Changes in Version 1.7-0 (DA,BS,RN)

• Bug fix: rsdc_viterbi() no longer errors on a single-observation input (the forward recursion indexed out of bounds).
• rsdc_minvar()/rsdc_maxdiv() validate their inputs before reconstructing the covariance: volatilities must be finite and positive and correlations finite with modulus below one, so a malformed column now raises an informative error naming the offending row instead of silently producing a wrong-signed covariance matrix.
• rsdc_maxdiv() checks that Rsolnp::solnp() actually converged (and that its weights are finite and sum to one) rather than accepting whatever it returns. Both optimizers now count the periods in which they fell back to equal weights, report the count as $n_fallback, and warn once if it is positive; previously the fallback was silent.
• rsdc_forecast_ahead() validates horizon, and rsdc_forecast() validates control$threshold, instead of failing later with a cryptic message.
• Bug fix (standard errors): the numerical Hessian is now rejected when the optimum sits within the finite-difference step of the feasible boundary (a correlation within ~2e-3 of +/-1, or a nearly singular correlation matrix). The objectives return a large finite penalty there, so the differenced Hessian was huge but finite and its inverse produced absurd, silently reported standard errors (e.g. 2e-08). vcov()/summary()/ confint() now return NA/NULL with an explanatory warning instead.
• vcov(type = "opg") no longer requires the Hessian-based covariance: OPG needs only the per-observation scores, which are now stored whenever compute_se = TRUE. Scores are still withheld (and all covariance estimators refused) when a "noX" transition row had to be projected onto the simplex, since that point is not a stationary optimum.
• Depends is now R (>= 4.0): the package compiles C++11 (from R 4.0 the default C++ standard) and declaring CXX_STD explicitly would raise an R CMD check NOTE.
• rsdc_estimate() now rejects fewer than two observations instead of returning a degenerate fit.
• Windows build fix: src/Makevars/src/Makevars.win now link BLAS/LAPACK explicitly ($(LAPACK_LIBS) $(BLAS_LIBS) $(FLIBS)), as required by RcppArmadillo. Without it, R CMD INSTALL on Windows failed at the link step with undefined references (dgemv_, dgemm_, dsyev_, …); macOS linkers tolerate the undefined symbols, which is why the omission was invisible on the development machine.
• The package article (methodology, a strictly out-of-sample five-industry study, and a Monte Carlo validation) is developed separately and is reproducible from data shipped with the package; it is no longer shipped as a package vignette. See citation("RSDC") for the current reference. The package ships the lightweight “Getting started” vignette (rsdc-intro.Rmd).
• Reparameterized global search: rsdc_estimate()’s DEoptim stage now explores canonical partial correlations (Joe, 2006, J. Multivariate Analysis 97(10), 2177-2189) instead of raw pairwise correlations. Every point of the search box maps to a valid positive-definite correlation matrix, so the out-of-the-box global search is feasible at any K (the natural box has essentially no feasible point from K ~ 5 on). For method = "noX" with N >= 3 the transition head is searched as bounded per-row softmax logits (roughly the historical [0.01, 0.99] box on the log-odds scale), which removes the row-simplex penalty wall while keeping every regime reachable; the N = 2 head and all natural-parameter results are unchanged. Within each search, the top three mutually distant members of the final DE population are refined by L-BFGS-B and the best refined point is kept (the best DE point is not always in the best basin).
• Replication diagnostic: with control$n_starts >= 2, a warning is issued when the best log-likelihood is reached by only one of the independent searches (tolerance 2.5 log-points), signaling a possibly non-global optimum. control = list(n_starts = 4, cores = 4) is the recommended protocol for multimodal surfaces (e.g. "noX" with N >= 3); the default remains a single search.
• rsdc_starts() is unchanged and remains a fast deterministic complement: when both routes agree, the optimum is cross-certified.
• Behavior change: cores > 1 on a plain single fit no longer enables the internal parallelType of DEoptim; parallelism lives across the independent searches of the multi-start forms.
• New dataset ff5ind: daily value-weighted returns of five Fama-French industry portfolios (Manuf, Enrgy, HiTec, Hlth, Utils; Fama & French, 1997, J. Financial Economics 43(2), 153-193; data (c) Fama and French, Kenneth R. French Data Library) with the MCCC Aggregate index (2025 Sentometrics release, monthly, forward-filled across trading days) and the CBOE VIX (via FRED, VIXCLS), aligned on 5155 U.S. trading days, 2005-01-03 to 2025-06-30. Reproducible build in data-raw/ff5ind.R.

