Citadel·Data Scientist·Technical Phone Screen·Senior

Senior

Jul 2026

Summary

Citadel data scientist interview with a meaty ML system design question about housing market prediction. One question, lots of surface area, and they clearly wanted to see if you could hold the whole thing together without falling apart on any single piece.

Questions Asked (1)

You have monthly housing market time series data across regions (price indices, mortgage rates, inventory, days-on-market, macro indicators). Design a classification system to predict whether a buyer should purchase now or wait up to k months. Walk through label definition, preprocessing, feature engineering, train/val splits, model choices, evaluation, leakage prevention, concept drift handling, and how you'd surface results to users.

System DesignData ModelingTechnical Trade-offs

Author's notes

This question has so many sub-parts that I spent the first two minutes just trying to figure out where to start.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and constraints (e.g., definition of 'should purchase', cost of waiting, risk tolerance) to frame the problem as a binary classification with a time horizon. Then walk through the ML lifecycle systematically, emphasizing temporal data handling, leakage prevention, and evaluation tailored to the decision's asymmetric costs. Conclude with how you'd operationalize and communicate predictions to users.

Pro tip: Emphasize that in finance, the cost of a false positive (buying when you should wait) often far outweighs a false negative, so you'd optimize for precision or use a cost-sensitive metric rather than plain accuracy. Also, mention that you'd validate with a rolling-origin evaluation to mimic real-time deployment.

1. Define the Prediction Target and Business Constraints

Clarify what 'should purchase now' means: e.g., buying now is better if the expected net benefit (price appreciation minus costs) over the next k months exceeds waiting. Define the label as a binary outcome based on future data, and incorporate asymmetric costs (e.g., missing a good deal vs. buying at a peak).

2. Data Preparation and Feature Engineering

Handle missing values, align time series across regions, and create features like lagged indicators, rolling statistics, momentum, seasonality, and macro interactions. Ensure all features are computed using only past data to avoid leakage.

3. Temporal Validation and Model Selection

Use time-based splits (e.g., expanding window or rolling-origin) for training/validation. Choose models that handle temporal dependencies and non-stationarity, such as gradient boosting with time-series features, or sequence models (LSTM, Temporal Fusion Transformer). Consider ensembling.

4. Evaluation and Leakage Prevention

Evaluate with metrics aligned to business costs (e.g., precision-recall AUC, expected cost). Prevent leakage by ensuring no future information is used in features or labels, and by carefully handling target encoding and cross-validation.

5. Deployment, Monitoring, and User Communication

Deploy with a pipeline that updates monthly, monitor for concept drift via performance metrics and data drift detection, and retrain periodically. Surface results as a simple recommendation (Buy/Wait) with confidence and key drivers, tailored to user risk profiles.

Key Points to Mention

Label definition: use future price changes and costs to define a binary target, ensuring the horizon k is aligned with business needs.
Leakage prevention: strictly use only historical data for features and labels; avoid look-ahead bias in time series.
Temporal validation: use rolling-origin or expanding window splits instead of random splits to respect time order.
Concept drift: monitor for changes in market regimes and retrain models; consider online learning or frequent updates.
Model choice: gradient boosting (XGBoost, LightGBM) with time-series features or deep learning models for sequence data; handle class imbalance.
Evaluation: use cost-sensitive metrics (e.g., expected cost, precision at high recall) and backtest on historical periods.
User interface: provide clear recommendations with uncertainty estimates and explanations (e.g., SHAP values) to build trust.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.