Feature Scaling and Normalization

You deploy a fraud detection model for American Express and observe a consistent failure pattern: the model flags almost exclusively high-dollar transactions as fraudulent, while many low-amount transactions that are later confirmed as fraud go undetected. Despite this, overall accuracy during offline evaluation appeared strong.

The model was trained on features including transaction amount, customer age, account tenure, number of transactions per day, and several binary behavioral indicators. No additional preprocessing or transformations were applied to the raw feature values.

How would you investigate the cause of this behavior? What aspects of the feature space and model behavior would you examine first, and how would that investigation guide your next steps?