Suppose you have developed a model to predict whether a customer will make a purchase based on their browsing behavior on an e-commerce site. You train your model using historical data collected over the past year, assuming that the distribution of customer behavior will remain the same in the future. However, after deploying the model, you notice a sudden drop in performance. This coincides with a major site redesign and a new marketing campaign, both of which have changed how customers interact with the website. As a result, the distribution of input features — such as time spent on pages, click patterns, and product categories viewed — has shifted compared to the training data. Your model, which performed well under the original data distribution, now makes less accurate predictions because it is encountering patterns it has not seen before.

When the IID assumption fails, standard evaluation metrics such as accuracy, precision, or recall can no longer be trusted as indicators of real-world performance. In the scenario above, the test set — drawn from the original data distribution — no longer reflects the conditions the model faces after deployment. As a result, metrics calculated on this set may suggest the model is highly effective, while actual performance on new data is much worse. This disconnect can cause you to overlook issues, deploy unreliable models, and make poor decisions based on misleading evaluation results.