Shopify · ML System Design

ML Modeling Round (Forecasting / Targeting / Fraud)

TrueInterview

August 21, 2026 · 1 min read

Requirements

You receive a business ML problem and must design the model itself, not the surrounding system. You should own the entire modeling story:

Problem framing — classification, regression, or ranking; what precisely constitutes the label.
Data & features — which data you would gather, plus feature engineering and selection.
Algorithm choice — candidate models with their trade-offs for this specific problem.
Evaluation — offline metrics, and how you define and gauge model success.
Deployment & monitoring — what you would track after launch and how you would improve the model.

Common prompts in this round:

Predict package/shipping delivery time (regression) — appears most often.
Decide whether to offer a Shopify loan to a user / a basic loan-recommendation model (classification).
Fraud detection — emphasis on features and model selection.
Revenue forecasting (finance organization).
Product ranking — standard classification or ranking model.

Notes

These are the everyday questions a working MLE tackles; depth is everything — name concrete features, justify your algorithm choice, and connect metrics to business outcomes. Staying too high-level is a typical failure mode.
Interviewer-fit risk is real: candidates report getting a prompt far outside their specialty (e.g., a recommendation/search person asked to do revenue forecasting), or interviewers with shallow ML knowledge ("no labeled data, how would you collect it?"). Stay structured and steer the conversation.
This is separate from the ML system-design round, which focuses on serving architecture; this one stays on modeling decisions, though some monitoring and serving questions still arise.

Preparation

Build a reusable end-to-end modeling template (framing → data/labels → features → model + trade-offs → offline/online eval → deployment → monitoring → iteration) and practice applying it cold to an unfamiliar prompt.
Prepare delivery-time/ETA regression, loan/fraud classification, and a ranking problem specifically, including the features and evaluation metrics you would use for each.
Rehearse handling weak interviewer prompts (missing labels, vague success criteria) by proposing reasonable assumptions and stating them aloud.

Requirements

You receive a business ML problem and must design the model itself, not the surrounding system. You should own the entire modeling story:

Problem framing — classification, regression, or ranking; what precisely constitutes the label.

Data & features — which data you would gather, plus feature engineering and selection.

Algorithm choice — candidate models with their trade-offs for this specific problem.

Evaluation — offline metrics, and how you define and gauge model success.

Deployment & monitoring — what you would track after launch and how you would improve the model.

Common prompts in this round:

Predict package/shipping delivery time (regression) — appears most often.

Decide whether to offer a Shopify loan to a user / a basic loan-recommendation model (classification).

Fraud detection — emphasis on features and model selection.

Revenue forecasting (finance organization).

Product ranking — standard classification or ranking model.

Notes

These are the everyday questions a working MLE tackles; depth is everything — name concrete features, justify your algorithm choice, and connect metrics to business outcomes. Staying too high-level is a typical failure mode.

Interviewer-fit risk is real: candidates report getting a prompt far outside their specialty (e.g., a recommendation/search person asked to do revenue forecasting), or interviewers with shallow ML knowledge ("no labeled data, how would you collect it?"). Stay structured and steer the conversation.

This is separate from the ML system-design round, which focuses on serving architecture; this one stays on modeling decisions, though some monitoring and serving questions still arise.

Preparation

Build a reusable end-to-end modeling template (framing → data/labels → features → model + trade-offs → offline/online eval → deployment → monitoring → iteration) and practice applying it cold to an unfamiliar prompt.

Prepare delivery-time/ETA regression, loan/fraud classification, and a ranking problem specifically, including the features and evaluation metrics you would use for each.

Rehearse handling weak interviewer prompts (missing labels, vague success criteria) by proposing reasonable assumptions and stating them aloud.