Congratulations to the team for this major release full of new features! See you tomorrow for those who will be at the Summer School, Le dimanche 11 juin 2017 à 00:58 +0200, Stéphane Adjemian a écrit :
Dear Dynare users and friends,
We are pleased to announce the release of Dynare 4.5.0.
This major release adds new features and fixes various bugs.
The Windows packages are already available for download at:
http://www.dynare.org/download/dynare-stable
The Mac and Debian/Ubuntu packages should follow soon.
All users are strongly encouraged to upgrade.
This release is compatible with MATLAB versions ranging from 7.3 (R2006b) to 9.2 (R2017a) and with GNU Octave version 4.2.1.
Here is the list of major user-visible changes:
- Ramsey policy
+ Added command `ramsey_model` that builds the expanded model with FOC conditions for the planner's problem but doesn't perform any computation. Usefull to compute Ramsey policy in a perfect foresight model,
+ `ramsey_policy` accepts multipliers in its variable list and displays results for them.
- Perfect foresight models
+ New commands `perfect_foresight_setup` (for preparing the simulation) and `perfect_foresight_solver` (for computing it). The old `simul` command still exist and is now an alias for `perfect_foresight_setup` + `perfect_foresight_solver`. It is no longer possible to manipulate by hand the contents of `oo_.exo_simul` when using `simul`. People who want to do it must first call `perfect_foresight_setup`, then do the manipulations, then call `perfect_foresight_solver`,
+ By default, the perfect foresight solver will try a homotopy method if it fails to converge at the first try. The old behavior can be restored with the `no_homotopy` option,
+ New option `stack_solve_algo=7` that allows specifying a `solve_algo` solver for solving the model,
+ New option `solve_algo` that allows specifying a solver for solving the model when using `stack_solve_algo=7`,
+ New option `lmmcp` that solves the model via a Levenberg- Marquardt mixed complementarity problem (LMMCP) solver,
+ New option `robust_lin_solve` that triggers the use of a robust linear solver for the default `solve_algo=4`,
+ New options `tolf` and `tolx` to control termination criteria of solvers,
+ New option `endogenous_terminal_period` to `simul`,
+ Added the possibility to set the initial condition of the (stochastic) extended path simulations with the histval block.
- Optimal simple rules
+ Saves the optimal value of parameters to `oo_.osr.optim_params`,
+ New block `osr_params_bounds` allows specifying bounds for the estimated parameters,
+ New option `opt_algo` allows selecting different optimizers while the new option `optim` allows specifying the optimizer options,
+ The `osr` command now saves the names, bounds, and indices for the estimated parameters as well as the indices and weights of the variables entering the objective function into `M_.osr`.
- Forecasts and Smoothing
+ The smoother and forecasts take uncertainty about trends and means into account,
+ Forecasts accounting for measurement error are now saved in fields of the form `HPDinf_ME` and `HPDsup_ME`,
+ New fields `oo_.Smoother.Trend` and `oo_.Smoother.Constant` that save the trend and constant parts of the smoothed variables,
+ new field `oo_.Smoother.TrendCoeffs` that stores the trend coefficients.
+ Rolling window forecasts allowed in `estimation` command by passing a vector to `first_obs`,
+ The `calib_smoother` command now accepts the `loglinear`, `prefilter`, `first_obs` and `filter_decomposition` options.
