompose the metric into the rates that drive it, and say which one you would check first.
State what result would change your recommendation, so the answer is falsifiable.
Restate the decision this analysis has to support, and who acts on the answer.

Follow-up

How would you detect that the metric is being gamed rather than genuinely improving?
Which segment would you cut first, and what would that rule out?

How would you measure the success of a new feature that allows users to collaborate on projects in real-time?

Medium · product sense and metrics

Approach

Decompose the metric into the rates that drive it, and say which one you would check first.
Restate the decision this analysis has to support, and who acts on the answer.
State what result would change your recommendation, so the answer is falsifiable.

Follow-up

Which segment would you cut first, and what would that rule out?
What would you do if the primary metric and the guardrail moved in opposite directions?

How do you determine the required sample size for an A/B test before launching?

Medium · a/b testing and experimentation

Approach

Say whether units interfere with each other, and switch design if they do.
State the primary metric and the minimum effect worth shipping, then size the test.
Name the guardrails that would stop a launch even on a positive primary result.

Follow-up

How would you handle interference between treated and control units?
What would you conclude if the result is positive but the test is underpowered?

What are the most common experimentation pitfalls you have encountered, and how do you mitigate them?

Medium · a/b testing and experimentation

Approach

Name the randomisation unit first; it decides the variance and what the test can detect.
Say whether units interfere with each other, and switch design if they do.
Decide the analysis before seeing data, including how long it runs and when you look.

Follow-up

What would you do if you could not randomise at all?
How would you handle interference between treated and control units?

Estimate the marginal effect of approval at a score cutoff

Hard · regression discontinuity · instrumental variables · selection

Policy auto-approves applications with bureau_score at or above 660 and routes 640 to 659 to manual review, which approves about 35 percent of them. You cannot randomise approvals. Using fct_loan_application and a bureau-sourced panel that reports 12-month serious delinquency on any trade line for approved and declined applicants alike, estimate the effect of approval at the margin. State the identifying assumptions you would test rather than assert, the estimator and its tuning choices, and exactly what the estimate does and does not license.

Approach

Set the design up as a fuzzy regression discontinuity, because crossing 660 changes approval probability sharply but not from zero to one. Let Z = 1{bureau_score >= 660} be the instrument, approval be the treatment, and 12-month serious delinquency be the outcome. The estimate is the ratio of the jump in the outcome to the jump in approval probability, which is two-stage least squares with Z as the instrument and identifies a local average treatment effect for compliers at the cutoff.
Insist on the bureau panel as the outcome source. Funded loans exist on both sides of 660 only because manual review approves some applicants below it, and those are selected on whatever the reviewer saw; an outcome defined only on funded loans would reintroduce exactly the selection the design is meant to remove.
Test the assumptions instead of listing them. Check density continuity of bureau_score at 660 with a McCrary-style or local-polynomial density test, since brokers and applicants can sometimes trigger a bureau refresh and a heaped density at the threshold kills the design. Check that pre-determined covariates, including declared_annual_income_minor, channel, product_code and model_version, are continuous at the cutoff. Check policy_rule_hits and the pricing table for any other rule that fires at exactly 660, because a second discontinuity at the same point is not separable from the first.
Choose the estimator deliberately. Local linear regression with a triangular kernel and an MSE-optimal bandwidth, with bias-corrected robust confidence intervals. Do not fit a high-order global polynomial, which puts weight on observations far from the cutoff and produces artefacts at the boundary. Because bureau_score is an integer, the running variable has mass points, so cluster standard errors by score value or use an approach designed for discrete running variables.
Respect outcome maturity. Include only application cohorts with a full 12 months of bureau observation; a recent cohort with partial observation will look cleaner and will drag the estimate.
State the limits in the deliverable. The estimate is the effect at 660 on applicants whose approval status is determined by the threshold. It does not license moving the cutoff to 620, because both the first stage and the outcome relationship differ away from 660 and because a policy change at scale shifts the applicant mix. Report bandwidth sensitivity alongside the point estimate.

