Build a time-series forecasting model
Company: Jane Street
Role: Machine Learning Engineer
Category: Machine Learning
Difficulty: hard
Interview Round: Technical Screen
You are given a univariate or multivariate time-series dataset with timestamps and numeric targets. Predict the next H steps. Describe your preprocessing: resampling and alignment, timezone handling, missing-data imputation, outlier treatment, detrending/seasonality decomposition, and scaling. Propose at least two modeling approaches—one classical (e.g., ARIMA/ETS) and one deep learning (e.g., LSTM/TCN/Transformer)—and justify your choice of loss (MAE/RMSE/Huber/Quantile), multi-step strategy (recursive/direct/multi-output), and how to incorporate covariates such as holidays and known future features. Define a non-leaky evaluation protocol with walk-forward validation/backtesting and appropriate metrics (MAE, RMSE, sMAPE/MAPE, quantile loss for P50/P
90). Explain how you would detect concept drift and recalibrate or retrain. Provide concise pseudocode for training and inference.
Overview: Build a time-series forecasting model evaluates core ML concepts, assumptions, math intuition, training/evaluation trade-offs, and practical failure modes in a realistic interview setting. A strong answer states assumptions, handles edge cases, explains trade-offs, and shows how to validate the result clearly.