Machine Learning for Time Series with Python - Second Edition

6

Multivariate and Hierarchical Forecasting

In previous chapters, we focused on univariate forecasting: predicting one time series from its own historical data. In Chapter 3, we used ARIMA to forecast Google Trends data based solely on its past values. In Chapter 5, we built features from each M5 item-store pair’s sales history to predict its future sales. While we trained one model across all 30,490 series, we treated each item-store as conditionally independent, assuming one product’s sales don’t directly influence another’s orders.

That approach, known as univariate forecasting, is the foundation for everything we have done so far, but it relies on a simplifying assumption that breaks down in two scenarios: when variables actively influence ...