NEWS.md

In dsge: Dynamic Stochastic General Equilibrium Models

dsge 1.2.0

This release collects the additions made between May and September 2026; the May-August ones were previously listed under 1.1.1. Dates are taken from the git history. Entries within each section are in chronological order.

New features

Nonlinear perfect foresight (2026-05-20)

• New perfect_foresight_nonlinear(): stacked-time Newton (LBJ) solver for deterministic transition paths of nonlinear DSGE models. Solves the full nonlinear equilibrium conditions simultaneously over the entire horizon, giving paths exact to Newton tolerance even for large shocks. Uses a block-bidiagonal Jacobian assembled by numerical finite differences and solved by O(T n³) block back-substitution; Armijo backtracking stabilises convergence. Returns the same dsge_perfect_foresight class as the linearized perfect_foresight(), so all existing plot, print, and summary methods apply unchanged. Reference: Juillard et al. (1998), Journal of Economic Dynamics and Control, 22, 1291–1318.

Variance decomposition (2026-05-21)

• New variance_decomposition() function with methods for dsge_solution, dsge_fit, and dsge_bayes. Computes either:
• the unconditional decomposition (default), giving each shock's share of the long-run steady-state variance of every observable via per-shock discrete-Lyapunov solves; or
• the forecast-error variance decomposition (FEVD) when a vector of integer horizons is supplied, giving each shock's share of the h-step-ahead forecast-error variance.
• Returns a dsge_variance_decomposition object with both raw contributions (in squared units) and percent shares; row sums equal 100 by construction.
• New plot.dsge_variance_decomposition() draws horizontal stacked bars for the unconditional case and one stacked-bar panel per observable (horizons on the x-axis) for FEVD. Uses the standard package theme palette and styling.

Closing the most-cited Dynare gaps: five features (2026-05-21)

• osr() -- Optimal Simple Rules. Finds the parameters of a user-specified, restricted policy rule that minimise an unconditional quadratic welfare loss subject to the model's rational-expectations equilibrium. Complements the existing fully-flexible ramsey_policy(). Implements Dennis (2007).
• conditional_forecast() -- forecasts conditional on a pre-specified path for a subset of observables (e.g. holding the policy rate fixed for k periods). Implements the Waggoner & Zha (1999) minimum-norm shock approach. Returns a dsge_conditional_forecast object that inherits from dsge_forecast so existing plot methods work.
• irf_match() -- impulse-response matching estimation. Estimates the model's structural parameters and shock SDs by minimising the weighted squared distance between model IRFs and a target IRF data frame (typically from a VAR). Follows Christiano, Eichenbaum & Evans (1999).
• bayes_dsge_var() -- Bayesian VAR with DSGE-implied prior (Del Negro & Schorfheide 2004). Combines a Bayesian VAR(p) with prior moments centred on the DSGE's second-moment implications, controlled by a single hyperparameter lambda. Returns posterior draws of the VAR coefficient matrix and innovation covariance and an approximate log marginal likelihood for choosing lambda.
• Multi-period / "news-shock" paths in perfect foresight. Both perfect_foresight() and perfect_foresight_nonlinear() already accept vector-valued shock paths; documentation now explicitly calls out this capability. The nonlinear stacked-time solver correctly models anticipated shocks (agents adjust at t=1 in expectation of a future shock); the linearised version uses the recursive policy and treats the path as a sequence of period-by-period surprises.

Full DSGE-VAR workflow, Dynare parity (2026-05-21)

• bayes_dsge_var_mh() -- joint Metropolis-Hastings estimation of the DSGE structural parameters, shock standard deviations, and the DSGE-prior weight lambda jointly. The VAR coefficients are analytically marginalised at every iteration via the closed-form Normal-inverse-Wishart conjugate posterior, so only (theta_DSGE, sigma, lambda) are sampled. Reuses the same adaptive RWMH sampler as bayes_dsge(). This brings the package to feature-parity with Dynare's estimation(..., dsge_var) command.
• forecast.dsge_dsgevar() and forecast.dsge_dsgevar_mh() -- unconditional fan-chart forecasts from a DSGE-VAR posterior. Iterates the VAR forward for each posterior draw, drawing innovations from N(0, Sigma) and aggregating across draws to produce posterior summary statistics.
• conditional_forecast.dsge_dsgevar() and the matching _mh() method -- DSGE-VAR conditional forecasts on a user-specified path for a subset of variables. Each posterior draw is conditioned via a per-period minimum-norm innovation injection (Sigma-metric minimum-norm gap closure), parallel to the Dynare Dvars_forecast.m workflow.

