New Estimation Interface

Suggestions for a simplification of the estimation interface in Dynare and breaking very general commands with lots of options in several commands with more targeted scope.

set_time

This command allows to specify an initial date and a frequency for the data used by dynare or the data generated by dynare (this command is noit specific to estimation). There is only one input argument. The frequency is implicit.

Syntax examples

If the initial period is the third quarter of 1950 (quaterly data):

   set_time(1950Q3)

If the initial period is 1960 (yearly data implicitely assumed):

   set_time(1960)

If the initial period is the fifth month of 1971 (monthly data):

   set_time(1971M5)

Preprocessor generated code

Instantiate an object from the dynDates class. This object, say initial_period, is stored in options_. If the command set_time is absent, default initial period is 1 (with yearly frequency). For the preceeding examples, the preprocessor should write in the main matlab file:

   options_.initial_period = dynDate('1950Q3');

   options_.initial_period = dynDate('1960');

   options_.initial_period = dynDate('1971M5');

the default being

   options_.initial_period = dynDate(1);

dataset

This command initializes the dataset (an object instantiated by dynSeries). The following options are available:

file
first_obs
last_obs
nobs
xls_sheet
xls_range

Syntax examples

Call with only one input argument:

   data(file=/home/stepan/Works/MZE/data/data.m)

   data(file=/home/stepan/Works/MZE/data/data.mat)

   data(file=/home/stepan/Works/MZE/data/data.xls)

In this case the implicit initial date is given by the set_time command, and all the observations are used. Input argument file is mandatory.

Call with more than one input argument:

   data(file=/home/stepan/Works/MZE/data/data.xls,first_obs=1950Q1)

In this first example the beggining of the sample is specified by first_obs. By default, all the observations are used after this initial date. The frequency (quaterly, monthly or yearly data) is implicit. If the implicit frequency in the command data is inconsistent with the implicit frequency declared in set_time an error message must be issued by the preprocessor. If the initial date is anterior to the date specified by set_time, an error will be issued by the matlab code (using the dynDates class).

   data(file=/home/stepan/Works/MZE/data/data.xls,first_obs=1950Q1,last_obs=2000Q4)

   data(file=/home/stepan/Works/MZE/data/data.xls,first_obs=1950Q1,nobs=204)

In these two examples the last arguments (nobs or last_obs) are used to specify the last observation in the sample. If the implicit frequency is inconsistent with the implicit frequency declared in set_time or if nobs is not strictly positive, an error must be issued by the preprocessor.

Preprocessor generated code

All the options in the data command are saved as fields of options_.dataset. For the previous examples we would have:

   options_.dataset.file = '/home/stepan/Works/MZE/data/data.m';
   options_.dataset.first_obs = dynDate('1950Q1');
   options_.dataset.last_obs = dynDate('2000Q4');
   options_.dataset.nobs = 204;

The default values (written by the preprocessor, not defined in global_initialization) are:

   options_.dataset.first_obs = options_.initial_period;
   options_.dataset.last_obs = NaN;
   options_.dataset.nobs = NaN;
   options_.dataset.xls_sheet = NaN;
   options_.dataset.xls_range = NaN;

estimated_params

Rely on options names rather than argument position in the command line. This new interface would replace estimated_params_bounds and estimated_params_init. The following option names would be necessary

For the declaration of the priors
- shape*, name of the prior distribution.
- mean*, mean of the prior distribution.
- median, median of the prior distribution.
- truncate, bounds of the prior distribution.
- stdev*, standard deviation of the prior distribution.
- variance*, variance of the prior distribution.
- mode*, mode of the prior distribution.
- interval, specification of the prior distribution from the cumulative distribution function defining an interval covering an arbitrary percentage of the prior mass. See below for an example.
- shift, to shift the domain of a distribution. See below for an example (useless if the domain of the distribution is the set of real numbers).
- domain*, specify the domain of distributions with bounded support
For the declaration of the priors for a MS-DSGE model, in addition to the starred entries above
- regimes, provide a vector of regimes (ordered by chain order) for which the prior is being specified
For optimization options (starred entry to be used for MS-DSGE)
- bounds, enforce the optimizer to look for a mode in a specified interval.
- init*, initial value for the optimizer.
For the mcmc options
- jscale, scale paramemeter for the jumping distribution specific to a parameter.

Syntax examples

The following instructions given as examples are to be used within the estimated_params block.

Declaration of a prior for estimated parameter alpha, specified using mean and variance:

     alpha.prior(shape=beta,mean=0.3,variance=0.1^2);
     beta.prior = alpha.prior;

Declaration of a prior for estimated parameter alpha, specified using the cumulative distribution function:

     alpha.prior(shape=beta,interval=[0.2,0.4,.9]);

with this syntax the prior is such that 90% of the prior mass lies between 0.2 and 0.4.

Shift the domain of a prior distribution:

     beta.prior(shape=gamma,mode=4,variance=10,shift=+2);
     sigma.prior(shape=gamma,mode=-1,variance=5,shift=-2);

In the first case the default domain of the gamma distribution (the set of real positive numbers) is shifted towards infinity in the second case the same distribution is shifted towards minus infinity.

