NEWS

---

dsp 1.8.0

• Added accessor functions so that a fitted model can be inspected without referring to the structure of the object: par_names() lists the parameters with stored draws, fitted() returns pointwise posterior summaries of one parameter, DIC() returns the deviance information criterion under both estimators of the effective number of parameters, get_spec() returns the model specification and nobs() returns the number of time points.
• summary() now returns an object of class summary.dsp with a print method. It reports the model specification, the MCMC settings, the DIC and one line per parameter, and fits on one screen. Effective sample sizes are computed by coda::effectiveSize(); pointwise ones for time-varying parameters are opt in through ess = TRUE. The pointwise posterior summaries that summary() previously returned as a list are now returned by fitted(), with quantile columns named as by quantile() ("2.5%" rather than "q_2.5").
• Added credible_bands(), which returns the simultaneous credible bands of a time-varying parameter and its simultaneous band scores (SimBaS), the multiplicity-adjusted evidence at each time point that the parameter is nonzero. Previously the bands were available only on plots and SimBaS not at all. The result is a data frame of class dsp_bands with one row per time point, labelled by the optional times. Its print method reports the global Bayesian p-value and the periods in which the band excludes zero.
• The simultaneous credible bands drawn by plot(include_joint_bands = TRUE) were undefined when any time point had no posterior variation, such as a fixed initial value. Such points are now excluded when the bands are calibrated.
• A constant observation variance (obsSV = "const", and every B-spline model) is now stored as one draw per iteration rather than copied to every time point. summary() reports it with the scalar parameters, and plot() still draws it over time for comparison with time-varying variances.
• summary() reports the first D entries of evol_sigma_t2, the variances of the initial states, on their own line.
• The anomaly component of the changepoint model (useAnom = TRUE) now has the horseshoe+ prior of Wu et al. (2024), as documented. Earlier versions used a horseshoe prior, because the horseshoe+ lines were commented out.
• The anomaly zeta and its variance zeta_sigma_t2 are now returned with T columns aligned with y, rather than T - D. The anomaly is defined for t > D only, as in Wu et al. (2024), so both are zero at the first D time points.
• plot() no longer drops the first D entries of zeta_sigma_t2, which has no initial states, or the first entry of evol_sigma_t2 when D = 0.
• as.mcmc() is re-exported from coda, so as.mcmc(fit) works after library(dsp) alone.
• predict(cp_prop = TRUE) now returns the changepoint probabilities aligned with the observed series, with NA for the first D time points, so that cp_prop[cp_t] gives the probabilities at the detected changepoints.
• print() for fitted models reported nskip as the thinning interval. It now reports nskip + 1, the interval coda::thin() gives.
• plot() now rejects type = "DIC" and type = "p_d", which are not draws.
• Fixed a missing "and" in the printed specification of B-spline models.

dsp 1.7.0 (2026-10-16)

• Added a coda::as.mcmc() method for fitted dsp objects, so that the convergence diagnostics and plotting functions in coda can be applied directly. Parameters are flattened to one column per index, named so that individual time points can be addressed, and the result records the sampler's iteration index, including burn-in and thinning.
• Added is.dsp() and documented the dsp class, including the components of a fitted object and their dimensions.
• plot() now works without an explicit type, defaulting to "mu". The default was previously unreachable because the method tested whether the argument had been supplied rather than testing its value. Supplying NULL, an unrecognised name, a vector or a non-character value now gives one message naming the quantities the fitted object holds.
• Removed RemoteType, RemotePkgRef and RemoteUrl from DESCRIPTION, which were left behind by a local installation.
• Removed two unused internal functions, getEffSize() and ergMean().

dsp 1.6.0 (2026-07-21)

• Fixed random number generation so that set.seed controls results; btf_nb reset the seed and the state sampler drew from a separate stream.
• Added a seed argument to dsp_fit.
• Fixed r_user, which did not hold the overdispersion parameter fixed as documented.
• Negative binomial models now use the faster Polya-Gamma sampler when the overdispersion is integer valued.
• dsp_spec no longer accepts r_init and r_sample directly; both are set through r_user.
• Failed Cholesky factorizations are now detected and redrawn with a warning, rather than returning invalid states.
• Fixed plot method dropping every changepoint after the first and drawing them D observations early.
• Fixed pluralization of "degree" in the print method.
• Documented the slice and mh overdispersion samplers as experimental.

dsp 1.5.1 (2026-07-16)

• minor changes to colors used in plot for consistency

dsp 1.5.0 (2026-06-14)

• dsp_spec now abstracts overdispersion sampling from the user by simplifying to options to default or Poisson approximation.

dsp 1.4.1

• Changed typo in print method for family "negbinomial"

dsp 1.4.0 (2026-05-12)

• Changed argument name t01 to times for consistency
• Changed variable name yhat to ypred for correctness

dsp 1.3.0 (2026-04-20)

• Fixed inconsistencies in model specification family names.
• Enhanced the plotting method.

dsp 1.2.0 (2025-08-19)

dsp 1.1.0

dsp 1.0.0

• Initial CRAN submission.