Describe the bug
The current implementation of the multimodel statistics preprocessor was not designed to work for daily data, but it doesn't raise an error either. This leads to misleading results, because prior to the statistics calculations, all time points are reset to the first of the month:

This error has propagated into the reference datasets for the multimodel-statistics regression tests. While the sample data have 62 time points (days), the reference statistics cubes only have 2 (the first days of the 2 months).

This issue will be solved in #685, but until then, daily multimodel-statistics output should not be trusted.