Time series stationarity and models for heteroskedasticity

Hi guys,

I am new at the community and I use KNIME for time series analysis about 2 weeks and I have few questions:

Is there a possibility to check if a timeseries itself is stationary except of “Analyze ARIMA Residuals”. I have following time series after removing the saisonality and I am not 100% sure if it is stationary:

The node Analyze ARIMA Residuals says for the first 10 lags it is stationary but the higher the lag the higher the autocorrelation.

And another question: Does KNIME have models for heteroskedasticity (like ARCH and GARCH)? I know this question came up already last year in July but maybe in the meantime there was a change.

Thanks in advance and if there are any unclear points, feel free to reach out.

Best regards,

Hi Franziska,

Thank you for the example! We don’t have another component for checking the stationarity. But your question whether the time series is stationary or not, I’ll check that!

And ARCH and GARCH models are not yet available in KNIME.


1 Like

Could you share

  • the granularity of the data and
  • the number of data points used for the model?

Hi Maarit,

thank your for your answer and efforts.

I used following dataset (only store 1):
Walmart Dataset (Retail) | Kaggle

It includes weekly sales and after removing the saisonality (with lag 52) 91 rows.

Hi Franziska,

below the answer from Prof. Daniele Tonini who is teaching the time series course at KNIME:

With seasonal series, typically you check the ACF function up to the lag corresponding to twice the seasonal period. So, for instance, with monthly data you check until 24 lags. Having weekly data you should check until 104/106 lags, but with only 91 obs you have a lot of instability of the auto correlation function at high lag values, just because there are few observations available to compute the covariance between Yt and Yt-k… so it’s quite normal to see that high variability of the AFC over there. If it’s not possible to collect more datapoints, I would consider to do the Ljung-Box test only up to 52 lags.

Thank you for the question, I also learned something! I hope this helps you!



Would differencing be an option? Then ust apply it e.g by using a lag column note and you would be sure (Please note I am not a data scientist so please correct me if I am wrong. Just trying to help here)

Yes, differencing should handle non-stationarity in data. However, if I read the answer correctly, @Franziska_W had already differenced the data at lag 52 to remove yearly seasonality.

Hi Maarit,

thank you so much for your help!


Yes, I already differenced the data. But anyway thank you for your help.

This topic was automatically closed 182 days after the last reply. New replies are no longer allowed.