Automatic Statistician — statistical report

Unemployment Analysis

Structural change, trends and uncertainty over time — the series that shows why a model must be allowed to say that behaviour changed.

Monthly unemployment rate · n = 408

Download PDF All examples

Key takeaways

– The series contains at least one clear structural break; a stationary model fitted to the whole record is misleading.
– Before and after the break, both the level and the volatility differ.
– Trend estimates are highly sensitive to whether the break is modelled.
– Forecast intervals that ignore the possibility of further breaks are too narrow.

Executive summary

A changepoint model separates the record into regimes with different levels and different noise. Within each regime a smooth trend describes the movement adequately. Attempting to fit one trend across the break produces residuals that are strongly autocorrelated and an interval that understates real uncertainty.

Dataset

Monthly unemployment rate, 408 observations. Seasonally unadjusted values are used so that any seasonal component is estimated rather than assumed away.

Structural components

Component	Structure	Interpretation
1	CP	Regime change at the detected date
2	SE	Smooth trend within each regime
3	PER	Mild annual seasonality
4	WN	Regime-dependent noise

TABLE 1 — Discovered components

Model criticism and limitations

Changepoint location is itself uncertain, and reporting a single date overstates precision. The model cannot anticipate future breaks; a forecast from it is conditional on the current regime continuing, which is exactly the assumption most likely to fail.