A retailer has three years of weekly sales history with yearly peaks around holidays and a gradual upward trend. The team needs a forecast of the next twelve weeks, produced entirely inside BigQuery. Which model should they train?
Show answer and explanation
Answer: A
ARIMA_PLUS is a BigQuery ML time series model that decomposes history into trend, seasonal, and holiday components and produces multi-step forecasts with prediction intervals. It trains from a time column and a value column, all in SQL without leaving BigQuery. KMEANS finds clusters, matrix factorization handles recommendations, and logistic regression predicts categories, so none of them produces a multi-week numeric forecast.
Why the other options are wrong
- B. KMEANS groups rows by feature similarity without a time axis. It can segment weeks but cannot project future sales values.
- C. Matrix factorization learns latent factors from user-item interaction matrices for recommendations. It has no notion of time-ordered sales and cannot forecast weekly totals.
- D. Logistic regression predicts a binary or categorical label. Weekly sales are a continuous quantity, so this model type does not match a numeric forecasting requirement.