NEWS

arima2 3.4.2.9000

Added Durbin-Levinson sampling scheme for randomly sampling ARMA coefficients. The algorithm is implement in the internal function .DLsample, which calls the C function ARIMA_transPars. The new sampling scheme can be used externally with the function sample_ARMA_coef by setting method = "DL", or internally as part of the multi-start algorithm by setting init_method = "DL" in the arima function.

Updated root sampling scheme. Previously, if $n$ roots were sampled, floor($n/2$) roots were sampled as conjugate pairs, and if $n$ is even the remaining root was sampled as a real number. This scheme did not allow the ability to sample any real roots (other than the final odd root), leading to the inability of sampling positive evenly numbered AR components. The updated root sampling scheme is the same, but allows for this possibility by randomly choosing some paired roots to be real rather than complex. The rate at which real roots are sampled is selected so that, roughly, the marginal probability that any AR/MA coefficient is positive is $0.5$.

Fixed minor bug in arima.R function that occured when include.mean = TRUE and some parameters are already fixed.

Fixed minor spelling errors in documentation.
Refactored sample_ARMA_coef.R. The new function is faster than the original, includes the capability to manually set the minimum acceptable distance between inverted roots, and is easier to read.
Package tests have been updated with the new version of the sample_ARMA_coef.R function.

Added max_inv_root and min_inv_root_dist arguments to arima:
- max_inv_root controls the maximum size of the inverted AR / MA polynomial roots. Defaults to 1.
- min_inv_root_dist controls the minimum allowed distance between AR and MA polynomial roots, in order to avoid nearly canceling roots. Defaults to 0.

Added a check to ensure a random restart is considered an improvement only if it returns a model that has a proper variance matrix for estimated coefficients.