Worked solution (40 min)

Assemble the application-level panel: bureau_score, decision, funded status, and the bureau-sourced 12-month serious delinquency flag, restricted to cohorts with 12 full months of outcome observation.
Plot the first stage, being approval probability against bureau_score in one-point bins, and confirm a visible jump of roughly 0.65 at 660 (about 1.00 above against about 0.35 below).
Plot the reduced form, being delinquency against bureau_score in the same bins, and read the jump at 660 off a local linear fit on each side.
Estimate the Wald ratio as reduced form over first stage, equivalently two-stage least squares with Z = 1{score >= 660}, using a triangular kernel, an MSE-optimal bandwidth, bias-corrected robust intervals, and standard errors clustered by integer score.
Run the assumption battery: density continuity at 660, covariate continuity, placebo cutoffs at 650 and 670, and bandwidth sensitivity at 0.5x and 2x the selected bandwidth.
Write the result as a complier average effect at 660 with its interval, its bandwidth sensitivity table, and an explicit sentence on what it does not license.

Expected result: A fuzzy RDD estimate delivered as the Wald ratio, jump in 12-month serious delinquency divided by a first-stage jump in approval probability of about 0.65, fitted by local linear regression with a triangular kernel at an MSE-optimal bandwidth, with robust bias-corrected intervals and standard errors clustered on the integer score. It is reported as a local average treatment effect for compliers at 660 only, accompanied by the density test, the covariate-continuity table and the bandwidth sensitivity, with an explicit statement that it does not support moving the cutoff.

Verify it yourself

Reduced form divided by first stage must equal the two-stage least squares coefficient to numerical precision; if it does not, the sample or the bandwidth differs between the two fits.
Placebo cutoffs at 650 and 670 should show no jump in either the first stage or the outcome.
Pre-determined covariates should be continuous at 660; a jump in income or channel mix at the threshold means applicants or brokers are sorting and the design is compromised.
The point estimate should be stable when the bandwidth is halved and doubled; large movement means the local linear fit is being driven by observations far from the cutoff.

Follow-up

Manual reviewers below the cutoff see documents the score does not. What does that do to the monotonicity assumption, and how would you look for defiers?
You are offered a randomised approval band across 650 to 659 for one quarter. What does it buy you that the discontinuity does not, and how would you price the expected loss of running it?
The density test shows a modest pile-up just above 660. What are the possible mechanisms, and which of them leave the design usable?

Manual review rate doubles on the day two changes ship

Hard · definition change · model version · population stability

Manual review rate on fct_loan_application, defined as decided_by = 'manual' over decisions in ('approve','decline'), stepped from 9 to 18 percent on a single day. That day a new model_version went live, and the same release changed which referral paths are recorded as manual. Columns: submitted_at, channel, bureau_score, model_pd_12m, model_version, policy_rule_hits, decision, decided_by, decision_at. Attribute the step across definition, model and population, with a number against each.

Approach

Establish the shape before the cause. A definition or code change produces a step at the deploy timestamp; a population change produces a ramp. Plot at hourly granularity either side of the release and read which one this is.
Dual-run the definition. Recompute the metric under both the old and the new recording rules across several months of history. If the historical series under the new rule sits near 18 percent throughout, that part of the step is definitional and should be restated, not investigated.
Hold the model constant and decompose the residual by policy_rule_hits identifier. A referral rule keyed to a score threshold shows its own step, and naming the rule is what makes the finding actionable to the policy owner.
Test the score, not the applicants. Compare model_pd_12m distributions by model_version on the same applications where shadow scores exist; otherwise compare adjacent weeks with a population stability index, treating the customary 0.1 and 0.25 marks as heuristics rather than tests, and assess calibration separately from ranking.
Test the applicants, not the score. Compare channel mix, bureau_score distribution and null-bureau share across the boundary. If those are stable, population is not a driver and should be reported as approximately zero rather than left unquantified.
Deliver three components that sum to the observed step, each with the team that owns it.

Follow-up

If the new model is better calibrated, is a higher referral rate the correct outcome or a threshold nobody retuned?
How would you have caught the recording change before it reached the dashboard?
What is the swap set here, and which swap group would you inspect first?

Suggested practice checklist

For a candidate whose interviews will centre on A/B testing, metric movement and causal claims. Design comes before arithmetic, arithmetic before analysis, and the week ends by rehearsing the readout rather than the derivation.