- Estimation
+ New options: `logdata`, `consider_all_endogenous`, `consider_only_observed`, `posterior_max_subsample_draws`, `mh_conf_sig`, `diffuse_kalman_tol`, `dirname`, `nodecomposition`
+ `load_mh_file` and `mh_recover` now try to load chain's proposal density,
+ New option `load_results_after_load_mh` that allows loading some posterior results from a previous run if no new MCMC draws are added,
+ New option `posterior_nograph` that suppresses the generation of graphs associated with Bayesian IRFs, posterior smoothed objects, and posterior forecasts,
+ Saves the posterior density at the mode in `oo_.posterior.optimization.log_density`,
+ The `filter_covariance` option now also works with posterior sampling like Metropolis-Hastings,
+ New option `no_posterior_kernel_density` to suppress computation of kernel density of posterior objects,
+ Recursive estimation and forecasting now provides the individual `oo_` structures for each sample in `oo_recursive_`,
+ The `trace_plot` command can now plot the posterior density,
+ New command `generate_trace_plots` allows generating all trace plots for one chain,
+ New commands `prior_function` and `posterior_function` that execute a user-defined function on parameter draws from the prior/posterior distribution,
+ New option `huge_number` for replacement of infinite bounds with large number during `mode_compute`,
+ New option `posterior_sampling_method` allows selecting the new posterior sampling options: `tailored_random_block_metropolis_hastings` (Tailored randomized block (TaRB) Metropolis-Hastings), `slice` (Slice sampler), `independent_metropolis_hastings` (Independent Metropolis-Hastings),
+ New option `posterior_sampler_options` that allow controlling the options of the `posterior_sampling_method`, its `scale_file`- option pair allows loading the `_mh_scale.mat`-file storing the tuned scale factor from a previous run of `mode_compute=6`,
+ New option `raftery_lewis_diagnostics` that computes Raftery/Lewis (1992) convergence diagnostics,
+ New option `fast_kalman_filter` that provides fast Kalman filter using Chandrasekhar recursions as described in Ed Herbst (2015),
+ The `dsge_var` option now saves results at the posterior mode into `oo_.dsge_var`,
+ New option `smoothed_state_uncertainty` to provide the uncertainty estimate for the smoothed state estimate from the Kalman smoother,
+ New prior density: generalized Weibull distribution,
+ Option `mh_recover` now allows continuing a crashed chain at the last save mh-file,
+ New option `nonlinear_filter_initialization` for the `estimation` command. Controls the initial covariance matrix of the state variables in nonlinear filters.
+ The `conditional_variance_decomposition` option now displays output and stores it as a LaTeX-table when the `TeX` option is invoked,
+ The `use_calibration` to `estimated_params_init` now also works with ML,
+ Improved initial estimation checks.
- Steady state
+ The default solver for finding the steady state is now a trust-region solver (can be triggered explicitly with option `solve_algo=4`),
+ New options `tolf` and `tolx` to control termination criteria of solver,
+ The debugging mode now provides the termination values in steady state finding.
- Stochastic simulations
+ New options `nodecomposition`,
+ New option `bandpass_filter` to compute bandpass-filtered theoretical and simulated moments,
+ New option `one_sided_hp_filter` to compute one-sided HP- filtered simulated moments,
+ `stoch_simul` displays a simulated variance decomposition when simulated moments are requested,
+ `stoch_simul` saves skewness and kurtosis into respective fields of `oo_` when simulated moments have been requested,
+ `stoch_simul` saves the unconditional variance decomposition in `oo_.variance_decomposition`,
+ New option `dr_display_tol` that governs omission of small terms in display of decision rules,
+ The `stoch_simul` command now prints the displayed tables as LaTeX code when the new `TeX` option is enabled,
+ The `loglinear` option now works with lagged and leaded exogenous variables like news shocks,
+ New option `spectral_density` that allows displaying the spectral density of (filtered) endogenous variables,
+ New option `contemporaneous_correlation` that allows saving contemporaneous correlations in addition to the covariances.
- Identification
+ New options `diffuse_filter` and `prior_trunc`,
+ The `identification` command now supports correlations via simulated moments,
- Sensitivity analysis
+ New blocks `irf_calibration` and `moment_calibration`,
+ Outputs LaTeX tables if the new `TeX` option is used,
+ New option `relative_irf` to `irf_calibration` block.
- Conditional forecast
+ Command `conditional_forecast` now takes into account `histval` block if present.
