Looking at the arima.c code related to arima fitting I noticed that the code
is mainly a merge of:
- Gardner, G, Harvey, A. C. and Phillips, G. D. A. (1980) Algorithm AS154.
An algorithm for exact maximum likelihood estimation of
autoregressive-moving average models by means of Kalman filtering. Applied
Statistics 29, 311322.
- Jones, R. H. (1980) Maximum likelihood fitting of ARMA models to time
series with missing observations. Technometrics 20 389395.
The first is used to fit the initial P0 matrix, and the second to do the
forecasts.
The AS154 implementation of P0 computation is O(r^4/8) in memory
requirements, where r is roughly the period length.
This is the origin of the ugly:
src/library/stats/src/arima.c:838: if(r > 350) error(_("maximum
supported lag is 350"));
I noted on the same AS154 paper that the initial P0 verify this equation:
P0 = T P0 T' + R R'
So I modified the arima.c code to find iteratively the solution of this
equation (starting from P0 = I)
The resulting code finds a solution very similar to the one of the original
code in a fraction of the occupied memory and in a time that is similar for
small lags and faster for bigger lags (without the r<350 limit).
Here the modified code: https://gist.github.com/911292
The question is, there are theoretical guarantees that the iterative
solution is the right solution?
Some hints/directions/books?
Matteo Bertini
[[alternative HTML version deleted]]
______________________________________________
R-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-devel