Changes in Version 1.6-0 (DA,BS,RN)

• New rsdc_starts(): data-driven warm starts for high-dimensional problems. In moderate dimensions the DEoptim global search works well, but for K >= 5 a random vector of pairwise correlations is almost never positive definite, so its initial population contains no feasible point. rsdc_starts() builds feasible starting vectors from the empirical correlations of low/…/high average-correlation sub-periods (positive definite by construction), with n_starts shrinkage variants as protection against local optima.
• rsdc_estimate() accepts an rsdc_starts object via control$start: each start is refined by the local optimizer (the global search is skipped), the highest-likelihood fit is kept, and the returned object carries start_logliks (spread across starts) and start_pars (per-start optima). control$cores parallelizes both multi-start forms (n_starts seeds or rsdc_starts warm starts); for a plain single fit it is forwarded to the backends, where cores > 1 enables parallel DEoptim (parallelType = 1), as before.
• New cores argument for rsdc_bootstrap() and rsdc_corr_bands(): task-level parallelism via the base parallel package (forked workers on Unix-alikes, a PSOCK cluster on Windows). All random draws are generated up front from one uninterrupted RNG stream and each parallelized task is deterministic given its inputs, so results are identical for any number of cores; only the wall time changes.
• Bug fix: the estimation backends no longer call set.seed() when a warm start skips the global search. Previously every warm-started refit reset the global RNG stream, so rsdc_bootstrap() — which interleaves simulation and warm-started re-estimation — produced nearly identical replicates from the second one on, drastically understating bootstrap standard errors and intervals (also via vcov(type = "bootstrap") and confint(type = "bootstrap")).

Changes in Version 1.5-0 (DA,BS,RN)

• Bug fix (audit): rsdc_forecast() now returns a scalar NA BIC (instead of a silent zero-length numeric(0)) when final_params$log_likelihood is missing.
• Bug fix (audit): a warm start (control$start) lying outside the default L-BFGS-B search box — as produced by the identifiability relabeling for N >= 3 (re-referenced softmax betas, former row-complement transition entries) — is no longer silently projected onto the box; the box is widened element-wise to keep the supplied start feasible (affects multi-start refits and rsdc_bootstrap() warm starts).
• CRAN hygiene: stray local session files under inst/simulation/results/ are now build-ignored, and the declared Rdpack import is used (importFrom(Rdpack, reprompt)), clearing both R CMD check NOTEs.
• rsdc_minvar()/rsdc_maxdiv(): when value_cols is a character vector and y carries column names, misaligned columns are now detected — reordered to value_cols (with a warning) when the names match as a set, otherwise a warning states that positional order is assumed. Previously the names were silently overwritten.
• rsdc_forecast_ahead() now validates that X_future has the same number of columns as the estimation covariate matrix X.
• plot.rsdc_fit() error message corrected: stored regime probabilities do not depend on compute_se.
• Documentation: rsdc_corr_bands() documents the truncation of invalid Gaussian draws near parameter bounds.
• Removed dead requireNamespace("mvtnorm") guards (mvtnorm is a hard dependency in Imports).
• Multi-start estimation: control$n_starts repeats the global+local search from several seeds and keeps the highest-likelihood fit, storing start_logliks. Since DEoptim is a stochastic global search, this is primarily a stability diagnostic across seeds (is the optimum reproducible?), not a replacement for it.
• rsdc_forecast_ahead(): genuine multi-step-ahead forecasts of the regime distribution and implied correlations, propagating the terminal filtered state through the Markov chain (X_future for the time-varying case).
• rsdc_corr_bands(): pointwise uncertainty bands for the predicted correlation path, by drawing parameters from the asymptotic sampling distribution and re-running the filter.
• rsdc_viterbi(): most likely (MAP) regime path via the Viterbi algorithm.
• broom tidiers tidy(), glance(), augment() and a ggplot2::autoplot() method for rsdc_fit (ggplot2 is a soft dependency in Suggests).
• Documentation: the “Getting started” vignette gains a section demonstrating the new features above (multi-start, bands, Viterbi, multi-step forecasts, tidy output), and the broom/autoplot methods gain runnable examples.

