Every floor · The machinery

Data Scientist (Banking)

Turns the bank's data into decisions — pricing, targeting, fraud detection, credit scoring, early warning.

BY MICHAEL SHANG · SIXTEEN YEARS IN COMMERCIAL BANKING · UPDATED 2026-10-11

// 01 — WHAT A DATA SCIENTIST DOES

The modelling is the visible part; the work is data quality, definitions that differ between systems, and the long argument about whether a result is real. Anything that touches a customer outcome will be validated independently and questioned by a regulator, which changes how you build.

// 02 — WHO ARRIVES HERE

Statistics, mathematics, engineering and economics backgrounds, and analysts moving up from reporting seats.

// 03 — WHAT THE INTERVIEW ASKS

These are the shapes the round takes, not a question bank — banks phrase them differently and the wording is never the point.

01How would you build an early-warning signal for a lending book?
02Your model is accurate and unfair to a group. What do you do?
03How do you know a result is real and not the data?
04Explain your model to someone who will act on it.
05When would you not use a model at all?
06How would you monitor it after it ships?

What the panel is listening for. Scepticism about your own output, and whether you can explain a model to the person whose decision it will change.

The finance, technology and operations quiz has no question for this seat yet. Its five were asked for the software engineer, financial accountant, operations analyst, procurement manager and people partner seats, in the same block, and the temperament they score is the one this panel scores too.

Take the quiz →

// 04 — ONE QUESTION, WORKED

QHow would you build an early-warning signal for a lending book?

A strong answer starts with the outcome and the horizon — a downgrade or a default within a set period, not “risk” — and with what a portfolio manager would do differently on a warning, because a signal nobody acts on is a report. It reaches for behaviour before financials: account conduct, late management accounts, limit utilisation, arrears on the small facilities, since accounts arrive months late. It names the leakage risk and the class imbalance, how the signal is validated out of time, how it is monitored after it ships, and how a desk overrides it.

A weak one names an algorithm and a set of financial ratios, and never says who will act on the output or how they would know it worked.

The seat is judged on scepticism about your own output and on explaining a model to the person whose decision it will change. Early-warning and scoring models are trained on what happened to borrowers; the drills show what a credit desk sees before it happens, and the portfolio review — where the signals move before the numbers — is the case closest to the modelling problem.

START HERE · Issue 09 — the Tuesday list

THE IN-TRAY · ONE MINUTE