Derived parameters in dsge_model() (2026-05-22)

• dsge_model() gains a new derived = function(p) list(...) argument that maps primitive parameters to derived ones (analogue of Dynare's # macro substitutions). Called by solve_dsge() at every solve, so derived parameters track the current primitives during estimation. This enables encoding rich structural DSGEs (e.g. Smets-Wouters 2007) that depend on derived steady-state ratios like Rk, W, K_Y, beta_bar, etc.

Closing 9 more Dynare-feature gaps (2026-05-26)

Each is implemented as a standalone function with full documentation and tests:

• calibrated_smoother() -- run the Kalman smoother on a calibrated (un-estimated) model. Convenience wrapper around new smooth_states.dsge_solution(), smooth_shocks.dsge_solution(), shock_decomposition.dsge_solution() methods.
• predetermined() -- alias for state(..., shock = FALSE), matching Dynare's predetermined_variables declaration for users porting models across implementations.
• perfect_foresight_expect_err() -- perfect-foresight simulation with expectation errors: agents form subjective expectations at each period and nature delivers the actual shock path (potentially different). Analogue of Dynare's perfect_foresight_with_expectation_errors_solver.
• extended_path() -- Fair-Taylor / Adjemian-Juillard extended-path stochastic simulation. Works on dsge_solution (linear) and dsgenl_model (nonlinear) objects.
• endogenous_prior() -- Christiano-Trabandt-Walentin endogenous prior on model-implied second moments. Plugs into bayes_dsge() via a new endogenous_prior argument.
• global_sensitivity() -- global sensitivity analysis with Sobol' indices (first-order and total-effect) and Morris elementary effects. Targets include model-implied moments, IRF magnitudes, or log-likelihood. Companion to local sensitivity via parameter_sensitivity().
• discretionary_policy() -- time-consistent (no-commitment) optimal policy via the Soederlind / Dennis fixed-point iteration. Companion to ramsey_policy() (commitment) and osr() (restricted simple rules).
• gmm_estimate() and smm_estimate() -- generalised / simulated method-of-moments estimation matching empirical to model moments (SDs, variances, covariances, autocorrelations, correlations). Supports one-step and two-step (optimal) weighting matrices.
• bayes_smc() -- tempered Sequential Monte Carlo sampler (Chopin 2002; Herbst & Schorfheide 2014). Robust alternative to bayes_dsge() (RWMH) and bayes_particle() (PMMH), especially for multimodal posteriors. Returns a dsge_smc object inheriting from dsge_bayes so all existing diagnostics and post-estimation methods work.

Four further Dynare-feature gaps closed (2026-08-20)

• model_latex() -- export a model's equations as LaTeX (the analogue of Dynare's write_latex_dynamic_model). Handles both linear and nonlinear models, substitutes Greek-letter parameter names, puts time subscripts on variables and wraps leads in a conditional expectation operator. Options cover the align/gather environments, numbering, \label emission and a standalone compilable document. Verified by compiling the generated output with pdflatex.
• kalman_filter_skewed() -- likelihood evaluation when the structural shocks are skew-normal rather than Gaussian. The Kalman gain remains the optimal linear filter, and the filter additionally propagates the third cumulant of the state exactly (the observation equation carries no measurement error, so the update is a deterministic linear projection). The predictive density is built from univariate skew-normals matched to the exact model-implied marginal skewness, and reduces exactly to the Gaussian filter when all skewness is zero.
• ms_filter() -- Markov-switching shock volatility via the Kim (1994) filter, with Hamilton regime probabilities, the K^2-to-K collapse, the Kim backward smoother, and print/plot methods. At first order the policy functions are certainty-equivalent, so the regime rescales only the shock loading and a single solve suffices.
• pac_weights(), pac_target_loading(), pac_simulate() -- FRB/US-style polynomial adjustment cost equations. Factors the Euler equation's characteristic polynomial into stable and unstable roots, returns the implied lag coefficients and forward weights (normalised so long-run homogeneity holds exactly), and collapses the infinite forward sum into closed form when the target follows a linear state process.

