Choosing Maximum Lag for Granger Causality Tests

Summary

The document asks how to choose the maximum lag in Granger causality testing on several daily time series covering roughly three years. The researcher wants to identify both whether predictive relationships may exist and the delay at which they appear. They observe that changing the maximum lag across a range of values produces substantially different results, and wonder whether testing more lags makes a causality finding stronger or instead helps locate the relevant delay.

No response, method, or test results are included, so the document does not resolve lag selection or validate the assumption that a larger maximum lag strengthens evidence. It is best read as a methodological question about specifying lag length and interpreting sensitivity in a time-series test. The text gives the approximate sample size and candidate delays as context, but supplies no diagnostics, model-selection criterion, or caveats beyond the observed variation in outcomes.

Key ideas

• The question concerns maximum lag selection for Granger causality tests on daily series.
• The stated objective is to assess possible predictive links and their delay.
• Test outcomes reportedly change when the maximum lag changes.
• The document does not explain a selection method or provide test results.
• A larger lag limit is raised as a hypothesis, not established as stronger evidence.

Tags

# Selecting the correct MaxLag for Granger's Causality


# Selecting the correct MaxLag for Granger's Causality












I have a developer who has created a python script to determine the granger's causality of several datasets that are approximately 3 years worth of daily data (approx 1100 data points for each time series). The script seems to run well but we are not sure what MaxLag we should choose. Our goal is to determine possible causalities AND to determine the lag time in the causality (1 day, 2 days, 7 days, 14 days, etc). Obviously, when we change the maxlag number from 1 to 15 we get very different numbers.

It is my understanding that the higher the MaxLag the more "analyses" are done on the time series which results in the high MaxLag numbers providing stronger causality results. That seemingly would be very helpful but only if we know what the actual "lag" is for the causality.

Appreciate any insights someone can provide on this.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.