The filter control file (which must be named kalman.prm) specifies the covariance matrix for the driving noise and the initial covariance of the state vector. The current implementation of the filter assumes that the driving noise directly impacts the levels. The specification of the noise uses the following format:
Lev_1/.34/1000
Lev_2/LEV_2_VAR/LEV_2_IVAR
Lev_1 | Lev_2/ LEV_1_2_COVAR
In the first line, the noise influencing Lev_1 has a variance of 0.34, and the initial variance of Lev_1 is 1000. In the second line, the noise influencing Lev_2 has a variance given by the model variable LEV_2_VAR, and the initial variance of Lev_1 is given by LEV_2_IVAR.In the third line, Lev_1 and Lev_2 have a covariance given by the model variable LEV_1_2_COVAR and no initial covariance is specified, so it will be assumed to be 0. Model variables are typically constants and are valid search parameters.
The initial value of the state covariance is a measure of how uncertain the initial states computed by the model are. If you leave this blank Vensim will use 1E6 on the diagonal, and 0 off. Generally it is better to enter this explicitly, since with a high variance, the first thing the filter will do is move the model dramatically toward the data. If you have some confidence the initial states are within 20% or 30%, you should use this (for example take (.3*initial value)2).