Importing Dynare .mod files (2026-09-24)

• New read_dynare() reads a Dynare .mod file (or model code passed as text) and translates it into a dsgenl_model, together with the calibration, shock standard deviations, measurement errors, priors, optimal-policy problem and occasionally binding constraints it declares. No Dynare, MATLAB or Octave installation is needed. solve_dsge(), estimate(), bayes_dsge(), osr() and simulate_occbin() accept the imported object directly.
• Model and timing: var, varexo, varexo_det, parameters, predetermined_variables, varobs, parameter assignments, the model block (model(linear), equation tags, # model-local variables), leads and lags of any length on variables and shocks, and STEADY_STATE(). Lags become auxiliary state variables (x_lag1, ...), leads beyond one period become auxiliary controls (x_lead1, ...), and each shock becomes an exogenous state holding the current innovation, the same approach Dynare uses internally, so no manual re-timing is needed.
• Macro processor: @#define, @#if/@#elseif/@#else, @#ifdef/@#ifndef, @#for (arrays, ranges, tuples, when), @#include, @#echo/@#error, macro functions and @{...} are expanded in R (a defines argument plays the role of Dynare's -D). Dynare's own bkk and agtrend macro examples expand to exactly the equations Dynare generates.
• Steady state and shocks: steady_state_model, initval, shock standard deviations, variances, covariances and correlations (correlated shocks are orthogonalised by Cholesky factorisation in declaration order, as in Dynare's impulse responses) and deterministic shock paths.
• Observables and measurement errors: models may now have fewer observed variables than shocks (dsgenl_model() previously required equal numbers). A stderr on an observed variable is a measurement error: y is observed as y_obs = y + y_me, and estimate() / bayes_dsge() map a data column y to y_obs automatically.
• Priors: Dynare's mean/standard-deviation priors are converted exactly to dsge's parameterisation, including inv_gamma_pdf via the new "inv_gamma1" prior family (Dynare's type-1 inverse gamma on a standard deviation; prior("inv_gamma1", mean = , sd = ) works too). Anything not translated is reported.
• Estimation settings: presample, first_obs and nobs from the file's estimation command are applied by estimate() and bayes_dsge(), which gain presample arguments; bayes_dsge() also gains shock_start (defaulting to the file's shock standard deviations). Models declared model(linear) are linearised with an exact Jacobian (about 4x faster for Smets-Wouters).
• Optimal policy: ramsey_model / ramsey_policy add the planner's first-order conditions (derived symbolically) and Lagrange multipliers, with the Ramsey steady state found by concentrating out the multipliers; discretionary_policy closes a linear model with the time-consistent rule from the Dennis (2007) algorithm; osr_params, osr_params_bounds and optim_weights are passed to osr(), which now also accepts dsgenl_model objects.
• OccBin: occbin_constraints with bind/relax-tagged equations are solved by simulate_occbin() with the piecewise-linear Guerrieri-Iacoviello (2015) algorithm used by Dynare's occbin_solver (anticipated regime durations, surprise shocks).
• solve_dsge() on an imported model honours values supplied in params for parameters that are otherwise fixed.
• Smets & Wouters (2007), end to end: the standard replication file (40 variables, 7 shocks, 36 estimated parameters) imports unchanged and, at the published posterior mode on the US data, reproduces Dynare's 280 impulse responses to 2.5e-12 and its log-likelihood (-1714.061158377), log-prior (-23.994069948) and posterior kernel to all nine printed decimals.
• Validated against Dynare 6.0 (scripts in dev/dynare-validation/): first-order impulse responses of 12 models (160 responses, including Dynare's example1, example2, agtrend and bkk, Ramsey, discretion, STEADY_STATE(), shock leads, long leads/lags and correlated shocks) agree to within 2e-6, and to within 5e-8 for all but bkk; log-likelihoods with measurement errors and with fewer observables than shocks agree to Dynare's printed precision; OSR optimum and loss agree (loss to 1e-13); OccBin paths agree to 3e-13.
• New example file inst/examples/rbc.mod.

Closer to Dynare on real-world files (2026-09-24)

