DifferentiableStateSpaceModels
Warning: This package is a proof of concept. While the code will remain working with proper use of a Julia manifest, you should not rely on this for research unless you are prepared to modify the source and help maintain it.
For a more complete example with the code below and Bayesian estimation, see the rbc estimation notebook.
Development and Benchmarking
See development.md for contributing code and running benchmarks
Model Class
The model follows Schmitt-Grohe and Uribe (2004) timing convention. The system takes a nonlinear expectational difference equation including all first-order conditions for decisions and the system evolution equations,
$$ \mathbb{E}_{t}\mathcal{H}\left(y',y,x',x;p\right)=0 $$
where $y$ are the control variables, $x$ are the states, and $p$ is a vector of deep parameters of interest. Expectations are taken over forward-looking variables and an underlying random process $\epsilon'$.
In addition, we consider an observation equation - which might be noisy, for
$$ z = Q \cdot \begin{bmatrix}y &x\end{bmatrix}^{\top} + \nu $$
where $\nu$ may or may not be normally distributed but $\mathbb{E}(\nu) = 0$ and $\mathbb{V}(\nu) = \Omega(p) \Omega(p)^{\top}$.
Assume that there is a non-stochastic steady state of this problem as $y_{ss}, x_{ss}$.
Perturbation Solution
Define the deviation from the non-stochastic steady state as $\hat{x} \equiv x - x_{ss}, \hat{y} \equiv y - y_{ss},$ and $\hat{z} \equiv z - z_{ss}$.
The solution finds the first or second order perturbation around that non-stochastic steady state, and yields
$$ x' = h(x; p) + \eta \ \Gamma(p)\ \epsilon' $$
where $\eta$ describes how shocks affect the law of motion and $\mathbb{E}(\epsilon') = 0$. Frequently this would be organized such that $\mathbb{V}(\epsilon)= I$, but that is not required. In addition, it could instead be interpreted as for $x' = h(x; p) + \eta \ \epsilon'$ with $\mathbb{V}(\epsilon') = \Gamma(p) \Gamma(p)^{\top}$.
and with the policy equation,
$$ y = g(x; p) $$
and finally, substitution in for the observation equation
$$ z= Q \begin{bmatrix} g(x;p) \\ x \end{bmatrix} + \nu $$
First Order Solutions
Perturbation approximates the above model equations, where $h$ and $g$ are not available explicitly, by a Taylor expansion around the steady state. For example, in the case of the 1st order model the solution finds
$$ \hat{x}' = A(p)\ \hat{x} + B(p) \epsilon' $$
and
$$ \hat{y} = g_x(p) \ \hat{x} $$
and
$$ \hat{z} = C(p)\ \hat{x} + \nu $$
where $C(p)\equiv Q \begin{bmatrix} g_x(p) \\ I\end{bmatrix}$, $B(p) \equiv \eta \Gamma(p)$, and $\mathbb{V}(v) = D(\nu) D(p)^{\top}$. Normality of $\nu$ or $\epsilon'$ is not required in general.
This is a linear state-space model and if the priors and shocks are Gaussian, a marginal likelihood can be evaluated with classic methods such as a Kalman Filter. The output of the perturbation can be used manually, or in conjunction with DifferenceEquations.jl.
Second-order solutions are defined similarly. See the estimation notebook for more details.
Gradients
All of the above use standard solution methods. The primary contribution of this package is that all of these model elements are differentiable. Hence, these gradeients can be composed for use in applications such as optimization, gradient-based estimation methods, and with DifferenceEquations.jl which provides differentiable simulations and likelihoods for state-space models. That is, if we think of a perturbation solver mapping $p$ to solutions (e.g. in first order $\mathbf{P}(p) \to (A, B, C, D)$, then we can find the gradients $\partial_p \mathbf{P}(p), \partial_p A(p)$ etc. Or, when these gradients are available for use with reverse-mode auto-differentiation, it can take "wobbles" in $A, B, C, D$ and go back to the "wiggles" of the underlying $p$ through $\mathbf{P}$. See ChainRules.jl for more details on AD.
Examples
Model Primitives
Models are defined using a Dynare-style DSL using Symbolics.jl. The list of primitives are:
- The list of variables for the controls $y$, state $x$, and deep parameters $p$.
