Time series and forecasting · insight-0235

Test one global model across many related series before maintaining thousands of local ones

Situation
A team fits and monitors a separate ARIMA for each of 4,000 store-SKU series.
Action
Benchmark a global model trained on all series (with lags, calendar features and series identifiers or scale features) against the local models in the same rolling-origin backtest, by segment.
Reason
Global models are not more restrictive than local ones, keep constant complexity as the number of series grows, and have performed well even on heterogeneous collections.
Boundary
Large panels: retail, routes, sensors, branches.

Symptom or risk

Local models on short noisy series overfit, maintenance cost grows with every series, and new items have no model at all.

A handful of long, high-value series with distinct dynamics may still deserve local models.

A global model is a single point of failure; one bad feature affects every series.

Leaking future information across series, for example using another series' same-day value.

Montero-Manso and Hyndman show global methods are not more restrictive than local ones and maintain constant complexity as the number of series grows.
Principles and algorithms for forecasting groups of time series: Locality and globality