Specify the (non default) domain of a prior distribution with bounded support:

     plouf.prior(shape=uniform,domain=[.5,2]);
     fuolp.prior(shape=beta,mode=1.5,stdev=.05,domain=[1,2]);

The default domain for the beta and uniform distribution is [0,1].

Set specific optimization and mcmc options

     alpha.options(bounds=[0 2],init=1,jscale=.1);
     beta.options(bounds=[0 1],init=.5,jscale=.3);
     sigma.options(init=.7,jscale=.1);

Declaration of the subsamples (must preceed the declaration of the priors):

     alpha.subsamples(name1=1950Q3:1957Q4, name2=1958Q1:1983Q2, name3=1983Q3:2011Q2);
     std(e).subsamples(name1=1950Q3:1957Q4, name2=1958Q1:1983Q2, name3=1983Q3:2011Q2);
     corr(e,u).subsamples(name1=1950Q3:1957Q4, name2=1958Q1:1983Q2, name3=1983Q3:2011Q2);
     beta.subsamples = alpha.subsamples;
     std(u).subsamples = beta.subsamples;

Declaration of alpha's prior over different subsamples (must follow the declaration of the subsamples):

     alpha.name1.prior(shape=normal,mode=0.30,stdev=.01);
     alpha.name2.prior(shape=normal,mode=0.33,stdev=.01);
     alpha.name3.prior(shape=normal,mode=0.40,stdev=.01);
     beta.name1.prior(shape=normal,mode=0.30,stdev=.01);
     beta.name2.prior(shape=normal,mode=0.33,stdev=.01);     
     beta.name3.prior(shape=normal,mode=0.40,stdev=.01);

Declaration of alpha's prior over certain regimes would be specified as:

     alpha.prior(shape=normal, mode=0.30, stdev=.01, regimes=[1,1,ALL]);
     alpha.prior(shape=normal, mode=0.33, stdev=.01, regimes=[1,2,ALL]);
     alpha.prior(shape=normal, mode=0.33, stdev=.01, regimes=[1,3,ALL]);

where the ordering of the entries in the argument to regimes is the same as the markov chain ordering (forced to be consecutive, positive integers indexed at 1).

Declaration of a prior for the estimated standard deviation of a structural shock e, specified using mean and variance:

     std(e).prior(shape=beta,mean=0.3,variance=0.1^2);

e must be declared as an exogenous variable (varexo).

Declaration of a prior for the estimated correlation between two structural shocks e and u, specified using mean and variance:

     corr(e,u).prior(shape=beta,mean=0.3,variance=0.1^2);

e and u must be declared as an exogenous variable (varexo).

Declaration of a prior for the estimated standard deviation of a measurement error on a variable y, specified using mean and variance:

     std(y).prior(shape=beta,mean=0.3,variance=0.1^2);

y is an endogenous variable and must be declared as an observed variable (varobs).

Declaration of a prior for the estimated correlation between two measurement errors on y and c, specified using mean and variance:

     corr(y,c).prior(shape=beta,mean=0.3,variance=0.1^2);

y and c are endogenous variables and must be declared as observed variables (varobs).

Preprocessor generated Matlab code

The preprocessor fills a new matlab structure called estimation_info. This structure is organized as follows:

The first field is a structure specifying the parameters to be estimated.

  estimation_info.parameters.deep.nb = 3;
  estimation_info.parameters.list = char('alpha','beta','sigma')
  estimation_info.parameters.structural_innovation.nb = 2;
  estimation_info.parameters.structural_innovation.list = ĉhar('e','u');
  estimation_info.parameters.structural_innovation_corr.nb = 1;
  estimation_info.parameters.structural_innovation_corr.list = char('e___u');
  estimation_info.parameters.measurement_error.nb = 2;
  estimation_info.parameters.measurement_error.list = ĉhar('y','c');
  estimation_info.parameters.measurement_error_corr.nb = 1;
  estimation_info.parameters.measurement_error_corr.list = char('y___c');

If the some parameters are not stable across the whole sample, these parameters are specific to each subsample. For instance if we estimate the model with one structual break on parameter alpha, we would have instead:

  estimation_info.parameters.deep.nb = 5;
  estimation_info.parameters.list = char('alpha$1','alpha$2','alpha$3','beta','sigma');
  estimation_info.parameters.structural_innovation.nb = 2;
  estimation_info.parameters.structural_innovation.list = ĉhar('e','u');
  estimation_info.parameters.structural_innovation_corr.nb = 1;
  estimation_info.parameters.structural_innovation_corr.list = char('e___u');
  estimation_info.parameters.measurement_error.nb = 2;
  estimation_info.parameters.measurement_error.list = ĉhar('y','c');
  estimation_info.parameters.measurement_error_corr.nb = 1;
  estimation_info.parameters.measurement_error_corr.list = char('y___c');

where alpha$1, alpha$2 and alpha$3 correspond to alpha for the first, second and third subsamples.