Day 1 — Design one test end to end on paper

Take a single feature change and write the full design: randomization unit, the exact point of exposure, the primary metric with its grain, guardrails, allocation, planned duration, and the decision rule committed before any data exists.
Write why the randomization unit must sit at or above the level where treatment can spill over, and give one case where user-level randomization is still contaminated (shared accounts or devices, or two participants in the same marketplace).
State in advance what you will do if the primary metric is flat while a secondary metric is significant.

Deliverable: A one-page test design with a decision rule written before launch.

Day 2 — Power arithmetic until it is automatic

Compute required sample size per arm for a binary metric with the normal approximation, n is approximately 2 times (z for alpha/2 plus z for power) squared times p(1 minus p) divided by delta squared, for baselines of 2, 10 and 40 percent at a 5 percent relative lift, and note that for a fixed relative lift the requirement falls as the baseline rises because delta grows proportionally with p.
Redo the calculation for a continuous metric using variance in place of p(1 minus p), and show why a heavy-tailed quantity such as revenue per user needs either far more traffic or a capped version with a stated cap.
Convert one of the results into weeks given a weekly eligible traffic figure, then list the two honest ways to shorten it (accept a larger detectable effect, or reduce variance) and write why quietly lowering the power target is a decision to miss more real wins, not a speedup.

Deliverable: A small script or sheet that maps baseline, minimum detectable effect, alpha and power to sample size and weeks, cross-checked against a published calculator.

Day 3 — Variance and the unit-of-analysis problem

Take a ratio metric whose denominator is not the randomization unit (clicks per session, randomized by user) and compute the standard error twice, once naively at session level and once by the delta method or a user-level bootstrap, then record how much the naive version understates it.
Implement CUPED on simulated data: choose a pre-period covariate X measured before assignment, estimate theta as Cov(Y, X) divided by Var(X), and analyse Y minus theta times (X minus its mean) in place of Y. Confirm the variance of the adjusted outcome equals the raw variance multiplied by one minus the squared correlation between Y and X, so a correlation of 0.45 removes about 20 percent of the variance and not 80.
Now run that simulation a few hundred times and confirm the adjusted effect estimate is unbiased for the same effect rather than numerically identical to the raw one. Within any single run the two differ, sometimes by a large fraction of the true effect, because the two arms' pre-period covariate means never coincide exactly in a finite sample; they agree in expectation, which is the property that matters and the one to state out loud.

Deliverable: A notebook showing the adjusted estimator with a measurably smaller variance than the raw one, plus a repeated-simulation table showing the two estimators agreeing on average while differing run by run.

Day 4 — Validity threats you can actually test for

Run a sample ratio mismatch check as a chi-square goodness-of-fit test against the intended allocation, and write the three causes you would chase first (assignment logged before exposure, an arm-specific redirect or load failure, bot filtering applied asymmetrically).
Simulate peeking: generate A/A data, test daily at alpha 0.05 across 14 looks, record the inflated false positive rate, then apply an alpha-spending boundary or commit to a fixed horizon and confirm the rate returns to nominal.
Write how you would separate a novelty effect from a durable lift using the treatment effect plotted against days since first exposure, and what shape would change your recommendation.

Deliverable: One table showing the peeking false positive rate before and after correction, plus a written SRM triage list.

Day 5 — When randomization is not available

Write the identifying assumption for difference-in-differences (parallel trends in the absence of treatment), then plot pre-period trends for two candidate control groups and justify rejecting one of them.
Design a switchback test for a change where user-level randomization would leak across participants, choosing a time-block length against the carryover you expect and saying how you would detect carryover in the data.
List what an interrupted time series or a synthetic control buys you and the one thing neither can rule out: an unobserved shock that coincides with the launch.

Deliverable: A one-page memo recommending a single quasi-experimental design and naming its weakest assumption explicitly.

Day 6 — The readout query

Write the assignment-to-exposure join that returns exactly one row per unit per experiment, and handle units appearing in both arms by excluding and counting them rather than silently keeping one.
Compute the per-arm metric, its variance and the relative lift with a confidence interval in SQL, then reproduce the identical numbers in a notebook as a cross-check.
Add a segment breakdown and write the sentence that keeps it from being p-hacking: segments declared in advance, everything else reported as exploratory and corrected for multiplicity.

