You have time-stamped order-book observations containing buy and sell information, instrument, price, and quantity. Design a model and evaluation procedure to predict the return over the next five minutes. ### Constraints & Assumptions - Observations arrive at irregular event times and may include multiple instruments. - The target must be computable using information available after the five-minute horizon without leaking that future information into features. - Trading costs and class or target imbalance must be considered when judging usefulness. ### Clarifying Questions to Ask - Is the target a raw, log, or market-adjusted return? - Is the prediction made on every event or on a regular decision grid? - Are crossed books, canceled orders, and market halts represented? ```hint Split by time before tuning Random row splits leak adjacent market regimes and often place nearly identical book states in train and validation data. ``` ### What a Strong Answer Covers - Point-in-time labels, book reconstruction, feature windows, and leakage prevention. - Baselines, model choice, time-based validation, and realistic performance metrics. - Instrument and regime drift, missing data, execution costs, and monitoring. ### Follow-up Questions - How would overlapping five-minute labels affect validation uncertainty? - Which features capture order-book imbalance without using future updates? - How would you decide whether the model is economically useful?