Changes in Version 1.4-0 (DA,BS,RN)

• Performance: the Hamilton-filter log-likelihood used during estimation is now evaluated in C++ (Rcpp/RcppArmadillo), matching the R reference to ~1e-8 and cutting estimation time by roughly an order of magnitude. The pure-R rsdc_hamilton() is retained (and used as the equivalence reference).
• Optimizer control: control now also accepts cores (parallel DEoptim via parallelType = 1) and start (a warm-start parameter vector that skips the global search and goes straight to local refinement).
• Parametric bootstrap: rsdc_bootstrap() simulates from the fitted model, re-estimates each replicate warm-started at the MLE, and returns bootstrap standard errors / percentile intervals that respect the parameter bounds.
• Robust standard errors: vcov()/confint() gain a type argument — "hessian" (default), "opg" (outer product of gradients), or "sandwich" (QML). Per-observation scores are stored on the fitted object.
• summary() now reports delta-method standard errors on the natural scale of the transition probabilities, plus regime diagnostics (stay probability, expected duration, ergodic distribution).
• Added a plot() method for fitted models (smoothed/filtered regime probabilities).
• Fitted objects now store the estimation residuals/covariates and the filtered/smoothed regime probabilities.
• Input-validation hardening (audit): rsdc_estimate()/rsdc_hamilton() now reject single-series input (K < 2) and a covariate matrix X whose row count does not match the data; rsdc_hamilton() validates that a user-supplied fixed transition matrix P is row-stochastic (finite, non-negative, rows summing to 1).
• confint() now uses the same safe square-root as summary(), so non-positive variance estimates map to NA instead of producing inconsistent NaN intervals.
• For the fixed-transition noX model with N >= 3, a transition row that the optimizer leaves infeasible (free probabilities summing above 1) is still projected onto the simplex, but the fit is now flagged as non-converged and its standard errors are suppressed (the projected point is not a valid basis for inference).

Changes in Version 1.3-0 (DA,BS,RN)

• rsdc_estimate() now returns an object of class "rsdc_fit" with standard S3 methods: print, summary, coef, logLik, nobs, vcov, confint, predict, and simulate. AIC()/BIC() work out of the box via logLik.
• Standard errors: the estimator now returns the observed-information variance-covariance (vcov) and standard errors (summary/confint), computed from the numerical Hessian of the negative log-likelihood at the MLE.
• Arbitrary number of regimes: N >= 4 is now supported for "noX" and "tvtp" (previously capped at N = 3).
• control now forwards optimizer settings (itermax, NP, parallelType, steptol, maxit, and compute_se) to DEoptim/optim, enabling faster runs.
• Robustness: warns when a transition probability is pinned to its bound and when a "tvtp" covariate matrix X has no intercept/constant column.
• Added a “Getting started” vignette.
• Moved the 596 KB source workbook out of inst/extdata into build-ignored data-raw/ to shrink the package tarball.

Changes in Version 1.2-0 (BS,DA,RN)

• Added mccc dataset: daily Media Climate Change Concerns (Aggregate) index, forward-filled from monthly and aligned row-for-row to greenbrown. Supplies the exogenous TVTP covariate used in the vignette, so the empirical workflow is fully reproducible from packaged data alone (no external Excel files).
• Added lightweight raw source inst/extdata/mccc-monthly.csv and builder data-raw/mccc.R.
• Fixed greenbrown documentation: date range is 2014-01-02 to 2022-12-30.

Changes in Version 1.1-2 (BS,DA)

• DESCRIPTION improved following the CRAN guidelines
• Seed is now a control parameter
• Removed examples for unexported functions
• Fixed do-not-run examples
• Added green-brown-ptf data in extdata
• Data added under the right .rda format
• Data Documentation added

Changes in Version 1.1-1 (DA)

• URLs fixed

Changes in Version 1.1-0 (DA)

• First release public version