Banking and credit case interview questions for analysts

Round 1: vocabulary you are expected to use correctly

Q1. What does "30+ DPD" mean, and why track several buckets?

Days past due: how many days a payment is late. "30+ DPD" is the share of accounts (or of outstanding balance) at least 30 days late. Buckets such as 1–30, 31–60, 61–90 and 90+ show whether accounts are drifting deeper into arrears or curing. A rising 1–30 bucket is an early warning; a rising 90+ bucket is a loss that has mostly already happened. For regulatory classifications such as non-performing assets, say you would use the lender's policy and the applicable RBI norms rather than a remembered rule of thumb.

Q2. Count-based or balance-based delinquency?

Both, because they answer different questions. Count-based says how many customers are struggling; balance-based says how much money is at risk. If balance-based delinquency rises faster than count-based, the problem is concentrated in larger loans.

Q3. What is a roll rate?

The share of accounts in one bucket this month that move to the next worse bucket next month — for example, 25% of 1–30 DPD accounts rolling to 31–60. Roll rates are used for short-term loss forecasting and for judging collections: if the 1–30 roll rate rises, early-stage collections may be weakening.

Round 2: the portfolio analysis question

Q4. "Delinquency on our personal-loan book rose from 3.1% to 4.0% this quarter. Is underwriting getting worse?"

Not necessarily, and this is the key trap. The overall rate mixes loans of different ages. A book that grew fast a year ago now has many loans reaching the age at which defaults typically peak, so overall delinquency rises even if every monthly batch performs the same.

The right tool is vintage analysis: group loans by origination month and compare delinquency at the same months-on-book. If the March vintage at six months on book is worse than the September vintage at six months, underwriting or the applicant mix changed. If vintages overlap at equal age, the rise is an ageing effect.

Q5. "The newest vintage looks excellent. Good news?"

Too early. At two months on book few borrowers have had time to miss enough payments. Compare only at equal months-on-book, and be wary of drawing conclusions from vintages younger than the typical time-to-default for the product.

Q6. "Approvals from one digital partner have twice the early delinquency. Credit or fraud?"

Look at first-payment default: borrowers who miss the very first instalment. A high first-payment default rate, especially with borrowers unreachable at the contact details given, suggests fraud or poor identity checks rather than genuine inability to pay. Genuine credit stress usually appears later, after some payments have been made. Checking device, address and employer data for repeated values across applications is the next step.

Round 3: the decision question

Q7. "We score applicants 300–900. Where should the approval cut-off be?"

A cut-off is a profit decision, not a model-accuracy decision. For each possible cut-off, estimate the approval rate, expected bad rate among approved applicants, and expected profit per applicant. The model below uses synthetic score bands and simple per-loan economics you would replace with the lender's own.

import pandas as pd

# Synthetic back-book: applicants per score band and observed bad rate
bands = pd.DataFrame({
    "band_min":   [300, 550, 600, 650, 700, 750, 800],
    "applicants": [800, 1200, 1500, 1800, 1700, 1400, 1600],
    "bad_rate":   [0.30, 0.16, 0.10, 0.06, 0.035, 0.02, 0.01],
})
profit_if_good = 6_000    # hypothetical net interest and fees over the loan life
loss_if_bad = 45_000      # hypothetical loss after recoveries

rows = []
for cut in bands.band_min:
    appr = bands[bands.band_min >= cut]
    approved = appr.applicants.sum()
    bads = (appr.applicants * appr.bad_rate).sum()
    profit = (approved - bads) * profit_if_good - bads * loss_if_bad
    rows.append({"cut_off": cut,
                 "approval_rate": approved / bands.applicants.sum(),
                 "bad_rate": bads / approved,
                 "profit_lakh": profit / 1e5})
print(pd.DataFrame(rows).round(3).to_string(index=False))

With these assumptions, profit peaks at the 600 cut-off, approving 80% of applicants. The break-even bad rate for a band is profit_if_good / (profit_if_good + loss_if_bad), about 11.8% here. The 550–599 band, at 16%, loses money on average, so approving it lowers profit; the 600–649 band, at 10%, is below break-even and still adds profit even though its bad rate looks high. That is the point to make aloud: a cut-off follows from the economics, not from how uncomfortable a bad rate feels.

What to add aloud: the back-book only shows performance for applicants who were approved in the past, so bad rates below the old cut-off are uncertain — this is the reject-inference problem. Suggest a small, controlled test approval of lower-score applicants to measure it, with a loss budget agreed in advance.

Q8. "Collections can only call 2,000 of 10,000 overdue customers this week. Who?"

Rank by expected recovery from calling: the probability that a call changes the outcome times the amount at stake. Customers likely to pay anyway (one day late, always pay by day five) and customers very unlikely to pay regardless gain least from a call. Those in between, with meaningful balances, gain most. If you have past data on called and uncalled customers, estimate the uplift from a call rather than just the probability of paying.

Q9. "A new model has a higher AUC. Replace the old one?"

Compare them at the cut-offs you would actually use: approval rate and bad rate at the operating point, and profit using the same economics as Q7. Check stability across recent months and segments, whether any input feature raises fairness or regulatory concerns, and whether the model can be explained to a credit committee. AUC is a ranking summary across all thresholds; a lender operates at one.

How to prepare

Build the Q7 table for a product you know — a credit card, a two-wheeler loan — with your own assumed economics. Then change the loss-if-bad figure and see how far the optimal cut-off moves. The ability to explain that sensitivity is what interviewers in lending look for.

Metric choice from business costs, as in Q9, is covered further in the model evaluation metrics bank. Credit scoring, vintages and cost-based thresholds are worked through with synthetic lending data in the Data Science programme.