- The set of equations $H$ as a function of $p, y(t), y(t+1), x(t),$ and $x(t+1)$. No $t-1$ timing is allowed.
- The loading of shocks $\eta$ as a fixed matrix of constants
- The shock covariance Cholesky factor $\Gamma$ as a function of parameters $p$
- The observation equation $Q$ as a fixed matrix.
- The Cholesky factor of the observation errors, $\Omega$ as a function of parameters $p$. At this point only a diagonal matrix is supported.
- Either the steady state equations for all of $y$ and $x$ in closed form as a function of $p$, or initial conditions for the nonlinear solution to solve for the steady state as functions of $p$
Defining Models
Install this package with ] add DifferentiableStateSpaceModels, then the full code to create the RBC model is
∞ = Inf
@variables α, β, ρ, δ, σ, Ω_1
@variables t::Integer, k(..), z(..), c(..), q(..)
x = [k, z] # states
y = [c, q] # controls
p = [α, β, ρ, δ, σ, Ω_1] # parameters
H = [1 / c(t) - (β / c(t + 1)) * (α * exp(z(t + 1)) * k(t + 1)^(α - 1) + (1 - δ)),
c(t) + k(t + 1) - (1 - δ) * k(t) - q(t),
q(t) - exp(z(t)) * k(t)^α,
z(t + 1) - ρ * z(t)] # system of model equations
# analytic solutions for the steady state. Could pass initial values and run solver and use initial values with steady_states_iv
steady_states = [k(∞) ~ (((1 / β) - 1 + δ) / α)^(1 / (α - 1)),
z(∞) ~ 0,
c(∞) ~ (((1 / β) - 1 + δ) / α)^(α / (α - 1)) -
δ * (((1 / β) - 1 + δ) / α)^(1 / (α - 1)),
q(∞) ~ (((1 / β) - 1 + δ) / α)^(α / (α - 1))]
Γ = [σ;;] # matrix for the 1 shock. The [;;] notation just makes it a matrix rather than vector in julia
η = [0; -1;;] # η is n_x * n_ϵ matrix. The [;; ]notation just makes it a matrix rather than vector in julia
# observation matrix. order is "y" then "x" variables, so [c,q,k,z] in this example
Q = [1.0 0 0 0; # select c as first "z" observable
0 0 1.0 0] # select k as second "z" observable
# diagonal cholesky of covariance matrix for observation noise (so these are standard deviations). Non-diagonal observation noise not currently supported
Ω = [Ω_1, Ω_1]
# Generates the files and includes if required. If the model is already created, then just loads
overwrite_model_cache = true
model_rbc = @make_and_include_perturbation_model("rbc_notebook_example", H, (; t, y, x, p, steady_states, Γ, Ω, η, Q, overwrite_model_cache))
After generation of the model, they can be included as any other julia files in your code (e.g. include(joinpath(pkgdir(DifferentiableStateSpaceModels), ".function_cache","my_model.jl"))) or moved somewhere more convenient.
Inclusion through the @make_and_include_perturbation_model creates the model automatically; after direct inclusion through a julia file, you can create a model with m = PerturbationModel(Main.my_model).
Solving Perturbations
Assuming the above model was created and loaded in one way or another as m
p_f = (ρ = 0.2, δ = 0.02, σ = 0.01, Ω_1 = 0.01) # Fixed parameters
p_d = (α = 0.5, β = 0.95) # Pseudo-true values
sol = generate_perturbation(model_rbc, p_d, p_f) # Solution to the first-order RBC
sol_2 = generate_perturbation(model_rbc, p_d, p_f, Val(2)); # Solution to the second-order RBC
@show sol.retcode, sol_2.retcode, verify_steady_state(m, p_d, p_f) # the final call checks that the analytically provided steady-state solution is correct
The perturbation solution (in the canonical form described in the top section) can be queried from the resulting solution. A few examples for the first order solution are below,
@show sol.y, sol.x # steady states y_ss and x_ss These are the values such that y ≡ ŷ + sol.y and x ≡ x̂ + sol.x
@show sol.g_x # the policy
@show sol.A, sol.B # the evolution equation of the state, so that x̂' = A x̂ + B ϵ
@show sol.C, sol.D; # the evolution equation of the state, so that z = C x̂ + ν with variance of ν as D D'.
