Stripe wants a country×industry daily GMV forecast for the next 90 days (2025-09-01 to 2025-11-29) using 3+ years of history. You have features: day-of-week, country holidays, marketing_spend_usd, avg_risk_score, FX rates to USD, CPI, and known product launch flags. Design an end-to-end, hierarchical solution: - Modeling: Compare ETS/Prophet-like additive seasonality vs gradient-boosted trees on TS features vs global RNN/Temporal Fusion Transformer. Pick one primary approach and specify how you’ll reconcile segment forecasts to the country and global totals (e.g., MinT, BU, TOPDOWN). Provide concrete formulas or references for reconciliation and why they fit Stripe’s cross-sectional structure. - Cross-validation: Specify rolling-origin CV with initial window, step size, and number of folds; include leakage-avoidant feature construction. Define the primary metric as wMAPE weighted by segment GMV; justify choice over RMSE/MAPE/Pinball loss. - Intermittent/sparse series: Propose a method (e.g., Croston/TSB, zero-inflated models) and how you’ll blend it with the main model via meta-learning. - Cold-start segments: Outline partial pooling or hierarchical Bayesian shrinkage across industries within a country; define priors and hyperparameters. - Exogenous regressors: Which to include, how to lag/transform FX, CPI, and marketing; how to handle non-stationarity and scale. - Outliers/regime shifts: Detect and treat events (e.g., policy change on 2024-07-01) using robust loss or event dummies; explain your decision rules. - Uncertainty: Produce calibrated 90% prediction intervals (conformal, quantile regression, or simulation); describe calibration diagnostics. - Monitoring/retraining: Define drift tests, alert thresholds, retrain cadence, and rollback criteria. - Limitations: List 3 failure modes and mitigations. Answer with a specific pipeline (data prep steps, model classes, key hyperparameters, reconciliation method, and evaluation design).

This question evaluates skills in hierarchical time-series forecasting, covering model selection and reconciliation, cross-validation design, intermittent and cold-start series handling, exogenous regressor treatment, uncertainty quantification, and operational monitoring within the Machine Learning/time-series forecasting domain.

How do I approach Machine Learning interview questions?

Machine Learning questions require understanding of core concepts and practice. PracHub provides solutions with explanations to help you master machine learning interviews.

What difficulty level is this interview question?

This is a medium difficulty Machine Learning question, commonly asked during Onsite rounds at Stripe.

What role is this question designed for?

This question is commonly asked for Data Scientist candidates at Stripe during technical interviews.

Design a hierarchical forecast for transactions

Stripe wants a country×industry daily GMV forecast for the next 90 days (2025-09-01 to 2025-11-29) using 3+ years of history. You have features: day-of-week, country holidays, marketing_spend_usd, avg_risk_score, FX rates to USD, CPI, and known product launch flags. Design an end-to-end, hierarchical solution:

Modeling: Compare ETS/Prophet-like additive seasonality vs gradient-boosted trees on TS features vs global RNN/Temporal Fusion Transformer. Pick one primary approach and specify how you’ll reconcile segment forecasts to the country and global totals (e.g., MinT, BU, TOPDOWN). Provide concrete formulas or references for reconciliation and why they fit Stripe’s cross-sectional structure.
Cross-validation: Specify rolling-origin CV with initial window, step size, and number of folds; include leakage-avoidant feature construction. Define the primary metric as wMAPE weighted by segment GMV; justify choice over RMSE/MAPE/Pinball loss.
Intermittent/sparse series: Propose a method (e.g., Croston/TSB, zero-inflated models) and how you’ll blend it with the main model via meta-learning.
Cold-start segments: Outline partial pooling or hierarchical Bayesian shrinkage across industries within a country; define priors and hyperparameters.
Exogenous regressors: Which to include, how to lag/transform FX, CPI, and marketing; how to handle non-stationarity and scale.
Outliers/regime shifts: Detect and treat events (e.g., policy change on 2024-07-01) using robust loss or event dummies; explain your decision rules.
Uncertainty: Produce calibrated 90% prediction intervals (conformal, quantile regression, or simulation); describe calibration diagnostics.
Monitoring/retraining: Define drift tests, alert thresholds, retrain cadence, and rollback criteria.
Limitations: List 3 failure modes and mitigations. Answer with a specific pipeline (data prep steps, model classes, key hyperparameters, reconciliation method, and evaluation design).

Modeling: Compare ETS/Prophet-like additive seasonality vs gradient-boosted trees on TS features vs global RNN/Temporal Fusion Transformer. Pick one primary approach and specify how you’ll reconcile segment forecasts to the country and global totals (e.g., MinT, BU, TOPDOWN). Provide concrete formulas or references for reconciliation and why they fit Stripe’s cross-sectional structure.
Cross-validation: Specify rolling-origin CV with initial window, step size, and number of folds; include leakage-avoidant feature construction. Define the primary metric as wMAPE weighted by segment GMV; justify choice over RMSE/MAPE/Pinball loss.
Intermittent/sparse series: Propose a method (e.g., Croston/TSB, zero-inflated models) and how you’ll blend it with the main model via meta-learning.
Cold-start segments: Outline partial pooling or hierarchical Bayesian shrinkage across industries within a country; define priors and hyperparameters.
Exogenous regressors: Which to include, how to lag/transform FX, CPI, and marketing; how to handle non-stationarity and scale.
Outliers/regime shifts: Detect and treat events (e.g., policy change on 2024-07-01) using robust loss or event dummies; explain your decision rules.
Uncertainty: Produce calibrated 90% prediction intervals (conformal, quantile regression, or simulation); describe calibration diagnostics.
Monitoring/retraining: Define drift tests, alert thresholds, retrain cadence, and rollback criteria.
Limitations: List 3 failure modes and mitigations. Answer with a specific pipeline (data prep steps, model classes, key hyperparameters, reconciliation method, and evaluation design).

Design a hierarchical forecast for transactions

Quick Overview

Design a hierarchical forecast for transactions

Write your answer

Design a hierarchical forecast for transactions

Quick Overview

Design a hierarchical forecast for transactions

Write your answer