Dynare 6.0 Released

Posted on 05 February 2024

We are pleased to announce the release of Dynare 6.0.

This major release adds new features and fixes various bugs.

The Windows, macOS, MATLAB Online and source packages are already available for download at the Dynare website.

This release is compatible with MATLAB versions ranging from 9.5 (R2018b) to 23.2 (R2023b), and with GNU Octave versions ranging from 7.1.0 to 8.4.0 (NB: the Windows package requires version 8.4.0 specifically).

Major user-visible changes

  • The Sequential Monte Carlo sampler as described by Herbst and Schorfheide (2014) is now available under value hssmc for option posterior_sampling_method.

  • New routines for perfect foresight simulation with expectation errors. In such a scenario, agents make expectation errors in that the path they had anticipated in period 1 is not realized exactly. More precisely, in some simulation periods, they may receive new information that makes them revise their anticipation for the path of future shocks. Also, under this scenario, it is assumed that agents behave as under perfect foresight, i.e. they make their decisions as if there were no uncertainty and they knew exactly the path of future shocks; the new information that they may receive comes as a total surprise to them. Available under new perfect_foresight_with_expectation_errors_setup and perfect_foresight_with_expectation_errors_solver commands, and shocks(learnt_in=…), mshocks(learnt_in=…) and endval(learnt_in=…) blocks.

  • New routines for IRF matching with stochastic simulations:

    • Both frequentist (as in Christiano, Eichenbaum, and Evans, 2005) and Bayesian (as in Christiano, Trabandt, and Walentin, 2010) IRF matching approaches are implemented. The core idea of IRF matching is to treat empirical impulse responses (e.g. given from an SVAR or local projection estimation) as data and select model parameters that align the model’s IRFs closely with their empirical counterparts.

    • Available under option mom_method = irf_matching option to the method_of_moments command.

    • New blocks matched_irfs and matched_irfs_weights for specifying the values and weights of the empirical impulse response functions.

  • Pruning à la Andreasen et al. (2018) is now available at an arbitrary approximation order when performing stochastic simulations with stoch_simul, and at 3rd order when performing particle filtering.

  • New log option to the var statement. In addition to the endogenous variable(s) thus declared, this option also triggers the creation of auxiliary variable(s) equal to the log of the corresponding endogenous variable(s). For example, given a var(log) y; statement, two endogenous will be created (y and LOG_y), and an auxiliary equation linking the two will also be added (equal to y = exp(LOG_y);). Moreover, every occurrence of y in the model will be replaced by exp(LOG_y). This option is, for example, useful for performing a loglinear approximation of some variable(s) in the context of a first-order stochastic approximation; or for ensuring that the variable(s) stay(s) in the definition domain of the function defining the steady state or the dynamic residuals when the nonlinear solver is used.

  • New model editing features

    • Multiple model blocks are now supported (this was already working but not explicitly documented).

    • Multiple estimated_params blocks now concatenate their contents (instead of overwriting previous ones, which was the former undocumented behavior); an overwrite option has been added to provide the old behavior.

    • New model_options statement to set model options in a global fashion.

    • New model_remove command to remove equations.

    • New model_replace block to replace equations.

    • New var_remove command to remove variables (or parameters).

    • New estimated_params_remove block to remove estimated parameters.

  • Stochastic simulations

    • Performance improvements for simulation of the solution under perturbation and for particle filtering at higher order (⩾ 3).

    • Performance improvement for the first order perturbation solution using either cycle reduction (dr=cycle_reduction option) or logarithmic reduction (dr=logarithmic_reduction).

    • New nomodelsummary option to the stoch_simul command, to suppress the printing of the model summary and the covariance of the exogenous shocks.

  • Estimation

    • A truncated normal distribution can now be specified as a prior, using the 3rd and 4th parameters of the estimated_params block as the bounds.

    • New conditional_likelihood option to the estimation command. When the option is set, instead of using the Kalman filter to evaluate the likelihood, Dynare will evaluate the conditional likelihood based on the first-order reduced form of the model by assuming that the initial state vector is at its steady state.

    • New additional_optimizer_steps option to the estimation command to trigger the sequential execution of several optimizers when looking for the posterior mode.

    • The generate_trace_plots command now allows comparing multiple chains.

    • The Geweke and Raftery-Lewis convergence diagnostics will now also be displayed when mh_nblocks>1.

    • New robust, TolGstep, and TolGstepRel options to the optimizer available under mode_compute=5 (“newrat”).

    • New brooks_gelman_plotrows option to the estimation command for controlling the number of parameters to depict along the rows of the figures depicting the Brooks and Gelman (1998) convergence diagnostics.

    • New mh_init_scale_factor option to the estimation command tor govern the overdispersion of the starting draws when initializing several Monte Carlo Markov Chains. This option supersedes the mh_init_scale option, which is now deprecated.

  • Steady state computation

    • Steady state computation now accounts for occasionally-binding constraints of mixed-complementarity problems (as defined by mcp tags).

    • New tolx option to the steady command for governing the termination based on the step tolerance.