@show sol.x_ergodic_var; # covariance matrix of the ergodic distribution of x̂, which is mean zero since x̂ ≡ x - x_ss
Functions of Perturbation Solutions (and their Derivatives)
The core feature of this library is to enable gradients of the perturbation solutions with respect to parameters (i.e., anything in the p_d vector). To show this, we will construct a function which uses the resulting law of motion and finds the gradient of the results with respect to this value.
function IRF(p_d, ϵ_0; m, p_f, steps)
sol = generate_perturbation(m, p_d, p_f) # First-order perturbation by default, pass Val(2) as additional argument to do 2nd order.
x = sol.B * ϵ_0 # start after applying impulse with the model's shock η and Γ(p)
for _ in 1:steps
# A note: you cannot use mutating expressions here with most AD code. i.e. x .= sol.A * x wouldn't work
# For more elaborate simuluations, you would want to use DifferenceEquations.jl in practice
x = sol.A * x # iterate forward using first-order observation equation
end
return [0, 1]' * sol.C * x # choose the second observable using the model's C observation equation since first-order
end
m = model_rbc # ensure notebook executed above
p_f = (ρ=0.2, δ=0.02, σ=0.01, Ω_1=0.01) # not differentiated
p_d = (α=0.5, β=0.95) # different parameters
steps = 10 # steps ahead to forecast
ϵ_0 = [1.0] # shock size
IRF(p_d, ϵ_0; m, p_f, steps) # Function works on its own, calculating perturbation
Derivatives of the Perturbation Solvers
The perturbation solver fills a cache for values used for calculating derivatives.
For example,
using Zygote
function f(params; m, p_f)
p_d = (α=params[1], β=params[2]) # Differentiated parameters
sol = generate_perturbation(m, p_d, p_f) # Default is first-order.
return sum(sol.A) # An ad-hoc example: reducing the law-of-motion matrix into one number
end
# To call it
m = PerturbationModel(Main.my_model)
p_f = (ρ=0.2, δ=0.02, σ=0.01, Ω_1=0.01)
param_val = [0.5, 0.95] # as a vector, but not required
display(f(param_val; m, p_f)) # Function works on its own, calculating perturbation
# Query the solution
@assert f(param_val; m, p_f) ≈ 7.366206154679124
# But you can also get its gradient with Zygote/etc.
display(gradient(params -> f(params; m, p_f), param_val))
# Result check
gradient(params -> f(params; m, p_f), param_val)[1] ≈ [61.41968376547458, 106.44095661062319]
However, the real benefit is that this function can itself be differentiated, to find gradients with respect to the deep parameters p_d and the shock ϵ_0
# Using the Zygote auto-differentiation library already loaded above
p_d = (α=0.5, β=0.95) # different parameters
ϵ_0 = [1.0] # shock size
IRF_grad = gradient((p_d, ϵ_0) -> IRF(p_d, ϵ_0; m, p_f, steps), p_d, ϵ_0) # "closes" over the m, p_f, and steps to create a new function, and differentiates it with respect to other arguments
Simulating Data with DifferenceEquations
Simulating with DifferenceEquations.jl
The manual iteration of the state-space model from the perturbation solution is possible, but can be verbose and difficult to achieve efficiency for gradients. One benefit of this package is that it creates state-space models in a form consistent with DifferenceEquations.jl which can be easily simulated, visualized, and estimated.
To do this, we will calculate a perturbation solution then simulate it for various x_0 drawn from the ergodic solution.
