init_parameters! not working about dsge.jl HOT 8 CLOSED

Hi caioodantas,

The problem is that when declaring parameters, you have to use Float64 for all of the numbers. For example, the parameter gamma is currently initialized as

m <= parameter(:γ, 1, (1e-3, 10), (1e-3, 10), ModelConstructors.Exponential(), RootInverseGamma(1, 0.75), fixed=false,
                   description="γ: Risk aversion.",
                   tex_label="\\gamma")

but the parameter value should be 1., or 1.0. Using 1 gives the parameter an Int64, which does not satisfy the type restrictions of a Parameter type (we need Float64 everywhere to ensure type stability for MH or SMC sampling). Similarly, the bounds should be (1e-3, 10.), the transform parameterization should be (1e-3, 10.), and the prior should be RootInverseGamma(1., 0.75). Additionally note that the second tuple is not a value bound but a transform parameterization, i.e. it parameterizes the Exponential transform applied when finding the mode using csminwel. If the transform applied was either the SquareRoot or Untransformed transformations, then the parameters are actually just the value bounds, but in the Exponential case, the parameters are not the value bounds (I understand this may be a confusing point. I've added this subtlety to the docket for things to add to our documentation). However, it is typically fine to use the same parameters as the value bounds (it just works out that way), but you should still see the docstring for transform_to_model_space here.

You should also make sure to copy the observables.jl, pseudo_observables.jl, and init_subspec.jl files from AnSchorfheide and replace the type assertions to XGabaix in those files. Otherwise, your XGabaix will not instantiate without error. Once you do these four things, it should work (or at least it did on my machine).

Cool idea by the way! Would be interesting to see a full-blown Bayesian estimation of Gabaix's behavioral model. Definitely let us know if you make headway on estimating this model, and we can try to highlight it as an interesting application of the package.

Best,
William

from dsge.jl.

Ah right, you'll want to set the transform parameterization to (1e-2, 1e-5) so that you don't accidentally hit any bounds due to numerical imprecision. The issue is happening b/c after making a new guess in R, csminwel then transforms the guess back into model space. Depending on the transform's parameterization, you'll get values that are outside the permissible value bounds. An alternative you could try to ensure you don't run into this problem is the SquareRoot transform, whose transformation parameters are just the value bounds, which makes it straightforward to ensuring you don't hit the value bounds.

As an example, if you use (0., 0.), then log(-Inf) = 0., which is below the lower bound of 1e-3. So you want a >= 1e-3.

from dsge.jl.

Can you show me the entire gabaix.jl script please? It'll be easier for me to diagnose what's going on that way.

from dsge.jl.

Thank you very much.
https://github.com/caioodantas/Behavioral-New-Keynesian-Model/blob/main/gabaix.jl

from dsge.jl.

Thanks William, your suggestion did fix the issue with init_parameters. I'm now having trouble with the transform parametrization.

ERROR: LoadError: New value of γ (11.000045391546594) is out of bounds ((0.0010454453523997807, 11.0))

I tried to fix it after reading the transform_to_model_space docstring, as you suggested, but I seem to have gotten something wrong.

m <= parameter(:γ, 1., (1e-3, 10.), (1e-3 + exp(1e-3 - 10.), 10. + exp(10. - 10.)), ModelConstructors.Exponential(), RootInverseGamma(1., 0.75), fixed=false, description="γ: Risk aversion.", tex_label="\\gamma")

from dsge.jl.

When are you getting this error? Also the transform parameterization for the exponential takes the form (a, b), where a and b DO NOT correspond to the value bounds at all. They are now just parameters for taking x in (0., Inf) and mapping to R, namely f(x; a, b) = a + exp( x - b). So the standard exponential transform has parameters (0., 0.), corresponding to f(x; 0., 0.) = exp(x).

from dsge.jl.

The error occurs right after
Recalculating Hessian... Hessian element: (1, 1)

I tried using parameters (0., 0.), but the error persists.

ERROR: LoadError: New value of γ (10.000045399205783) is out of bounds ((0.001, 10.0))

Stacktrace: [1] parameter(::UnscaledParameter{Float64,Exponential}, ::Float64) at C:\Users\Samsung\.julia\packages\ModelConstructors\6QJpO\src\parameters.jl:553 [2] map!(::typeof(parameter), ::Array{AbstractParameter{Float64},1}, ::Array{AbstractParameter{Float64},1}, ::Array{Float64,1}) at .\abstractarray.jl:2197 [3] update!(::Array{AbstractParameter{Float64},1}, ::Array{Float64,1}; change_value_type::Bool) at C:\Users\Samsung\.julia\packages\ModelConstructors\6QJpO\src\parameters.jl:1024 [4] update! at C:\Users\Samsung\.julia\packages\ModelConstructors\6QJpO\src\parameters.jl:1013 [inlined] [5] update! at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\abstractdsgemodel.jl:417 [inlined] [6] posterior!(::XGabaix{Float64}, ::Array{Float64,1}, ::Array{Float64,2}; sampler::Bool, ϕ_smc::Float64, catch_errors::Bool) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\posterior.jl:88 [7] posterior! at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\posterior.jl:76 [inlined] [8] (::DSGE.var"#f_hessian#294"{XGabaix{Float64},Array{Float64,2},Array{Float64,1}})(::Array{Float64,1}) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\hessian.jl:37 [9] hess_diag_element(::DSGE.var"#f_hessian#294"{XGabaix{Float64},Array{Float64,2},Array{Float64,1}}, ::Array{Float64,1}, ::Int64; ndx::Int64, check_neg_diag::Bool, verbose::Symbol) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\hessizero.jl:112 [10] hessizero(::Function, ::Array{Float64,1}; check_neg_diag::Bool, verbose::Symbol, distr::Bool) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\hessizero.jl:32 [11] hessian!(::XGabaix{Float64}, ::Array{Float64,1}, ::Array{Float64,2}; check_neg_diag::Bool, verbose::Symbol) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\hessian.jl:41 [12] estimate(::XGabaix{Float64}, ::Array{Float64,2}; verbose::Symbol, proposal_covariance::Array{Any,2}, mle::Bool, sampling::Bool, filestring_addl::Array{String,1}, continue_intermediate::Bool, intermediate_stage_start::Int64, intermediate_stage_increment::Int64, save_intermediate::Bool, tempered_update_prior_weight::Float64, run_csminwel::Bool) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\estimate.jl:187 [13] estimate(::XGabaix{Float64}, ::Array{Float64,2}) at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\src\estimate\estimate.jl:105 [14] macro expansion at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\docs\examples\run_default.jl:46 [inlined] [15] top-level scope at .\timing.jl:174 [inlined] [16] top-level scope at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\docs\examples\run_default.jl:0 [17] include(::String) at .\client.jl:457 [18] top-level scope at none:1 in expression starting at C:\Users\Samsung\.julia\packages\DSGE\tJSkz\docs\examples\run_default.jl:43

from dsge.jl.

Thanks again, William. Even though the (1e-2, 1e-5) parametrization didn't work, the SquareRoot did.
You have been very helpful.

from dsge.jl.