What data and approach for a simple 30-day DAU forecast?

WHAT THIS TESTS: This question tests whether you can translate unstructured behavioral data into a structured forecasting pipeline and resist the urge to over-engineer a beginner-level problem. The interviewer cares about your understanding of sessionization, privacy-aware data handling, temporal feature extraction, and choosing an interpretable baseline model before considering complexity. They also want to see that you connect the forecast to business actions like content scheduling or ad inventory.

A GOOD ANSWER COVERS: First, data sources and privacy. You need time-stamped event logs such as clicks, scrolls, page loads, video plays, likes, and comments. From these raw logs you derive sessions by grouping events separated by at least 30 minutes of inactivity. User identifiers should be hashed or removed to adhere to GDPR and CCPA, retaining only necessary attributes. Second, feature engineering. Build session-level and user-history-level features including session start time, day of week, and hour. Third, modeling approach. Treat the daily outcome as a continuous regression problem using historical daily aggregates as a baseline. The reference framework notes regression or time-to-event modeling for future interval totals, and you should evaluate with MAE, RMSE, or R squared. Fourth, deployment and validation. Start with batch inference rather than online prediction, and plan an A/B test to measure whether the forecast improves business outcomes like ad placement or resource allocation.

COMMON WRONG ANSWERS: Proposing deep learning or complex neural networks without first establishing a simple baseline. Ignoring sessionization and using raw ungrouped events as if they were independent observations. Requesting personally identifiable information without mentioning hashing, consent, or regulatory compliance. Failing to name evaluation metrics or to explain how the forecast would be validated in production. Suggesting real-time online inference for a 30-day lookahead when batch forecasting is more appropriate.

LIKELY FOLLOW-UPS: How would you handle anomalous days such as holidays or viral content spikes? Would you use online or batch prediction, and what latency constraints actually matter for a 30-day forecast? How would you check that the model is fair across different user cohorts or regions? What would you do if the 30-minute sessionization rule stopped reflecting true user behavior after a product redesign?

ONE CONCRETE EXAMPLE: A streaming service wants to predict daily active users for the next month. You extract backend event logs containing video plays and clicks, then group events into sessions using the 30-minute inactivity rule. You hash user IDs to preserve privacy and count distinct users per day to create the target series. You build a regression baseline using day-of-week, hour of day, and the prior seven days of active users as features. You evaluate the model on a holdout period using RMSE and deploy it as a daily batch job. The forecasts feed into ad inventory planning and are validated through an A/B test that compares predicted versus actual inventory fill rates.