ModalForecast 0.2.0
New features
- Seasonal models.
fit_modal_arima() gains a
seasonal argument, with the same form as in
stats::arima(), to fit Modal SARIMA(p,d,q)(P,D,Q)[s]
models. The seasonal coefficients are reported as sar1,
sma1, and so on. Existing non-seasonal calls are
unchanged.
auto.modal.arima() searches over seasonal orders
(max.P, max.Q), chooses the seasonal
differencing order D with forecast::nsdiffs(),
and gains a trace argument.forecast() gains
point = c("joint", "marginal"). The default,
"joint", is the most probable future trajectory, as in
0.1.0. "marginal" gives the conditional mode of each future
value, which differs from the joint path for skewed errors and horizons
greater than one. The returned object also contains
mode_path and mode_shift.- Asymptotic prediction intervals now use the exact distribution of
the forecast error at the estimated parameters, computed by numerical
convolution of the SKD densities.
- New
vcov() and nobs() methods.
summary() reports a test of symmetry, H0: gamma = 1.
diagnostics() and envelope() support
seasonal models; for these the residual ACF, PACF and Ljung-Box test
cover two seasonal periods.envelope() gains refit = TRUE (the
default): each replication simulates a series from the fitted model and
re-estimates it, so the envelope accounts for parameter estimation.
Without it, the smallest distances of a correct Skewed Laplace fit fall
below the envelope, because its maximum likelihood fit pulls several
residuals to zero. refit = FALSE keeps the faster envelope
based on simulated innovations.- Sergio Luis Mercado Londoño and Víctor Hugo Lachos join as
contributors.
Changes in parameterization
- The Skew-Normal now uses the parameterization of Galarza et
al. (2017),
f(y) = 4p(1-p)/sqrt(2*pi*sigma^2) * exp(-2*rho_p((y-mu)/sigma)^2),
the same as the Skewed Student-t and Skewed Laplace. sigma
therefore has the same meaning for the three distributions. For
Skew-Normal fits, the log-likelihood, the systematic coefficients and
gamma are unchanged; the reported sigma equals
the 0.1.0 value divided by (gamma + 1/gamma)/2.
BIC() and logLik() use the number of
observations after differencing.
Bug fixes
- The analytical gradients of the Skewed Student-t and Skewed Laplace
log-likelihoods had errors, so the optimizer could stop far from the
maximum while reporting convergence. For example, a Skewed Laplace
ARIMA(2,1,1) for
log10(lynx) reached a log-likelihood of
-189.99 in 0.1.0 and -0.14 now. Estimates from these two distributions
may change substantially.
- The Skewed Laplace fit is refined with Nelder-Mead after BFGS,
because its log-likelihood is not differentiable when a residual is
zero, and its standard errors use the outer product of the scores
instead of the Hessian.
- Standard errors of
sigma, gamma and
nu in summary() were those of their
logarithms; they now use the delta method.
- The random generators used for bootstrap intervals and envelopes did
not match the fitted Student-t and Laplace densities (the Laplace draws
were half as wide even for
gamma = 1). All draws now come
from the fitted density.
- The Skewed Laplace CDF was wrong below the mode, which affected the
lower asymptotic prediction bounds and the randomized quantile
residuals.
- The theoretical quantiles of the Laplace envelope were half their
correct value.
- Removed an unrelated script,
inst/extras/devTests.R.
ModalForecast 0.1.0
- Initial release.
- Added
fit_modal_arima() for fitting Parametric Modal
ARIMA models using the generalized SKD family.
- Implemented S3 methods:
print, coef,
AIC, BIC, summary, and
plot.
- Added comprehensive simulation tests demonstrating parameter
recovery (c, ar1, sigma, gamma).