The second field is a structure specifying the initial conditions for the optimization routine.

  estimation_info.init.deep = [1.0; .5; .7 ];
  estimation_info.init.structural_innovation = [.01; .05];
  estimation_info.init.structural_innovation_corr = [.1];
  estimation_info.init.measurement_error = [.01; .07];
  estimation_info.init.measurement_error_corr = [-.1];

or with the structural break on alpha:

  estimation_info.init.deep = [1.0; 1.1; .9; .5; .7 ];
  estimation_info.init.structural_innovation = [.01; .05];
  estimation_info.init.structural_innovation_corr = [.1];
  estimation_info.init.measurement_error = [.01; .07];
  estimation_info.init.measurement_error_corr = [-.1];

If there is no estimated measurement errors in the model (ie estimation_info.parameters.measurement_error.nb=0) the estimation_info.init.measurement_error and estimation_info.init.measurement_error_corr are set to a empty matrices.

The third field is a structure specifying the bounds for the optimization routine.

  estimation_info.bounds.deep = [0, 2; 0, 1; -Inf, Inf];
  estimation_info.bounds.structural_innovation = [0, .01; 0, Inf];
  estimation_info.bounds.structural_innovation_corr = [-1 1];
  estimation_info.bounds.measurement_error = [0 Inf; 0 Inf];
  estimation_info.bounds.measurement_error_corr = [-1 1];

Default bounds for estimated standard deviations or variance are 0 and Inf. Default bounds for estimated correlations are -1 and 1. Default bounds for estimated deep parameters are -Inf and Inf (meaning no effective bounds for the optimization).

The fourth field is a structure describing the subsamples (if any). By default each estimated parameter is kept constant along the whole sample. For instance, if there is one unexpected break on the estimated parameter alpha, we would have:

  estimation_info.subsample.deep.alpha = [ '1950:Q3', '1957:Q4'; '1958Q1', '1983:Q2'; '1983Q3', '2011Q2'];

If there is no unexpected break on the estimated deep parameters then estimation_info.subsample.deep is an empty matrix.

An optional field is added if the model is estimated with a Bayesian approach. This field specifies the prior beliefs. For X = 'deep', 'structural_innovation', 'measurement_error', 'structural_innovation_corr' and 'measurement_error_corr' we have the following fields (n*1 vectors or cell arrays of doubles or n*2 vectors or cell arrays of doubles):

   estimation_info.prior.shape.X
   estimation_info.prior.mean.X
   estimation_info.prior.stdev.X
   estimation_info.prior.variance.X
   estimation_info.prior.mode.X
   estimation_info.prior.interval.X
   estimation_info.prior.shift.X
   estimation_info.prior.domain.X
   estimation_info.prior.jscale.X

To provide the structures above, the preprocessor will first generate code in the form:

   estimation_info.parameter(i).prior.name = alpha;
   estimation_info.parameter(i).prior.shape = 2;
   estimation_info.parameter(i).prior.mean = 3;
   ...
   estimation_info.structural_innovation(i).prior.name = e;
   estimation_info.structural_innovation(i).prior.shape = 2;
   estimation_info.structural_innovation(i).prior.mean = 0.3;
   ...
   estimation_info.measurement_error(i).prior.name = y;
   estimation_info.measurement_error(i).prior.shape = 2;
   estimation_info.measurement_error(i).prior.mean = 0.3;
   ...
   estimation_info.structural_innovation_corr(i).prior.name1 = e;
   estimation_info.structural_innovation_corr(i).prior.name2 = u;
   estimation_info.structural_innovation_corr(i).prior.shape = 2;
   estimation_info.structural_innovation_corr(i).prior.mean = 0.3;
   ...
   estimation_info.measurement_error_corr(i).prior.name1 = y;
   estimation_info.measurement_error_corr(i).prior.name2 = c;
   estimation_info.measurement_error_corr(i).prior.shape = 2;
   estimation_info.measurement_error_corr(i).prior.mean = 0.3;

This format is easier for the user to understand if she would like to modify the <modfilename>.m file or programmatically create her own process for interacting with the Dynare Matlab code. Further, self-contained structures are better organized than the parallel arrays mentioned above. A routine will be created to transform this input into the aforementioned parallel array format, which is easier to use in the backend.

Preprocessor generated C code for MS-DSGE

As C++ output needs to be produced for the MS-DSGE code, the prior declaration will be stored as an object of the Prior Class, StdPrior Class or the CorrPrior Class, depending on the form the prior statement took (param.prior(...), std(exo/endo var).prior(...) or corr(endo/exo,endo/exo).prior(...)). In addition to declaring one of the aforementioned objects for each of the prior statements and organizing them into vectors, the following will also be provided by the preprocessor:

a C++ Map, with paramater names as keys and parameter indices as values
a Markov_Switching object for every markov_switching(...) statement (See MarkovSwitchingInterface)
NB: shape will be passed as a string, not as a number. This means that the current method of translating prior distributions into numbers in the Bison file will not work. Rather, a separate prior shape class (or at the very least, a Map available at the Statement level) should be created that can translate between shapes and their associated numbers.