I am working on demand models (specifically cancellations, renewals for finacial products) and I have access to about 5 years of data. I will be using a GLM for the model and need to select the time period to model on.
Is there a way to make a best selection of the time period?
Ways I have thought about are:
In general, the more data you have the better. So, if you are trying to forecast demand going forward, then it would likely make sense to model on the five years of data.
You mentioned that there is potentially a lag present in the data. If you are using GLM, and you are specifically looking to model the lag between two variables, then this is known as cross-correlation, and you should be using a cross-correlation function to determine if a lag exists between two variables.
Autocorrelation describes a situation where there is a lag within one variable itself. If there is a possibility that autocorrelation is present in your data, then you should test for this condition using the Durbin-Watson test, and depending on whether it exists and the nature of the autocorrelation, then these could indeed be modelled using lags of the dependent variable as regressors. You might find the information in this post more informative on this particular issue.
Ultimately, the suggested course of action will depend on the type of correlation (if any) that exists between your data, and the time period for which you are trying to forecast. However, if you have 5 years of data, then I see no reason why you shouldn't use all the data available to you.
