Creating Models

The ModelConstructors.jl provides several two models as well as example scripts that show how to estimate a model.

  • A generic model GenericModel designed as a template for user-defined models and for users who only needs functionality to estimate parameters.
  • An example of a LinearRegression model with minimal fields.

To see examples of more complex models, we recommend looking at the source code of models in DSGE.jl. These models use most, if not all, of the fields of GenericModel.

GenericModel

Template for New Models

The idea of GenericModel is that it should have most fields required to define a generic mathematical model. For most purposes, it will have more fields than needed. If a user wants to create their own model but are not familiar with the process of defining a new model type, then we recommend that the user copy the source code of GenericModel into a new file and editing that file. Referencing the source code of models in DSGE.jl will be useful.

Estimating CAPM

Some users only need the ability to estimate parameters. GenericModel provides a convenient way to do that since it includes functionality to choose settings and observables. The script estimatecapm.jl in the examples/factormodel folder in docs shows how to use GenericModel in this way. We can write

capm = GenericModel()
capm <= parameter(:α, 0., (-1e5, 1e5), (-1e5, 1e5), Untransformed(), Normal(0, 1e3),
                  fixed = false)
# add in more parameters...

This code instantiates a GenericModel object and adds a parameter called α, which takes a default value of 0. and is assigned a normal prior with mean zero and variance 1000. The two (-1e5, 1e5) intervals specify the bounds of the parameter before and after transformation. The Untransformed() call indicates that we do not transform the parameter, hence why the bounds are the same. For more details on parameters, see the The Parameter type.

The example script then constructs a likelihood function. To use SMC.jl, the likelihood function needs to take two arguments: parameters and data. Using GenericModel makes this simple because we just use capm.parameters as the parameters provided as inputs to the likelihood function.

Finally, we just call smc(likelihood_fnct, capm.parameters, data) to estimate the parameters of the CAPM model we have defined!

LinearRegression

The LinearRegression type is an example of a model type that does not need to use all the fields provided in GenericModel when estimating. For comparison, the type definition for GenericModel is

mutable struct GenericModel{T} <: AbstractModel{T}
    parameters::ParameterVector{T}
    steady_state::ParameterVector{T}
    keys::Dict{Symbol, Int}
    endogenous_states::Dict{Symbol,Int}
    exogenous_shocks::Dict{Symbol,Int}
    observables::Dict{Symbol,Int}
    pseudo_observables::Dict{Symbol,Int}
    spec::String
    subspec::String
    settings::Dict{Symbol, Setting}
    test_settings::Dict{Symbol, Setting}
    rng::MersenneTwister
    testing::Bool
    observable_mappings::Dict{Symbol, Observable}
    pseudo_observable_mappings::Dict{Symbol, PseudoObservable}

while the type definition for LinearRegression is only

mutable struct LinearRegression{T} <: AbstractModel{T}
    parameters::ParameterVector{T}
    keys::Dict{Symbol, Int}
    spec::String
    subspec::String
    settings::Dict{Symbol, Setting}
    test_settings::Dict{Symbol, Setting}
    rng::MersenneTwister
    testing::Bool
end