How to validate production models?

You are interviewing for a fintech model-validation team that acts as a second line of defense for credit-risk and fraud models. A hiring manager asks: **How would you validate a machine learning model in production?** Assume the model is used for transaction fraud detection or credit decisioning, where false positives can block good users and false negatives can create financial loss and regulatory risk. Describe an end-to-end validation framework that covers: - the business objective and cost function - data quality checks, label definition, label delay, and leakage detection - train, validation, and test design for time-dependent data - how to assess conceptual soundness of the modeling approach - how to evaluate classification models under severe class imbalance - calibration, threshold setting, and decision-policy trade-offs - how to monitor the model after deployment for drift, calibration decay, and fairness issues - what documentation and governance a model-validation team should require - when a simple linear model may be preferred to a more complex model, and what assumptions linear regression relies on - how you would compare common classification models such as logistic regression, tree-based models, and ensemble methods in this setting