Time Series

How Automated Time-Series Analysis Works

Decomposition, detection and the order in which a machine should ask the questions.

Elizabeth Sramek Published 28 July 2026 1 MIN READ

Key takeaways

– Order matters: check the index, then level shifts, then seasonality, then trend.
– Seasonality detected before changepoints will absorb the break and hide it.
– Automatic frequency detection fails on irregular and gappy series.
– Every conclusion should come with the window it was computed over.

Start with the index

Irregular spacing, duplicated timestamps, timezone drift and daylight-saving boundaries produce spurious structure that no amount of modelling recovers from. This check is unglamorous and catches more real errors than the modelling that follows.

Then structural breaks

If the level or the variance changes partway through, everything estimated across the break is a blend of two regimes. Seasonal estimates in particular will absorb a step change and report it as a cycle.

Then seasonality

A candidate period comes from the domain, not from the spectrum alone. Weekly and annual periods are usually real; a detected period of 37 days usually is not.

Then trend

With breaks and cycles accounted for, the remaining slow movement is the trend. Reported without the other two, a trend estimate is an artefact of the window chosen.

Then criticise

Simulate from the fitted model and compare. Systematic divergence in a particular period is more informative than a global goodness-of-fit number — see the unemployment analysis.

Written by

Elizabeth Sramek

All articles