Build a time-series forecasting model
Overview
This question evaluates time-series forecasting competencies, covering preprocessing, handling covariates and temporal issues, selection between classical and deep-learning models, multi-step forecasting strategies, evaluation metrics, and production concerns like drift detection and retraining.
Jane Street
Aug 1, 2025, 12:00 AM
Machine Learning Engineer
Technical Screen
Machine Learning
0
0
Forecast the Next H Steps for Time Series
Context
You are given one or more time series with timestamps and numeric targets (e.g., demand, returns, sensor values). Your task is to forecast the next H time steps. Some covariates may be available, including both observed-past features and known-future features (e.g., holidays, schedules, prices).
Requirements
-
Preprocessing
- Resampling and alignment across all series and covariates
- Timezone and DST handling
- Missing-data imputation strategy (short vs. long gaps)
- Outlier detection and treatment strategy
- Detrending and seasonality decomposition
- Scaling/normalization
-
Modeling (propose at least two approaches)
- One classical (e.g., ARIMA/SARIMAX/ETS)
- One deep learning (e.g., LSTM/TCN/Transformer)
- Justify: loss choice (MAE/RMSE/Huber/Quantile), multi-step strategy (recursive/direct/multi-output), and how to incorporate covariates (holidays, known-future features)
-
Evaluation (non-leaky)
- Walk-forward validation/backtesting protocol
- Metrics: MAE, RMSE, sMAPE/MAPE, and quantile loss for P50/P90
-
Monitoring and maintenance
- How to detect concept drift
- How to recalibrate or retrain
-
Implementation
- Provide concise pseudocode for training and inference
Solution
Login required
Show