- Shock decomposition
+ New option `colormap` to `shocks_decomposition` for controlling the color map used in the shocks decomposition graphs,
+ `shocks_decomposition` now accepts the `nograph` option,
+ New command `realtime_shock_decomposition` that for each period `T= [presample,...,nobs]` allows computing the:
* realtime historical shock decomposition `Y(t|T)`, i.e. without observing data in `[T+1,...,nobs]`
* forecast shock decomposition `Y(T+k|T)`
* realtime conditional shock decomposition `Y(T+k|T+k)-Y(T+k|T)`
+ New block `shock_groups` that allows grouping shocks for the `shock_decomposition` and `realtime_shock_decomposition` commands,
+ New command `plot_shock_decomposition` that allows plotting the results from `shock_decomposition` and `realtime_shock_decomposition` for different vintages and shock groupings.
- Macroprocessor
+ Can now pass a macro-variable to the `@#include` macro directive,
+ New preprocessor flag `-I`, macro directive `@#includepath`, and dynare config file block `[paths]` to pass a search path to the macroprocessor to be used for file inclusion via `@#include`.
- Command line
+ New option `onlyclearglobals` (do not clear JIT compiled functions with recent versions of Matlab),
+ New option `minimal_workspace` to use fewer variables in the current workspace,
+ New option `params_derivs_order` allows limiting the order of the derivatives with respect to the parameters that are calculated by the preprocessor,
+ New command line option `mingw` to support the MinGW-w64 C/C++ Compiler from TDM-GCC for `use_dll`.
- dates/dseries/reporting classes
+ New methods `abs`, `cumprod` and `chain`,
+ New option `tableRowIndent` to `addTable`,
+ Reporting system revamped and made more efficient, dependency on matlab2tikz has been dropped.
- Optimization algorithms
+ `mode_compute=2` Uses the simulated annealing as described by Corana et al. (1987),
+ `mode_compute=101` Uses SOLVEOPT as described by Kuntsevich and Kappel (1997),
+ `mode_compute=102` Uses `simulannealbnd` from Matlab's Global Optimization Toolbox (if available),
+ New option `silent_optimizer` to shut off output from mode computing/optimization,
+ New options `verbosity` and `SaveFiles` to control output and saving of files during mode computing/optimization.
- LaTeX output
+ New command `write_latex_original_model`,
+ New option `write_equation_tags` to `write_latex_dynamic_model` that allows printing the specified equation tags to the generate LaTeX code,
+ New command `write_latex_parameter_table` that writes the names and values of model parameters to a LaTeX table,
+ New command `write_latex_prior_table` that writes the descriptive statistics about the prior distribution to a LaTeX table,
+ New command `collect_latex_files` that creates one compilable LaTeX file containing all TeX-output.
- Misc.
+ Provides 64bit preprocessor,
+ Introduces new path management to avoid conflicts with other toolboxes,
+ Full compatibility with Matlab 2014b's new graphic interface,
+ When using `model(linear)`, Dynare automatically checks whether the model is truly linear,
+ `usedll`, the `msvc` option now supports `normcdf`, `acosh`, `asinh`, and `atanh`,
+ New parallel option `NumberOfThreadsPerJob` for Windows nodes that sets the number of threads assigned to each remote MATLAB/Octave run,
+ Improved numerical performance of `schur_statespace_transformation` for very large models,
+ The `all_values_required` option now also works with `histval`,
+ Add missing `horizon` option to `ms_forecast`,
+ BVAR now saves the marginal data density in `oo_.bvar.log_marginal_data_density` and stores prior and posterior information in `oo_.bvar.prior` and `oo_.bvar.posterior`.
- Bugs and problems identified in version 4.4.3 and that have been
fixed in version 4.5.0:
- BVAR models
+ `bvar_irf` could display IRFs in an unreadable way when they moved from negative to positive values,
+ In contrast to what is stated in the documentation, the confidence interval size `conf_sig` was 0.6 by default instead of 0.9.