    • New fsolve_options option to the steady command for passing options to fsolve (in conjunction with the solve_algo=0 option).

    • New option from_initval_to_endval option to the homotopy_setup block, for easily computing homotopy from initial to terminal steady state (when the former is already computed).

    • New non_zero option to resid command to restrict display to non-zero residuals.

  • Perfect foresight

    • Significant performance improvement of the stack_solve_algo=1 option to the perfect_foresight_solver command (Laffargue-Boucekkine-Juillard algorithm) when used in conjunction with options block and/or bytecode of the model block.

    • New relative_to_initval option to the mshocks block, to use the initial steady state as a basis for the multiplication when there is an endval block.

    • New static_mfs option to the model block (and to the model_options command), for controlling the minimum feedback set computation for the static model. It defaults to 0 (corresponding to the behavior in Dynare version 5).

    • Various improvements to homotopy

      • New endval_steady option to the perfect_foresight_setup command for computing the terminal steady state at the same time as the transitory dynamics (and new options steady_solve_algo, steady_tolf, steady_tolx, steady_maxit and steady_markowitz for controlling the steady state nonlinear solver).

      • New homotopy_linearization_fallback and homotopy_marginal_linearization_fallback options to the perfect_foresight_solver command to get an approximate solution when homotopy fails to go to 100%.

      • New homotopy_initial_step_size, homotopy_min_step_size, homotopy_step_size_increase_success_count and homotopy_max_completion_share options to the perfect_foresight_solver command to fine tune the homotopy behavior.

      • Purely backward, forward and static models are now supported by the homotopy procedure.

    • The stack_solve_algo=1 and stack_solve_algo=6 options of the perfect_foresight_solver command were merged and are now synonymous. They both provide the Laffargue-Boucekkine-Juillard algorithm and work with and without the block and bytecode options of the model block. Using stack_solve_algo=1 is now recommended, but stack_solve_algo=6 is kept for backward compatibility.

  • OccBin

    • New simul_reset_check_ahead_periods option to the occbin_setup and occbin_solver commands, for resetting check_ahead_periods in each simulation period.

    • new simul_max_check_ahead_periods, likelihood_max_check_ahead_periods, and smoother_max_check_ahead_periods options to the occbin_setup command, for truncating the number of periods for which agents check ahead which regime is present.

  • Optimal policy

    • The osr command now accepts the analytic_derivation and analytic_derivation_mode options.

    • The evaluate_planner_objective command now computes the unconditional welfare for higher-order approximations (⩾ 3).

    • New periods and drop options to the evaluate_planner_objective command.

  • Semi-structural models

    • New pac_target_info block for decomposing the PAC target into an arbitrary number of components. Furthermore, in the presence of such a block, the new pac_target_nonstationary operator can be used to select the non stationary part of the target (typically useful in the error correction term of the PAC equation).

    • New kind option to the pac_model command. This option allows the user to select the formula used to compute the weights on the VAR companion matrix variables that are used to form PAC expectations.

    • Performance improvement to solve_algo=12 and solve_algo=14, which significantly accelerates the simulation of purely backward, forward and static models with the perfect_foresight_solver command and the routines for semi-structural models.

  • dseries classes

    • The remove and remove_ methods now accept a list of variables (they would previously only accept a single variable).

    • New MATLAB/Octave command dplot to plot mathematical expressions generated from variables fetched from (different) dseries objects.

  • Misc

    • New display_parameter_values command to print the parameter values in the command window.

    • New collapse_figures_in_tabgroup command to dock all figures.

    • Performance improvement for the use_dll option of the model block. The preprocessor now takes advantage of parallelization when compiling the MEX files.

    • New mathematical primitives available: complementary error function (erfc), hyperbolic functions (cosh, sinh, tanh, acosh, asinh, atanh).

    • New last_simulation_period option to the initval_file command.

    • The calib_smoother command now accepts the nobs and heteroskedastic_filter options.

    • Under the MATLAB Desktop, autocompletion is now available for the dynare command and other CLI commands (thanks to Eduard Benet Cerda from MathWorks).

    • Model debugging: The preprocessor now creates files for evaluating the left- and right-hand sides of model equations separately. For a model file called ramst.mod, you can call [lhs,rhs]=ramst.debug.static_resid(y,x,params); (for the static model) and [lhs,rhs]=ramst.debug.dynamic_resid(y,x,params,steady_state); (for the dynamic model), where y are the endogenous, x the exogenous, params the parameters, and steady_state is self-explanatory. NB: In the dynamic case, the vector y of endogenous must have 3n elements where n is the number of endogenous (including auxiliary ones); the first n elements correspond to the lagged values, the middle n elements to the contemporaneous values, and the last n elements to the lead values.

    • New interactive MATLAB/Octave command search for listing the equations in which given variable(s) appear (requires json command line option).

    • The model_info command allows to print the block decomposition even if the block option of the model block has not been used, by specifying the new options block_static and block_dynamic.

