Design a robust traffic forecasting pipeline
Overview
This question evaluates a candidate's competency in end-to-end time-series forecasting pipeline design, covering data cleaning and missing-value handling, anomaly detection and intervention strategies, feature engineering, probabilistic modeling with Unobserved Components Models, rolling-origin backtesting, model comparison, and scaling for multiple related series. It is commonly asked to assess practical and conceptual understanding of Machine Learning and time-series forecasting — including model assumptions, uncertainty quantification, evaluation metrics for quantiles, and productionization considerations — and tests both conceptual understanding and practical application within the Machine Learning / Time-Series Forecasting domain.
Forecasting Daily Amazon Retail Traffic: End-to-End Design
You are given 5 years of daily Amazon retail site traffic counts. Design an end-to-end forecasting pipeline that produces 1-, 7-, and 28-day-ahead forecasts along with 10th/50th/90th percentile prediction intervals.
Specify and justify the following:
(a) Data cleaning and missing-value strategies
- How you would standardize the time index, handle bots/outages/duplicates, apply transformations, and impute or carry missingness into the model.
(b) Anomaly detection and treatment
- Methods to detect point outliers and regime shifts, and how you would downweight, cap, or model them (e.g., interventions).
(c) Feature engineering
- Calendar and event features (holidays, Prime Day, Black Friday/Cyber Monday), promotions/price indices, day-of-week/weekend effects, moving averages and lags, and any external drivers.
(d) Model choice: Unobserved Components Model (UCM)
- Define the UCM structure (trend, seasonality, regressors), describe estimation via Kalman filtering and smoothing, and list key hyperparameters (e.g., state variances, seasonal complexity).
(e) Rolling-origin backtesting
- Your scheme to pick training window lengths, forecast horizons, refit frequency, metrics (including for quantiles), and guardrails against leakage.
(f) UCM vs. SARIMA comparison
- Contrast assumptions, interpretability, support for exogenous regressors, handling missing data, multi-seasonality, and computational cost.
(g) Scaling to hundreds of related series
- How you would productionize and parallelize, share information across series, and when you would switch to alternative models (e.g., global probabilistic or deep-learning approaches).
Solution
Login required