- Conditional forecasts
+ The `conditional_forecast` command produced wrong results in calibrated models when used at initial values outside of the steady state (given with `initval`),
+ The `plot_conditional_forecast` option could produce unreadable figures if the areas overlap,
+ The `conditional_forecast` command after MLE crashed,
+ In contrast to what is stated in the manual, the confidence interval size `conf_sig` was 0.6 by default instead of 0.8.
+ Conditional forecasts were wrong when the declaration of endogenous variables was not preceeding the declaration of the exogenous variables and parameters.
- Discretionary policy
+ Dynare allowed running models where the number of instruments did not match the number of omitted equations,
+ Dynare could crash in some cases when trying to display the solution,
+ Parameter dependence embedded via a `steady_state` was not taken into account, typically resulting in crashes.
- dseries class
+ When subtracting a dseries object from a number, the number was instead subtracted from the dseries object.
- DSGE-VAR models
+ Dynare crashed when estimation encountered non-finite values in the Jacobian at the steady state,
+ The presence of a constant was not considered for degrees of freedom computation of the Gamma function used during the posterior computation; due to only affecting the constant term, results should be be unaffected, except for model_comparison when comparing models with and without.
- Estimation command
+ In contrast to what was stated in the manual, the confidence interval size `conf_sig` for `forecast` without MCMC was 0.6 by default instead of 0.9,
+ Calling estimation after identification could lead to crashes,
+ When using recursive estimation/forecasting and setting some elements of `nobs` to be larger than the number of observations T in the data, `oo_recursive_` contained additional cell entries that simply repeated the results obtained for `oo_recursive_T`,
+ Computation of Bayesian smoother could crash for larger models when requesting `forecast` or `filtered_variables`,
+ Geweke convergence diagnostics were not computed on the full MCMC chain when the `load_mh_file` option was used,
+ The Geweke convergence diagnostics always used the default `taper_steps` and `geweke_interval`,
+ Bayesian IRFs (`bayesian_irfs` option) could be displayed in an unreadable way when they move from negative to positive values,
+ If `bayesian_irfs` was requested when `mh_replic` was too low to compute HPDIs, plotting was crashing,
+ The x-axis value in `oo_.prior_density` for the standard deviation and correlation of measurement errors was written into a field `mearsurement_errors_*` instead of `measurement_errors_*`,
+ Using a user-defined `mode_compute` crashed estimation,
+ Option `mode_compute=10` did not work with infinite prior bounds,
+ The posterior variances and covariances computed by `moments_varendo` were wrong for very large models due to a matrix erroneously being filled up with zeros,
+ Using the `forecast` option with `loglinear` erroneously added the unlogged steady state,
+ When using the `loglinear` option the check for the presence of a constant was erroneously based on the unlogged steady state,
+ Estimation of `observation_trends` was broken as the trends specified as a function of deep parameters were not correctly updated during estimation,
+ When using `analytic_derivation`, the parameter values were not set before testing whether the steady state file changes parameter values, leading to subsequent crashes,
+ If the steady state of an initial parameterization did not solve, the observation equation could erroneously feature no constant when the `use_calibration` option was used,
+ When computing posterior moments, Dynare falsely displayed that moment computations are skipped, although the computation was performed correctly,
+ If `conditional_variance_decomposition` was requested, although all variables contain unit roots, Dynare crashed instead of providing an error message,
+ Computation of the posterior parameter distribution was erroneously based on more draws than specified (there was one additional draw for every Markov chain),
+ The estimation option `lyapunov=fixed_point` was broken,
+ Computation of `filtered_vars` with only one requested step crashed Dynare,
+ Option `kalman_algo=3` was broken with non-diagonal measurement error,
+ When using the diffuse Kalman filter with missing observations, an additive factor log(2*pi) was missing in the last iteration step,
+ Passing of the `MaxFunEvals` and `InitialSimplexSize` options to `mode_compute=8` was broken,
+ Bayesian forecasts contained initial conditions and had the wrong length in both plots and stored variables,
