tsmarch 1.0.3

• Added support for joint ARMA(X)-GARCH first stage models in the DCC and Copula GARCH models, with the conditional mean propagated to the joint distribution (estimation, filtering, prediction and simulation).
• The cond_mean argument is now mutually exclusive with first stage ARMA dynamics at the model level: if any first stage series uses arma, supplying cond_mean raises an error in the specification, filter, simulate and predict methods.
• Added support for first stage mean equation regressors (xreg in garch_modelspec) in the DCC and Copula GARCH models: tsfilter and predict gain a newxreg argument and simulate an xreg argument, each taking a list with one element per series (fully named and possibly partial, or fully unnamed and complete).
• The cond_mean argument is now also mutually exclusive with first stage mean regressors, for the same reason as with ARMA dynamics.
• Added support for first stage variance equation regressors (vreg in garch_modelspec) in the DCC and Copula GARCH models: tsfilter and predict gain a newvreg argument and simulate a vreg argument, each taking a list with one element per series of raw regressor matrices with NROW(y), h and h + burn rows respectively. Regressors required by a first stage model but not supplied raise an error naming the affected series.
• Documented the first stage variance regressors in the vignettes, including the choice between the additive and multiplicative (exponential) forms, and that first stage mean and variance regressors supplied to the forward-looking methods are treated as a single deterministic path shared across all simulated paths, so regressor uncertainty is not propagated into the predictive distribution.
• Fixed simulate with burn > 0 failing for the DCC and Copula GARCH models: the burn-in period is now discarded once at the end of the joint recursion rather than in the correlation recursion before the first stage GARCH recursion could consume it, and the returned correlation, covariance and standardized residual arrays are correctly of length h.
• Minimum required version of tsgarch raised to 1.0.5.
• Fixed the univariate pre-sample trim length in the partitioned hessian/score calculation when the ARMA order exceeds the GARCH order.
• Fixed the joint hessian/score calculation failing when the first stage models have unequal numbers of estimated parameters (e.g. when variance targeting is used for only some series) (#7, reported by @kbuchardt).
• Fixed the sign of the second stage rows of the partitioned hessian and of the second stage columns of the score matrix in the DCC and Copula GARCH models, which were computed from the log-likelihood rather than the negative log-likelihood used for the first stage. As a result vcov(type = "H") returned negative variances for the second stage parameters and estfun returned second stage scores of the wrong sign.
• Fixed the QMLE and NW sandwich estimators in vcov (and hence summary with vcov_type = "QMLE" or "NW"), which used B M B rather than B M B'. Since the partitioned hessian of the two stage estimator is not symmetric, this understated the second stage standard errors and returned a non-symmetric matrix. The default OP standard errors used by summary are unaffected.
• The second derivatives in the partitioned hessian are now computed by central differences with steps relative to the parameter magnitude, improving the accuracy of the cross derivatives with respect to first stage parameters of small magnitude such as the GARCH intercept. This requires more likelihood evaluations, so estimation with return_hessian = TRUE is somewhat slower.
• The joint likelihood used in the partitioned hessian (and, for the constant correlation model, in the scores) of the Copula GARCH model now applies the same truncation of the probability integral transformed values as the estimation objective, so that the derivatives are those of the function actually minimised.
• Fixed pca_cov = "LW" failing in GOGARCH estimation due to a missing trace argument (#6, reported by @Caliani21).
• Replaced Rf_error with Rcpp::stop in the C++ code (#5, reported by @Enchufa2).

tsmarch 1.0.2

• Fix to DCC model simulation when order > 1.
• Bump requirement for R > 4.1.0 (required by tsgarch)
• Fix to Rf_error prompted from Rcpp team

tsmarch 1.0.1

• Small fixes to RADICAL algorithm.

• Added the FASTICA algorithm as an option to the GOGARCH model since the RADICAL algorithm still needs work for high dimensional systems and may be slow in those cases. This is a custom implementation which more closely follows the original Matlab code (https://research.ics.aalto.fi/ica/fastica/) rather than the other alternatives in R.

• Added the constant correlation test of Engle and Sheppard and a flextable method for pretty printing.