When not to use a forecasting foundation model

Foundation models for forecasting are useful when you have many related series and little time to fit each one. They are a poor default when you have one short series, a known seasonal pattern, and a stakeholder who needs to see the equation. In that setting ETS, ARIMA, or a simple regression with calendar features is easier to audit and often as accurate.

Zero-shot models still need a clean history. Missing timestamps, mixed frequencies, and level shifts from a policy change will confuse Chronos as surely as they confuse ARIMA. The difference is that the foundation model will still emit a confident quantile band. A local statistical model at least forces you to look at residuals and to state the differencing and seasonality you chose.

Use a pretrained forecaster when the task is many series, irregular but related, and you want a first distribution quickly—then compare it with a baseline on a true hold-out. Use Chronos-2 when covariates and multivariate structure are real, not when you only have a univariate classroom example. Fine-tune only if you have enough in-domain series that a hold-out still remains after fitting.

The teaching point is the same as in medical imaging. A large pretrained network is a tool, not a proof. Report the baseline, the split, and what the forecast is not allowed to decide. A warehouse forecast is not a diagnosis, and a quantile is not a guarantee.