GCP Professional ML Engineer Practice Question 11: BigQuery ML Demand Forecasting
Google Cloud Professional Machine Learning Engineer · Test 2 · question 1 of 10
Three years of daily unit-sales history for 2,000 SKUs already sit in a BigQuery table. Supply chain analysts, who write SQL every day but have never used Python, need a 60-day demand forecast per SKU refreshed weekly. Moving the data out of BigQuery would require a new security review, so the team wants the forecast produced where the data already lives.
Which approach meets these requirements with the least effort?
Show answer and explanation
Answer:
BigQuery ML lets analysts create and query models with ordinary SQL, and ARIMA_PLUS is its univariate time series model type: CREATE MODEL ... OPTIONS(model_type='ARIMA_PLUS', ...) trains on the historical column, and ML.FORECAST returns the future values. The modeling pipeline handles irregular intervals, missing values, outliers, step changes, holiday effects and seasonality automatically, and it can fit many time series in one statement, so 2,000 SKUs are covered by a single query. Nothing leaves BigQuery and nobody writes Python (option C).
- A is wrong because it moves the data out of BigQuery to Cloud Storage, which is exactly the security review the team is trying to avoid, and it adds an export-and-import step for every weekly refresh. - B is wrong because it makes SQL-only analysts maintain Python code, and looping a per-SKU statsmodels fit over 2,000 series in a notebook is far more work than one SQL statement. - D is wrong because a general-purpose large language model is not a forecasting model. It gives no confidence intervals, no reproducible fit, and no reliable numeric extrapolation over thousands of series.