+ Filtered variables obtained with `mh_replic=0`, ML, or `calibrated_smoother` were padded with zeros at the beginning and end and had the wrong length in stored variables,
+ Computation of smoothed measurement errors in Bayesian estimation was broken,
+ The `selected_variables_only` option (`mh_replic=0`, ML, or `calibrated_smoother`) returned wrong results for smoothed, updated, and filtered variables,
+ Combining the `selected_variables_only` option with forecasts obtained using `mh_replic=0`, ML, or `calibrated_smoother` leaded to crashes,
+ `oo_.UpdatedVariables` was only filled when the `filtered_vars` option was specified,
+ When using Bayesian estimation with `filtered_vars`, but without `smoother`, then `oo_.FilteredVariables` erroneously also contained filtered variables at the posterior mean as with `mh_replic=0`,
+ Running an MCMC a second time in the same folder with a different number of iterations could result in crashes due to the loading of stale files,
+ Results displayed after Bayesian estimation when not specifying the `smoother` option were based on the parameters at the mode from mode finding instead of the mean parameters from the posterior draws. This affected the smoother results displayed, but also calls to subsequent command relying on the parameters stored in `M_.params` like `stoch_simul`,
+ The content of `oo_.posterior_std` after Bayesian estimation was based on the standard deviation at the posterior mode, not the one from the MCMC, this was not consistent with the reference manual,
+ When the initialization of an MCMC run failed, the metropolis.log file was locked, requiring a restart of Matlab to restart estimation,
+ If the posterior mode was right at the corner of the prior bounds, the initialization of the MCMC erroneously crashed,
+ If the number of dropped draws via `mh_drop` coincided with the number of draws in a `_mh'-file`, `oo_.posterior.metropolis.mean` and `oo_.posterior.metropolis.Variance` were NaN.
- Estimation and calibrated smoother
+ When using `observation_trends` with the `prefilter` option, the mean shift due to the trend was not accounted for,
+ When using `first_obs`>1, the higher trend starting point of `observation_trends` was not taken into account, leading, among other things, to problems in recursive forecasting,
+ The diffuse Kalman smoother was crashing if the forecast error variance matrix becomes singular,
+ The multivariate Kalman smoother provided incorrect state estimates when all data for one observation are missing,
+ The multivariate diffuse Kalman smoother provided incorrect state estimates when the `Finf` matrix becomes singular,
+ The univariate diffuse Kalman filter was crashing if the initial covariance matrix of the nonstationary state vector is singular,
- Forecats
+ In contrast to what is stated in the manual, the confidence interval size `conf_sig` was 0.6 by default instead of 0.9.
+ Forecasting with exogenous deterministic variables provided wrong decision rules, yielding wrong forecasts.
+ Forecasting with exogenous deterministic variables crashed when the `periods` option was not explicitly specified,
+ Option `forecast` when used with `initval` was using the initial values in the `initval` block and not the steady state computed from these initial values as the starting point of forecasts.
- Global Sensitivity Analysis
+ Sensitivity with ML estimation could result in crashes,
+ Option `mc` must be forced if `neighborhood_width` is used,
+ Fixed dimension of `stock_logpo` and `stock_ys`,
+ Incomplete variable initialization could lead to crashes with `prior_range=1`.
- Indentification
+ Identification did not correctly pass the `lik_init` option, requiring the manual setting of `options_.diffuse_filter=1` in case of unit roots,
+ Testing identification of standard deviations as the only parameters to be estimated with ML leaded to crashes,
+ Automatic increase of the lag number for autocovariances when the number of parameters is bigger than the number of non-zero moments was broken,
+ When using ML, the asymptotic Hessian was not computed,
+ Checking for singular values when the eigenvectors contained only one column did not work correctly,
- Model comparison
+ Selection of the `modifiedharmonicmean` estimator was broken,
- Optimal Simple Rules
+ When covariances were specified, variables that only entered with their variance and no covariance term obtained a wrong weight, resulting in wrong results,
+ Results reported for stochastic simulations after `osr` were based on the last parameter vector encountered during optimization, which does not necessarily coincide with the optimal parameter vector,
+ Using only one (co)variance in the objective function resulted in crashes,
+ For models with non-stationary variables the objective function was computed wrongly.