Deliverable: A single query that outputs the full readout table, matched to a notebook recomputation.

Day 7 — Present it to someone who will not read the appendix

Give a 10-minute readout of a real or simulated experiment in the order decision, number, uncertainty, caveat.
Have your listener ask "can we ship it" in the case where the primary is flat and a guardrail moved, and answer with a recommendation rather than a request for more data.
Rewrite your opening line so the recommendation lands before any methodology.

Deliverable: A one-page readout whose first line is the recommendation.

Explaining a method to someone who will not follow the maths

Half of this section is about translation. Be ready to describe how you explained a result to someone who did not want the method, only the implication, and what you did when the simplified version started being repeated in a way that overstated it. Correcting your own simplification is a strong beat.

Explain the concept of statistical significance to a non-technical stakeholder.

Medium · behavioural and stakeholder questions

Approach

Close with what you would do differently, concretely.
Pick a story where you drove the decision, not one where you observed it.
Name the disagreement or constraint, and how you resolved it with evidence.

Follow-up

What would you do differently if you ran that project again?
What did you decide not to do, and why?

Allocate one analyst-week across three competing risk requests

Medium · prioritisation · decision deadlines · stakeholder negotiation

Three requests land in the same week and you have one analyst-week. Payments wants a merchant-level decline teardown before a contract renewal in nine days. Credit wants a swap-set analysis on a cutoff change scheduled to ship in six weeks. Insurance wants accident-quarter loss ratios at 12 months development for a reserving review with no fixed date. Each sponsor believes theirs is first, and each has escalated before. Produce the allocation, the reasoning you would say out loud to all three at once, and what you explicitly drop.

Approach

Score each request on the decision it unblocks rather than on effort or on how loudly it arrived: what changes if it is late, and is that change reversible.
Separate deadline from value. The nine-day renewal is a hard, irreversible date with a bounded prize; the six-week cutoff has slack but a much larger downside if it ships unmeasured; the reserving number has no date but feeds external reporting, which is its own kind of hard.
Hunt for the cheap partial in each: a decline teardown restricted to the top merchants by declined value usually answers the contract question at a fraction of the full cut.
Sequence by hard date first, then by largest irreversible downside, and deliver the trade-off to all three sponsors in one message rather than three, so nobody negotiates privately against a version you told someone else.
Name what is dropped and who now owns that consequence, in writing, so the trade-off is visible rather than silently absorbed by you.

Follow-up

The credit sponsor escalates to your manager. What do you change, and what do you refuse to change?
How would you make this allocation reproducible so the next contested week is a rule application rather than a negotiation?

Defend a vintage finding that contradicts the portfolio dashboard

Medium · vintage analysis · mix shift · stakeholder pushback

The lending dashboard shows blended 90-plus days-past-due falling for four consecutive quarters while originations grew 60 percent. Using fct_loan_performance_monthly, you build a vintage view keyed on origination_month by months_on_book and find the three most recent vintages are worse than their predecessors at the same age. The business lead presents that dashboard weekly and pushes back hard, suggesting you picked favourable cohorts. You get one meeting and the vintage table. Present the finding so it survives the cherry-picking objection and ends in a decision.

Approach

Reconcile before you contradict: show that aggregating your vintage table along the calendar diagonal reproduces the published blended series, so the disagreement is about age mix rather than about data quality.
Make the mechanism arithmetic rather than rhetorical: a loan cannot reach 90 days past due before it is 90 days old, so rapid origination growth shifts weight onto young months-on-book where the rate is structurally near zero.
Show every vintage rather than a selected pair, all indexed at months_on_book equal to 12, with cohort sizes printed beside each curve so nobody can claim the divergence rests on a thin cohort.
Handle restructuring explicitly, because restructured_flag resets days_past_due: count each loan on its worst pre-restructure state, or recent vintages will look better than they are.
Close on the decision rather than the chart: state what the divergence implies for the cutoff or the channel mix, and state in advance what evidence would make you withdraw the claim.

Follow-up

Two cohorts differ at month 12. How do you separate a seasoning effect from a genuine credit-quality effect?
Someone argues the recent vintages are simply a broker-channel mix shift. How do you test that, and what would confirm it?