When RLM is performed on a time series where the number of observations, n, is approximately equal to the number of parameters, k, there is the potential for the MAD calculation to estimate a penalty of zero, resulting in a division by 0, and errors in the relevant BLAS routine.
The BLAS error typically looks like:
On entry to DGELSD parameter number 4 had an illegal value
For example, a time series y = [1000, 1100, 1050, 950, 1003], the observations of 1100, 1050, and 950 may receive a weight of 0. When calculating the median absolute deviation, the median of [1100, 0, 0, 0, 1003] is zero. In order to prevent the inevitable division by zero, we should ensure that this scaling variable is always at least as large as the machine epsilon value.
When RLM is performed on a time series where the number of observations,
n, is approximately equal to the number of parameters,k, there is the potential for the MAD calculation to estimate a penalty of zero, resulting in a division by 0, and errors in the relevant BLAS routine.The BLAS error typically looks like:
For example, a time series
y = [1000, 1100, 1050, 950, 1003], the observations of1100,1050, and950may receive a weight of 0. When calculating the median absolute deviation, the median of[1100, 0, 0, 0, 1003]is zero. In order to prevent the inevitable division by zero, we should ensure that this scaling variable is always at least as large as the machine epsilon value.