Build and evaluate donation propensity model
Company: Capital One
Role: Business Analyst
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
You need a model to maximize expected net revenue from solicitations. Costs: online reach costs $1 per person; gala attendance costs $100 per attendee plus a $20,000 fixed venue cost. Outcomes: probability of donating and donation amount. Tasks:
1) Propose a two-stage model (conversion and amount) or a direct expected-revenue model. Specify features, handling of sparse/categorical variables, leakage risks, and treatment of highly skewed amounts (e.g., log-transform, zero-truncated models).
2) Define the profit-optimized thresholding policy for online (solicit vs suppress) and the invite policy for the gala with capacity 100. Show how you convert predicted probabilities and amounts into expected marginal profit and use it to rank donors.
3) Explain calibration and why it matters for profit optimization; pick a method (e.g., Platt/Isotonic) and an evaluation plot.
4) Describe offline policy evaluation when historical assignments were non-random. Include inverse propensity weighting or doubly robust estimators and the assumptions required.
5) If you instead model uplift (treatment effect) for the online campaign, define the objective, appropriate metrics (e.g., Qini/uplift AUC), and how you would guard against targeting bias that starves low-budget donors of engagement.
Overview: Build and evaluate a donation propensity model to choose whom to solicit online or invite to a gala for maximum expected net revenue. Account for campaign costs, a 100-person gala limit, prediction calibration, biased historical assignments, and fair targeting when modeling uplift.
Read the full Capital One Business Analyst interview experience this question came from
