I am trying to forecast a target time series in AutoGluon using global and global+covariates models.

The target time series ranges monthly from 1986-01-01 to 2024-05-01.
I have a collection of ~160 additional time series regressors I believe will inform the forecast of the target and want to include them in my AutoGluon model.

The issue lies in the additional regressors, which all vary greatly in date range due to availability, making them various degrees of censored.

For example, these are a few regressors I am working with (but there are many more spanning many different start and end dates):

As you can see, R1 and R2 are left-censored with respect to the target date range. Majority of my regressors are in this category. There is also a select few that are like R3 which are censored on both sides (but this is a minority within the overall collection).

Thus far I have tried doing a manual selection of regressors in 3 different date ranges 1986-01-01, 1997-01-01, and 2007-01-01 to 2022-01-01. Then I’ve run a model to forecast the target using the historic target data and these regressors and analyzed the mean evaluation of each forecast. This is a fine, but strenuous way of doing what I want especially considering the vast degrees of left and right centred data and it also doesn’t fully capture all of the regressors.

I am looking for advice on implementing a model that allows me to include as many regressors as possible, in the largest date range possible (1987-2024 would be ideal). I am preferably not looking to do any imputation/interpolation of these data points. What I am looking for is suggestions of creative ways to feed these regressors into AutoGluon, or suggestions on different time series forecasting packages that would allow this type of input.

Any advice is welcome, and please reach out if you need clarification, some of my terms may not be exactly right but hopefully you understand the gist. Thank you!