- Ramsey policy
+ If a Lagrange multiplier appeared in the model with a lead or a lag of more than one period, the steady state could be wrong.
+ When using an external steady state file, incorrect steady states could be accepted,
+ When using an external steady state file with more than one instrument, Dynare crashed,
+ When using an external steady state file and running `stoch_simul` after `ramsey_planner`, an incorrect steady state was used,
+ When the number of instruments was not equal to the number of omitted equations, Dynare crashed with a cryptic message,
+ The `planner_objective` accepted `varexo`, but ignored them for computations,
- Shock decomposition
+ Did not work with the `parameter_set=calibration` option if an `estimated_params` block is present,
+ Crashed after MLE.
- Perfect foresight models
+ The perfect foresight solver could accept a complex solution instead of continuing to look for a real-valued one,
+ The `initval_file` command only accepted column and not row vectors,
+ The `initval_file` command did not work with Excel files,
+ Deterministic simulations with one boundary condition crashed in `solve_one_boundary` due to a missing underscore when passing `options_.simul.maxit`,
+ Deterministic simulation with exogenous variables lagged by more than one period crashed,
+ Termination criterion `maxit` was hard-coded for `solve_algo=0` and could no be changed,
+ When using `block`/`bytecode`, relational operators could not be enforced,
+ When using `block` some exceptions were not properly handled, leading to code crashes,
+ Using `periods=1` crashed the solver (bug only partially fixed).
- Smoothing
+ The univariate Kalman smoother returned wrong results when used with correlated measurement error,
+ The diffuse smoother sometimes returned linear combinations of the smoothed stochastic trend estimates instead of the original trend estimates.
- Perturbation reduced form
+ In contrast to what is stated in the manual, the results of the unconditional variance decomposition were only stored in `oo_.gamma_y(nar+2)`, not in `oo_.variance_decomposition`,
+ Dynare could crash when the steady state could not be computed when using the `loglinear` option,
+ Using `bytcode` when declared exogenous variables were not used in the model leaded to crashes in stochastic simulations,
+ Displaying decision rules involving lags of auxiliary variables of type 0 (leads>1) crashed.
+ The `relative_irf` option resulted in wrong output at `order>1` as it implicitly relies on linearity.
- Displaying of the MH-history with the `internals` command crashed if parameter names did not have same length.
- Dynare crashed when the user-defined steady state file returned an error code, but not an conformable-sized steady state vector.
- Due to a bug in `mjdgges.mex` unstable parameter draws with eigenvalues up to 1+1e-6 could be accepted as stable for the purpose of the Blanchard-Kahn conditions, even if `qz_criterium<1`.
- The `use_dll` option on Octave for Windows required to pass a compiler flag at the command line, despite the manual stating this was not necessary.
- Dynare crashed for models with `block` option if the Blanchard- Kahn conditions were not satisfied instead of generating an error message.
- The `verbose` option did not work with `model(block)`.
- When falsely specifying the `model(linear)` for nonlinear models, incorrect steady states were accepted instead of aborting.
- The `STEADY_STATE` operator called on model local variables (so-called pound variables) did not work as expected.
- The substring operator in macro-processor was broken. The characters of the substring could be mixed with random characters from the memory space.
- Block decomposition could sometimes cause the preprocessor to crash.
- A bug when external functions were used in model local variables that were contained in equations that required auxiliary variable/equations led to crashes of Matlab.
- Sampling from the prior distribution for an inverse gamma II distribution when `prior_trunc>0` could result in incorrect sampling.
- Sampling from the prior distribution for a uniform distribution when `prior_trunc>0` was ignoring the prior truncation.
- Conditional forecasts were wrong when the declaration of endogenous variables was not preceeding the declaration of the exogenous variables and parameters.
On behalf of the Dynare Team, Stéphane.
--
Sébastien Villemot
Économiste
Sciences Po, OFCE