A retailer deployed a sales-forecasting model a year ago and has not changed its code or weights, yet its accuracy on live traffic has fallen steadily. The team wants to identify genuine reasons a stable, unchanged model degrades in production over time. Which two factors most plausibly explain this kind of gradual decline? Select TWO.
- AThe statistical distribution of incoming feature values has drifted away from the distribution the model was trained on. Correct
- BThe relationship between the input features and the target outcome has shifted, so the same inputs now map to different results. Correct
- CThe model's learned parameters slowly decay on their own each month while the endpoint is serving requests.
- DThe hardware accelerator running inference rounds floating-point numbers differently as the server ages physically.
- EThe training dataset stored in object storage shrinks automatically, leaving the model with less to predict from.
Why A is correct: Data drift means live inputs no longer resemble the training data, so a fixed model fits them less well and accuracy slips.
Why B is correct: Concept drift changes the underlying input-to-output mapping the model learned, degrading a model whose weights stay frozen.
Why C is wrong: Tempting because performance falls over time, but stored weights are static files and do not decay or change while serving.
Why D is wrong: Sounds technical, but inference arithmetic does not drift with hardware age, so this cannot explain a steady accuracy loss.
Why E is wrong: Misleading because data matters, yet the training set is not consulted at inference and stored data does not shrink by itself.