Transform a time series for a supervised learning model?

WHAT THIS TESTS: This question tests your practical understanding of feature engineering for time series data. The interviewer wants to see if you can correctly reframe a sequential problem into a tabular format that standard supervised models expect (an X_train, y_train structure). It specifically probes your knowledge of feature creation and, most importantly, your awareness of time-based data leakage and the correct validation strategies to prevent it.

A GOOD ANSWER COVERS: A strong answer has three parts. First, explain the core concept: creating a supervised learning dataset where the target y is the value at time t, and the features X are derived from data at times < t. Second, detail the three main categories of features you would create. These are lag features (the value at t-1, t-7), window features (rolling aggregates like a 7-day moving average or 30-day max), and calendar features (day of week, month, is_holiday). Third, and most critically, specify the validation strategy. You must state that a simple random split is incorrect and that you would use a time-ordered split (e.g., train on 2021-2022, test on 2023) or a forward-chaining procedure like scikit-learn's TimeSeriesSplit.

COMMON WRONG ANSWERS: A candidate who only mentions lag features is giving a junior-level answer. The biggest red flag, however, is failing to mention the validation strategy. Suggesting standard k-fold cross-validation or a random train-test split is an immediate signal that the candidate doesn't understand the risk of data leakage in time series, where the model would be trained on future data to predict the past, leading to inflated performance metrics and a useless model in production.

LIKELY FOLLOW-UPS: Expect follow-up questions like: "How do you choose the right lag or window size?" (Answer: domain knowledge, autocorrelation plots, or treat it as a hyperparameter). "What if you have multiple related time series?" (Answer: use their lags and window features as input features for the main series). "How would this change for a multi-step forecast?" (Answer: adjust the target to be a future value, e.g., y is value(t+7)).

ONE CONCRETE EXAMPLE: To forecast daily user signups, for a given date t, the target y is signups(t). The feature matrix X would include: lag features like signups(t-1) and signups(t-7); window features like mean_signups(t-7 to t-1) and stddev_signups(t-30 to t-1); and calendar features like day_of_week(t) and is_holiday(t). We would train the model on data up to Dec 31, 2023, and test its performance on data from Jan 1, 2024, onwards.