• MATLAB code in .mod files and _steadystate.m files: a built-in MATLAB interpreter (matrices, cell arrays, structures, indexing, control flow, subfunctions, anonymous functions, eval, fsolve, csolve, fzero, ...) runs the MATLAB statements of a .mod file (calibrations such as sigma = sqrt(V(1, 1)), verbatim blocks, set_param_value()) and a <model>_steadystate.m file, including helper functions in other .m files of the model's folder. Parameters that a steady-state file or steady_state_model sets are recomputed from the other parameters whenever the model is solved, as in Dynare. Parameters take the values they have at the file's first computing command.
• lik_init = 2: the Kalman filter can start, as Dynare does, from a covariance of 10 times the identity on Dynare's state vector (observed and predetermined variables). Smets & Wouters (2007) with its original lik_init = 2 now reproduces Dynare's log-likelihood (-1738.513893160) to all printed decimals.
• Leads of two or more periods inside nonlinear terms become auxiliary variables as in Dynare (exact at every order of approximation).
• Parsing: % inside MATLAB strings is no longer taken for a comment; MATLAB statements without a semicolon, multi-line matrices, time indices on parameters, steady_state() in lower case and vector values of deterministic shocks are handled; discretionary policy is imposed through the planner's targeting rule, which is determinate where the instrument's reaction to shocks alone was not.
• Validated on Johannes Pfeifer's DSGE_mod collection (68 files, dev/dynare-validation/extra/batch_dsge_mod.R): 67 import unchanged; the first-order impulse responses of all 53 stochastic models Dynare 6.0 runs in Octave agree with Dynare's (52 to within 1e-6 relative, one with linearly dependent states to 7e-5). Second- and third-order decision rules of 21 nonlinear models agree with Dynare's to within 1e-6 (relative), most to 1e-9 or better (dev/dynare-validation/extra/validate_higher_order.R).

Perfect foresight and MATLAB data files for imported models (2026-09-25)

• New simulate_perfect_foresight() solves the deterministic path of a model imported with read_dynare() as Dynare's perfect_foresight_setup / perfect_foresight_solver (and simul) do: initial and terminal conditions from initval, endval, histval, steady and oo_.endo_simul(..., 1) = ..., deterministic shocks (temporary and permanent), and complementarity conditions from mcp equation tags (Dynare's lmmcp option, e.g. a zero lower bound). The stacked system is solved by Newton's method with a sparse Jacobian from symbolic derivatives (using the Matrix package when installed), with Levenberg-Marquardt steps where the Jacobian is singular. On Johannes Pfeifer's DSGE_mod collection the paths of all six perfect-foresight models Dynare solves in Octave (Solow models, Ramsey-Cass-Koopmans, Ramsey policy from t0, a ZLB model solved with lmmcp) agree with Dynare's to within 2.5e-7, mostly 1e-13.
• The MATLAB interpreter used by read_dynare() now reads data files (load of Octave text files, MATLAB .mat files via R.matlab, plain numeric files; xlsread/readmatrix via readxl; csvread, dlmread) and provides fmincon (bounds, linear and nonlinear constraints), fminunc, lsqnonlin, hpfilter, ksdensity, interp1, polyfit/polyval, prctile, cov and corrcoef. Errors now name the MATLAB statement that failed first (e.g. a missing data file). Matrix, R.matlab and readxl are suggested packages.

Improvements

Publication-ready plot theme (2026-05-21)

• All plot.dsge_* methods now use a unified visual theme: consistent navy/brick/olive palette, light dotted gridlines, thin solid zero reference lines, semi-transparent confidence bands, and sans-serif typography. Centralised in a new internal R/plot-theme.R, so the look-and-feel of every figure -- IRFs, forecasts, smoothed states and fit, historical shock decomposition, perfect-foresight paths, occasionally-binding-constraint comparisons, identification diagnostics, sensitivity tornadoes, prior-posterior overlays, MCMC traces, posterior densities, running means, ACFs and posterior predictive checks -- is now uniform and publication-ready. No API changes; all existing user code keeps working.

Forecast fan chart + history overlay (2026-05-21)

• forecast.dsge_fit() now also returns the forecast standard deviation at each horizon (sd column in forecasts), computed by iterating the state covariance forward analytically. The result also carries the in-sample observed data as history.
• plot.dsge_forecast() now shows the historical series in grey, followed by the point forecast in navy with three nested fan bands at the 50/80/95% levels and a vertical separator at the forecast origin.

Smoothed state uncertainty bands (2026-05-21)

• The Kalman smoother now propagates and stores the smoothed state covariances. smooth_states() returns an additional matrix smoothed_states_var (T x n_states) with the per-period state variances.
• plot.dsge_smoothed(..., type = "states") automatically draws semi-transparent +/- 2 sigma bands around each smoothed state when the variance information is available.

Perfect-foresight comparison overlay (2026-05-21)

• plot.dsge_perfect_foresight() gains a new compare = ... argument that overlays a second perfect-foresight path on the same panels (typical use: pass the linearized perfect_foresight() result alongside a perfect_foresight_nonlinear() result to visualise the nonlinearity premium for large shocks).

