A production model's accuracy falls over three months even though monitoring shows the input feature distributions have not moved. What is the most likely explanation?
Show answer and explanation
Answer: C
If inputs look statistically the same but outcomes have changed, the mapping the model learned no longer holds. That is concept drift: for example, a shift in policy or customer behaviour alters which feature patterns lead to the label. Data drift is ruled out by the monitoring result, and training-time problems such as too few epochs or underfitting would have shown up as poor scores from the start rather than after months of stable operation.
Why the other options are wrong
- A. Data drift would appear as changed feature distributions, which the monitoring explicitly rules out.
- B. Epoch count is fixed at training time and cannot explain a decline that begins later.
- D. Underfitting is visible immediately as weak training and validation scores, not as a delayed production drop.
Exam tip. Ask two questions: did the inputs move, and did the input-to-output relationship move?