Automated Time-Series Analysis
Trend, seasonality, changepoints, anomalies and honest forecast intervals.
Time series are where automated analysis is most useful and most easily fooled. The structures are limited enough to enumerate — level, trend, seasonality, changepoints, noise — and the failure modes are well documented.
The four questions
Is there a trend, and is it real or an artefact of the window? Is there a cycle, and is its amplitude stable? Did the behaviour change at a point in time? How much of what remains is noise?
Uncertainty is the deliverable
A forecast without an interval is a guess with better typography. The interval is also a diagnostic: if it is implausibly narrow, the model has assumed the future resembles the past more closely than the data supports — see the unemployment analysis.
In this section
Time Series
Trend vs Seasonality vs Change Points
How to distinguish trend, seasonality and structural breaks, and why confusing them produces confident nonsense.
How Automated Time-Series Analysis Works
How automated time-series systems detect trend, seasonality and changepoints, and how to judge their output.