Bug fixes

Second- and third-order perturbation (2026-09-24)

• The second-order risk correction (g_ss, h_ss) left out the curvature of the equations in next-period variables and added a spurious h_xx term, so it could have the wrong sign and size; the third-order terms g_xxx, g_xss and g_sss were also wrong. Both orders are now computed following Schmitt-Grohe and Uribe (2004) and Andreasen (2012), with exact symbolic derivatives of the model equations and generalized Sylvester solvers that scale to large models (the previous dense system grew with the fourth power of the number of states). Solutions now match Dynare's to about 1e-10 on standard models.

More robust first-order solver (2026-09-24)

• solve_dsge() now solves linearised models by cyclic reduction and, failing that, an inverse-free spectral divide (Bai, Demmel and Gu 1997) instead of only a fixed-point iteration, which diverged on several standard models (e.g. Hansen 1985, Kiyotaki-Moore 1997, Schmitt-Grohe and Uribe 2003). Unit roots are allowed, as in Dynare.

Robust Lyapunov solver (2026-05-22)

• compute_unconditional_P() now falls back to the doubling algorithm (Smith / Anderson) when the Kronecker-form system (I - H ⊗ H) is near-singular -- typical for medium-scale DSGEs with highly persistent shocks (eigenvalues close to 1, as in Smets-Wouters). Previously the solver returned a fake fallback of diag(1e6), which inflated model_covariance() and DSGE-VAR prior moments by many orders of magnitude. Affected functions (model_covariance, variance_decomposition, bayes_dsge_var, bayes_dsge_var_mh) now produce sensible moments on highly persistent models.

VAR stability filtering in DSGE-VAR forecasts (2026-05-22)

• forecast.dsge_dsgevar(), forecast.dsge_dsgevar_mh() and the matching conditional-forecast methods now drop any posterior draw whose VAR companion matrix has eigenvalues outside the unit disc, emitting a message reporting the number of skipped draws. Prevents occasional explosive paths when the DSGE-VAR posterior has long tails.

Package maintenance

• .Rbuildignore now excludes all hidden top-level files and folders (such as local editor settings), so they are never bundled into the source tarball.

dsge 1.1.0

New features

Bayes factor model comparison

• New bayes_factor() for pairwise Bayesian model comparison. Accepts two or more dsge_bayes or dsge_marginal_likelihood objects, computes log Bayes factors, posterior model probabilities, and Kass-Raftery (1995) evidence labels. Custom prior model probabilities supported via prior_odds.

Parallel MCMC chains

• bayes_dsge() gains an n_cores argument. When n_cores > 1 chains run simultaneously using parallel::mclapply (POSIX) or a PSOCK cluster (Windows), with automatic fallback to sequential execution.

Third-order perturbation

• solve_dsge() now accepts order = 3 for nonlinear (dsgenl_model) models, computing cubic state coefficients (g_xxx, h_xxx), state-sigma^2 cross terms (g_xss, h_xss), and sigma^3 corrections (g_sss, h_sss) following Schmitt-Grohe and Uribe (2004).
• New simulate_3rd_order() for pruned simulation of third-order approximations (Andreasen, Fernandez-Villaverde and Rubio-Ramirez, 2018).

Bootstrap particle filter

• New particle_filter(): sequential importance resampling with systematic resampling and numerically stable log-sum-exp weights.
• New particle_filter_loglik(): convenience wrapper accepting a dsge_solution object.
• New bayes_particle(): Particle Marginal Metropolis-Hastings (PMMH; Andrieu, Doucet and Holenstein, 2010) for fully nonlinear Bayesian estimation without any linearization. Returns a dsge_particle object that inherits from dsge_bayes, so all existing diagnostics apply.

Ramsey optimal policy

• New ramsey_policy(): computes the welfare-maximising linear feedback rule by solving the discrete-time algebraic Riccati equation (DARE) via value-function iteration. Accepts quadratic welfare weights on states (Q_xx), controls (Q_yy), and cross terms (Q_xy).
• New welfare_loss(): evaluates expected welfare loss under an arbitrary (possibly non-optimal) feedback policy for comparison with the Ramsey solution.

dsge 1.0.0

Major release: complete DSGE estimation and analysis toolkit

Model-implied covariance reporting (v0.10.0)

• New model_covariance() for unconditional covariance and correlation matrices of model observables and controls.
• Supports ML fits, Bayesian fits, and raw solution objects.

