MLA-C01 Practice Question 74: DeepAR Related Time Series
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 8 · question 4 of 10
A home-goods retailer forecasts weekly demand for 3,000 products. The products follow shared seasonal patterns by category, but many have only a year or two of history. The current approach fits a separate ARIMA model to each product, and it forecasts short-history products especially poorly. The team wants one SageMaker AI built-in algorithm that learns from all products together and produces probabilistic forecasts.
Which algorithm should the team use?
Show answer and explanation
Answer:
DeepAR is SageMaker AI's forecasting algorithm for sets of related time series. Classical methods such as ARIMA fit one model per series. DeepAR trains a single model across the whole dataset, so each forecast draws on patterns learned from similar series, and AWS notes that it outperforms ARIMA and ETS when there are hundreds of related series. Categorical features let it learn how each group, here each product category, typically behaves, which particularly helps products with short histories. It produces probabilistic forecasts (option C).
- A is wrong because a linear regression on lags cannot capture the non-linear, shared seasonal behavior across thousands of series, and it does not produce probabilistic forecasts. - B is wrong because a 12-week window cannot see yearly seasonality, and a point regressor produces one number, not a forecast distribution. It is also not a purpose-built forecasting algorithm. - D is wrong because Random Cut Forest detects *anomalies* and gives anomaly scores. It does not forecast future values.
Reference: Use the SageMaker AI DeepAR forecasting algorithm