You have observations containing timestamp, stock identifier or index membership, sentiment score, relevance score, and stock prices. Design a model to predict a future stock return over a specified horizon. Explain target construction, point-in-time feature generation, model baselines, time-based validation, and how sentiment relevance should affect the design. The answer should distinguish statistical prediction from a profitable trading decision. ### Constraints & Assumptions - News or sentiment records may arrive after their event time. - Multiple records can refer to the same stock and story. - Transaction costs and turnover matter in evaluation. ### Clarifying Questions to Ask - What horizon and return definition are required? - Is the relevance score calibrated across sources and time? - Do timestamps represent publication time, ingestion time, or both? ```hint Point-in-time availability is part of every feature A sentiment value is unusable at decision time if it was ingested later, even when its article has an earlier publication timestamp. ``` ### What a Strong Answer Covers - Leakage-safe labels and joins, deduplication, aggregation windows, and market baselines. - Chronological evaluation with embargoes or purging when labels overlap. - Predictive and cost-adjusted metrics, uncertainty, drift, and monitoring. ### Follow-up Questions - How would you model market-wide sentiment separately from stock-specific sentiment? - How would delayed sentiment data be backfilled without changing historical training truth? - What result would convince you not to deploy the model?