p_f = (ρ = 0.2, δ = 0.02, σ = 0.01, Ω_1 = 0.01) # Fixed parameters
p_d = (α = 0.5, β = 0.95) # Pseudo-true values
m = model_rbc # ensure notebook executed above
sol = generate_perturbation(m, p_d, p_f) # Solution to the first-order RBC
# Simulate T observations from a random initial condition
T = 20
# draw from ergodic distribution for the initial condition
x_iv = MvNormal(sol.x_ergodic_var)
problem = LinearStateSpaceProblem(sol, x_iv, (0, T))
plot(solve(problem))
The LinearStateSpaceProblem type is automatically constructed from the underlying perturbation. However, we can override any of these options, or pass in our own noise rather than simulate it for a particular experiment
noise = Matrix([1.0; zeros(T-1)]') # the ϵ shocks are "noise" in DifferenceEquations for SciML compatibility
x_iv = [0.0, 0.0] # can pass in a single value rather than a distribution
problem = LinearStateSpaceProblem(sol, x_iv, (0, T); noise)
plot(solve(problem))
To demonstrate the composition of gradients between DifferenceEquations and DifferentiableStateSpaceModels lets adapt this function to simulates an impulse with fixed noise and looks at the final observable
function last_observable(p_d, noise, x_iv; m, p_f, T)
sol = generate_perturbation(m, p_d, p_f)
problem = LinearStateSpaceProblem(sol, x_iv, (0, T);noise, observables_noise = nothing) # removing observation noise
return solve(problem).z[end][2] # return 2nd argument of last observable
end
T = 100
noise = Matrix([1.0; zeros(T-1)]') # the ϵ shocks are "noise" in DifferenceEquations for SciML compatibility
x_iv = [0.0, 0.0] # can pass in a single value rather than a distribution
p_f = (ρ = 0.2, δ = 0.02, σ = 0.01, Ω_1 = 0.01) # Fixed parameters
p_d = (α = 0.5, β = 0.95) # Pseudo-true values
m = model_rbc # ensure notebook executed above
last_observable(p_d, noise, x_iv; m, p_f, T)
And, as before, we can calculate gradients with respect to the underlying p_d parameters, but also with respect to the noise which will demonstrate a key benefit of these methods, as they can let us do a joint likelihood of the latent variables in cases where they cannot be easily marginalized out (e.g., non-Gaussian or nonlinear). Note that the dimensionality of this gradient is over 100.
gradient((p_d, noise, x_iv) -> last_observable(p_d, noise, x_iv; m, p_f, T), p_d, noise, x_iv)
Finally, we can use a simple utility functions to investigate an IRF.
p_f = (ρ = 0.2, δ = 0.02, σ = 0.01, Ω_1 = 0.01) # Fixed parameters
p_d = (α = 0.5, β = 0.95) # Pseudo-true values
m = model_rbc
sol = generate_perturbation(model_rbc, p_d, p_f)
ϵ0 = [1.0]
T = 10
sim = irf(sol, ϵ0, T)
plot(sim)
Sneak Peak at SciML Compatible Functionality
Finally, there are a variety of features of SciML which are supported. For example, parallel simulations of ensembles and associated summary statistics.
# Simulate multiple trajectories with T observations
trajectories = 40
x_iv = MvNormal(sol.x_ergodic_var)
problem = LinearStateSpaceProblem(sol, x_iv, (0, T))
# Solve multiple trajectories and plot an ensemble
ensemble_results = solve(EnsembleProblem(problem), DirectIteration(), EnsembleThreads();
trajectories)
summ = EnsembleSummary(ensemble_results) # see SciML documentation. Calculates median and other quantles automatically.
summ.med # median values for the "x" simulated ensembles
plot(summ, fillalpha= 0.2) # plots by default show the median and quantiles of both variables. Modifying transparency as an example
Calculate sequence of observables
We can use the underlying state-space model to easily simulate states and observables
# Simulate T observations
T = 20
p_f = (ρ = 0.2, δ = 0.02, σ = 0.01, Ω_1 = 0.01) # Fixed parameters
p_d = (α = 0.5, β = 0.95) # Pseudo-true values
sol = generate_perturbation(model_rbc, p_d, p_f) # Solution to the first-order RBC
x_iv = MvNormal(sol.x_ergodic_var) # draw initial conditions from the ergodic distribution
problem = LinearStateSpaceProblem(sol, x_iv, (0, T))
sim = solve(problem, DirectIteration())
ϵ = sim.W # store the underlying noise in the simulation
# Collapse to simulated observables as a matrix - as required by current DifferenceEquations.jl likelihood
# see https://github.com/SciML/DifferenceEquations.jl/issues/55 for direct support of this datastructure
z_rbc = hcat(sim.z...)