Prediction and reporting tools (v0.10.0)

• New prediction_interval() for one-step-ahead prediction bands using Kalman filter innovation variance.
• New prediction_accuracy() for RMSE, MAE, and mean bias statistics.
• Enhanced fitted() method for ML-estimated models.

Robust standard errors (v0.10.0)

• New robust_vcov() for sandwich (Huber-White) variance-covariance estimation. Provides standard errors robust to model misspecification.

Bayesian diagnostics and model comparison (v0.10.0)

• New posterior_predictive() for posterior predictive checks with variance and autocorrelation statistics.
• New marginal_likelihood() via modified harmonic mean estimator for Bayesian model comparison.
• New geweke_test() for Geweke (1992) convergence diagnostic.
• New mcmc_diagnostics() for comprehensive MCMC health summary.

dsge 0.9.0

Occasionally binding constraints

• New simulate_occbin() for piecewise-linear simulation under inequality constraints (e.g., zero lower bound).
• New obc_constraint() helper for constraint specification.
• Shadow-shock algorithm with iterative regime detection.
• Plot method with constraint-binding period shading.

dsge 0.8.0

Second-order perturbation

• solve_dsge() now accepts order = 2 for second-order approximation of nonlinear models.
• New simulate_2nd_order() for pruned second-order simulation.
• New irf_2nd_order() for generalised impulse-response functions with asymmetric shock effects.
• Risk/precautionary corrections to the steady state (sigma-correction).

dsge 0.7.0

Perfect foresight / deterministic transition paths

• New perfect_foresight() for deterministic simulation of transition paths after temporary or permanent shocks.
• Supports both linear and linearized nonlinear models.
• Plot method with deviation-from-steady-state and level views.

dsge 0.6.0

Identification and sensitivity analysis

• New check_identification() for local identification diagnostics via Jacobian SVD of the autocovariance mapping.
• New parameter_sensitivity() for sensitivity of likelihood, IRFs, steady state, and policy matrices to parameter perturbations.
• New prior_posterior_update() for Bayesian informativeness diagnostics comparing posterior concentration to prior width.

dsge 0.5.0

Kalman smoother and shock decomposition

• New smooth_states() for Rauch-Tung-Striebel Kalman smoothing.
• New smooth_shocks() for extraction of smoothed structural shocks.
• New shock_decomposition() for historical decomposition of observed variables into individual shock contributions.
• Plot methods for smoothed states and shock decomposition.

dsge 0.4.0

Bayesian nonlinear DSGE estimation

• bayes_dsge() now accepts dsgenl_model objects for Bayesian estimation of nonlinear DSGE models via first-order perturbation.
• For each MCMC draw, the model's steady state is re-solved and the system re-linearized at the candidate parameters, ensuring correct parameter- dependent dynamics.
• Failed steady-state solves, linearization failures, and Blanchard-Kahn violations are handled gracefully --- proposals are rejected without crashing the sampler. Failure counts are reported in the output.
• Data is automatically centered around the parameter-specific steady state for nonlinear models.
• Posterior IRFs work for nonlinear models via the same irf() interface.

dsge 0.3.0

Bayesian estimation (linear)

• New bayes_dsge() function for Bayesian estimation of linear DSGE models via adaptive Random-Walk Metropolis-Hastings (RWMH).
• New prior() constructor for specifying prior distributions: normal, beta, gamma, uniform, and inv_gamma.
• Multiple chains with dispersed starting points from prior draws.
• Adaptive proposal covariance during warmup targeting ~25% acceptance rate.
• Parameter transformations (log, logit) with Jacobian corrections for constrained parameters.
• MCMC diagnostics: effective sample size (ESS), R-hat (split), MCSE, and acceptance rates.
• Posterior impulse-response functions via irf() with pointwise credible bands computed from posterior draws.

dsge 0.2.0

Nonlinear DSGE support

• New dsgenl_model() constructor for nonlinear DSGE models defined via string-based equations with VAR(+1) lead notation.
• steady_state() generic for computing the deterministic steady state via Newton-Raphson with damped line search.
• linearize() computes first-order Taylor expansion around steady state.
• solve_dsge() and estimate() now accept dsgenl_model objects.

dsge 0.1.0

Initial release.

• Formula-based linear model specification.
• Klein (2000) solution via method of undetermined coefficients.
• Maximum likelihood estimation via Kalman filter.
• Policy/transition matrix extraction with delta-method standard errors.
• Impulse-response functions, forecasting, stability diagnostics.

dsge documentation built on Sept. 25, 2026, 5:08 p.m.