QR Data Analysis Prediction Case

Open-ended data-analysis round: define features, targets, models, and evaluation for a prediction problem such as housing prices, CitiBike demand, rental price, stock price from search/social data, or alpha testing. Requirements You receive a prediction problem and must reason through the full modeling plan. Reported domains include: Predict housing prices from historical data. Predict CitiBike demand or station inventory at a specific location. Predict rental price. Use six years of Google search data with timestamp, device id, location, query, and URL to predict a company's stock price at arbitrary times. Use Twitter / social data to predict stock price. Test whether a momentum alpha works. Interviewers probe: Feature and outcome definition. Model choice: linear regression, random forest, nonlinear alternatives, and why each may work. Missing values, outliers, data cleaning, and special shocks such as COVID. Nonlinear relationships, heteroskedasticity, multicollinearity, seasonality, data drift, and variance drift. Data leakage and overfitting. Train / validation / test split. Evaluation metrics such as MSE, MAE, accuracy, ROC-AUC, or domain-specific performance metrics. Notes This is not a coding-heavy round. It is a structured statistical reasoning interview. Candidates who only give high-level ML names struggle; the interviewer asks for assumptions, failure modes, and exact evaluation logic. Several candidates observe that traditional statistics and linear-model reasoning are valued more than generic ML framing. Preparation Prepare a reusable case framework: target, features, leakage risks, model family, validation split, metric, drift monitoring. Drill linear-model diagnostics: heteroskedasticity, multicollinearity, omitted variables, and residual checks. Practice explaining how shocks such as COVID should affect train/test split and feature treatment.