HierarchicalForecast: MinTrace(mint_shrink) lead to error due to missing time stamps in Y_fitted_df

[flxschrml]

When I try to implement Nixtla's hierarchicalforecast package with the MinTrace(mint_shrink), it is required to add a Y_fitted_df that contains fitted values for past time stamps for the different time series that I want to reconcile. However, since some of my time series have later starting dates than others, they also have different starting dates in Y_fitted_df. Based on my observation, there is no major issue with the reconciliation if starting dates within Y_fitted_df only vary by few time stamps (e.g. 3 months if the time series are structured on a monthly basis) but if for certain time series, there are more time stamps missing in Y_fitted_df (e.g. 36 months out of 60), I get an error message saying:

"Exception: min_trace (mint_shrink) needs covariance matrix to be positive definite."

However, this error only occurs for the method MinTrace(mint_shrink), not for MinTrace(ols). But the issue with MinTrace(ols) is that apparently, that in my case all test runs led to the same reconciled dataframes, no matter if I emitted 0, 3 or 36 data points for certain time series within Y_fitted_df.

Can anyone please tell me how Nixtla's hierarchicalforecast handles those cases with missing time stamps in Y_fitted_df?

Also, are there any best practices (using ols instead of mint_shrink? filling up missing time stamps with 0? ...?) when it comes to missing time stamps in Y_fitted_df?

Thank you in advance.