    • There is now a default value for the global initialization file (GlobalInitFile option of the configuration file): the global_init.m in the Dynare configuration directory (typically $HOME/.config/dynare/global_init.m under Linux and macOS, and c:\Users\USERNAME\AppData\Roaming\dynare\global_init.m under Windows).

    • For those compiling Dynare from source, the build system has been entirely rewritten and now uses Meson; as a consequence, it is now faster and easier to understand.

  • References:

    • Andreasen, Martin M., Jesús Fernández-Villaverde, and Juan Rubio-Ramírez (2018): “The Pruned State-Space System for Non-Linear DSGE Models: Theory and Empirical Applications,” Review of Economic Studies, 85(1), 1-49.
    • Brooks, Stephen P., and Andrew Gelman (1998): “General methods for monitoring convergence of iterative simulations,” Journal of Computational and Graphical Statistics, 7, pp. 434–455.
    • Christiano, Eichenbaum and Charles L. Evans (2005): “Nominal Rigidities and the Dynamic Effects of a Shock to Monetary Policy,” Journal of Political Economy, 113(1), 1–45.
    • Christiano, Lawrence J., Mathias Trabandt, and Karl Walentin (2010): “DSGE Models for Monetary Policy Analysis,” In: Handbook of Monetary Economics 3, 285–367.
    • Herbst, Edward and Schorfheide, Frank (2014): “Sequential Monte Carlo Sampling for DSGE Models,” Journal of Applied Econometrics, 29, 1073-1098.

Incompatible changes

  • The default value of the mode_compute option of the estimation command has been changed to 5 (it was previously 4).

  • When using block decomposition (with the block option of the model block), the option mfs now defaults to 1. This setting should deliver better performance in perfect foresight simulation on most models.

  • The default location for the configuration file has changed. On Linux and macOS, the configuration file is now searched by default under dynare/dynare.ini in the configuration directories defined by the XDG specification (typically $HOME/.config/dynare/dynare.ini for the user-specific configuration and /etc/xdg/dynare/dynare.ini for the system-wide configuration, the former having precedence over the latter). Under Windows, the configuration file is now searched by default in %APPDATA%\dynare\dynare.ini (typically c:\Users\USERNAME\AppData\Roaming\dynare\dynare.ini).

  • The information stored in oo_.endo_simul, oo_.exo_simul, and oo_.irfs is no longer duplicated in the base workspace. New helper functions send_endogenous_variables_to_workspace, send_exogenous_variables_to_workspace, and send_irfs_to_workspace have been introduced to explicitly request these outputs and to mimic the old behavior.

  • The dynare_sensitivity command has been renamed sensitivity. The old name is still accepted but triggers a warning.

  • The syntax resid(1) is no longer supported.

  • The mode_compute=6 option to the estimation command now recursively updates the covariance matrix across the NumberOfMh Metropolis-Hastings runs, starting with the InitialCovarianceMatrix in the first run, instead of computing it from scratch in every Metropolis-Hastings run.

  • The periods command has been removed.

  • The Sigma_e command has been removed.

  • The block option of the model block no longer has an effect when used in conjunction with stoch_simul or estimation commands.

  • The Dynare++ executable is no longer distributed since almost all of its functionalities have been integrated inside Dynare for MATLAB/Octave.

  • A macro-processor variable defined without a value (such as @#define var in the .mod file or alternatively -Dvar on the dynare command line) is now assigned the true logical value (it was previously assigned 1).

  • The parallel_slave_open_mode option of the dynare command has been renamed parallel_follower_open_mode.

  • The static option of the model_info command is now deprecated and is replaced by the block_static option.

Bugs that were present in 5.5 and that have been fixed in 6.0

  • The mh_initialize_from_previous_mcmc option of the estimation command would not work if estimation was conducted with a different prior and the last draw in the previous MCMC fell outside the new prior bounds
  • When specifying a generalized inverse Gamma prior, the hyperparameter computation would erroneously ignore the resulting mean shift
  • When using the mh_recover option of the estimation command, the status bar always started at zero instead of showing the overall progress of the recovered chain
  • The model_diagnostics command would fail to check the correctness of user-defined steady state files
  • GSA: LaTeX output was not working as expected
  • Forecasts and filtered variables could not be retrieved with the heteroskedastic_shocks block
  • The OccBin smoother would potentially not display all smoothed shocks with heteroskedastic_filter option
  • The OccBin smoother would crash if the number of requested periods was smaller than the data length
  • The multivariate OccBin smoother would return wrong results if the constraint was binding in the first period
  • The plot_shock_decomposition command would fail with the init2shocks block if the initial_condition_decomposition was not run before
  • LaTeX output under Windows failed to compile for plot_priors=1 option of the estimation command and Brooks and Gelman (1998) convergence diagnostics
  • The plot produced by the shock_decomposition command was too big, making the close button inaccessible
  • Monthly dates for October, November and December (i.e. with a 2-digit month number) were not properly interpreted by the preprocessor
  • Theoretical moments computed by stoch_simul at order=2 with pruning would not contain unconditional and conditional variance decomposition