Interpreteing the outputs of the Decompose Signal component

I’m having a bit of trouble understanding the visual outputs of the Decompose Signal component. I’m relatively new to time series analysis so I might just be confused, but I want to make sure that the outputs are indeed what they say they are.

The four signal plots look like this, using the sample data in the Accessing Transforming and Modelling Time Series workflow.
image

Signal and Trend Removed make sense – one is the raw data, the other is the difference from the trend line. Except that when you dig into the component, the ‘Trend Removed’ plot actually shows the ‘Seasonality 1’ values, which are the detrended values minus the seasonality-removed values (which I think would represent the ‘seasonality component’ of the time series?).

I would expect ‘Seasonality 1 Differenced’ to show the result of subtracting the lagged values at the first peak of the ‘Trend Removed ACF’ line. Except it actually represents ‘Seasonality 2’, which in turn is the ‘Noise’ of the first seasonality lag calculation minus itself (because there was no second seasonality lag applied).

I would then expect ‘Seasonality 2 Differenced’ to show the result of applying the second seasonality lag, if one was detected, but instead it maps to ‘Noise’, which is the result of applying the first ACF lag.

If nothing else, some further documentation in the component to explain the logic of these plots and labels might help to reduce confusion. I know there are at least two KNIME blog posts discussing this component, but neither of them go to this level of detail. I’d appreciate any clarification you can offer.

Ok I might be getting closer. So if ‘Trend Removed’ is acutally the first seasonality component, and there is no second seasonality component, then ‘Seasonality 1 Differenced’ is supposed to show that there is no seasonality left to model once the trend and first seasonality are accounted for?

Whereas if there was a second seasonality component, ‘Seasonality 1 Differenced’ would show the residuals that remain after only the first seasonality component is accounted for? And ‘Seasonality 2 Differenced’ would show what was left once both were accounted for?

Hi @AngusVeitch1,

I’ve just noticed this thread and took a look into the component.
It appears to be plotting incorrect columns into those charts. I’ll look to update this component soon.

You’re correct about what it is showing:
*Signal is correct
*Trend removed is correct
*Seasonality 1 is showing the flat line because there is no seasonality (this should be fixed to show trend removed minus this seasonality)
*Seasonality 2 is showing the noise (although this should be the same as trend removed for this time series, I’m also looking into this)

In the mean time please check out this component:

2 Likes

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.