MarSwitching.jl is a package for estimating Markov switching dynamic models (also called regime switching) in Julia.

Author: Mateusz Dadej, m.dadej {at} unibs.it

Please check the documentation for examples and information on using the package.

citation: I encourage to cite the JOSS paper of the package when using it in your research. You can use the following BibTeX entry from the CITATION.bib file:

@article{Dadej2024, 
  doi = {10.21105/joss.06441}, 
  url = {https://doi.org/10.21105/joss.06441}, 
  year = {2024}, 
  publisher = {The Open Journal}, 
  volume = {9}, 
  number = {98}, 
  pages = {6441}, 
  author = {Mateusz Dadej}, 
  title = {MarSwitching.jl: A Julia package for Markov switching dynamic models}, 
  journal = {Journal of Open Source Software} 
}

Or paste the citation manually:

Dadej, M., (2024). MarSwitching.jl: A Julia package for Markov switching dynamic models. Journal of Open Source Software, 9(98), 6441, https://doi.org/10.21105/joss.06441

MarSwitching is in general registry. To install simply use following command:

] add MarSwitching

The Markov switching models are a class of models that allow for the parameters to change over time, depending on the unobservable state like economic recession, high volatility on financial markets or epidemiologic outbreak. The state follows a Markov process with a given probability transition matrix for each of $k$ states:

Satisfying standard Markovian properties. The general model is defined as follows:

Where $\mathbf{y}_t$ is vector of dependent variables, $\mathbf{\mu}_s$ and $\mathbf{\beta}_s$ are model parameters dependent on the state $S_t$, $\mathbf{\gamma}$ is a vector of parameters for exogenous variables. The error is distributed according to some distribution $f$ with state dependent covariance matrix $\mathcal{\Sigma}_s$.

Because of the unobserved nature of the state, the model is estimated by maximum likelihood. The likelihood function is calculated using the method described in Hamilton, 1989.

The package also provide time-varying transition probabilities (TVTP) (Filardo, 1994) which allows for the transition matrix to change over time, depending on exogenous variables. Each transition probability has a following form:

For more thorough introduction to the Markov switching models, see 9th chapter of Guidolin and Pedio, 2018.

• Currently available:
  • Markov switching model with $k$ regimes and combinations of switching/non-switching:
    • intercept
    • variance
    • exogenous variables
  • Model with time-varying transition probabilities (TVTP) (à la Filardo 1994)
  • Filtered probabilities
  • Smoothed probabilities (Kim, 1994)
  • Summary statistics of coefficients
  • Instantaneous and one step ahead predict()
  • Expected regime duration
  • Simulation of data both from estimated model and from given parameters
  • Variable and number of states selection (with random and grid search)
• Planned functionality:
  • Other error distributions (student-t, GED, etc.)
  • Markov Switching GARCH model
  • Markov Switching VAR model
  • Markov Switching model with lagged states. E.g. $y_t = \mu_{S_t} + \phi(y_{t-1} - \mu_{S_{t-1}})$

Future development is closely related to the package's popularity.

MarSwitching.jl is the fastest open source implementation of the model. The benchmark was done on artificially generated data with 400 observations, from the model with 3 regimes, 1 switching and 1 non switching exogenous variable. Table below shows mean absolute error of estimated parameters with respect to the actual parameters from generate_msm() function.

MarSwitching.jl is 6,7 times faster than statsmodels implementation in Python/Cython, 8,2 times faster than MSwM in R and 42 times faster than MS_Regress in MATLAB/MEX, although MATLAB package is also calculating standard errors during function call. Every implementation had virtually the same error of estimated parameters.

Software versions: MarSwitching.jl v0.2.2, Statsmodels v0.14.1, MSwM v1.5, MS_Regress v1.11. The programming languages versions were: Julia v1.10.1, Python v3.12.2, R v4.2.1 and MATLAB vR2024a. Calculations were run on: Windows 11 x64 Intel(R) Core(TM) i7-9850H CPU @ 2.60GHz, 2592 Mhz, 6 Core(s), 12 Logical Processor(s).

Code of the benchmarks can be found in benchmark folder.

Following example will first generate artificial data and then estimate a simple Markov switching model with regime dependent intercept, exogenous variable and variance. The model is defined as follows:

using MarSwitching
using Random
import Statistics: quantile

k = 2            # number of regimes
T = 400          # number of generated observations
μ = [1.0, -0.5]  # regime-switching intercepts
β = [-1.5, 0.0]  # regime-switching coefficient for β
σ = [1.1,  0.8]  # regime-switching standard deviation
P = [0.9 0.05    # transition matrix (left-stochastic)
     0.1 0.95]

Random.seed!(123)

# generate artificial data with given parameters
y, s_t, X = generate_msm(μ, σ, P, T, β = β) 

# estimate the model
model = MSModel(y, k, intercept = "switching", exog_switching_vars = X[:,2])

# we may simulate data also from estimated model
# e.g. for calculating VaR:
quantile(generate_msm(model, 1000)[1], 0.05)

# or more interestingly, output summary table
summary_msm(model)

The summary_msm(model) will output following summary table:

Markov Switching Model with 2 regimes
=================================================================  
# of observations:          400 AIC:                      1169.384 
# of estimated parameters:    8 BIC:                      1201.316 
Error distribution:    Gaussian Instant. adj. R^2:          0.4957 
Loglikelihood:           -576.7 Step-ahead adj. R^2:        0.3757 
-----------------------------------------------------------------  
------------------------------
Summary of regime 1:
------------------------------
Coefficient  |  Estimate  |  Std. Error  |  z value  |  Pr(>|z|)   
-------------------------------------------------------------------
β_0          |    -0.516  |       0.052  |   -9.874  |    < 1e-3  
β_1          |    -0.003  |       0.051  |   -0.058  |     0.954  
σ            |     0.843  |       0.022  |   38.751  |    < 1e-3  
-------------------------------------------------------------------        
Expected regime duration: 28.58 periods
-------------------------------------------------------------------        
------------------------------
Summary of regime 2:
------------------------------
Coefficient  |  Estimate  |  Std. Error  |  z value  |  Pr(>|z|)
-------------------------------------------------------------------        
β_0          |     0.824  |       0.132  |    6.256  |    < 1e-3  
β_1          |    -1.483  |        0.12  |  -12.308  |    < 1e-3  
σ            |     1.124  |       0.046  |   24.639  |    < 1e-3
-------------------------------------------------------------------        
Expected regime duration: 11.34 periods
-------------------------------------------------------------------        
left-stochastic transition matrix:
          | regime 1   | regime 2
---------------------------------------
 regime 1 |   96.501%  |    8.819%  |
 regime 2 |    3.499%  |   91.181%  |

As can be seen, the parameters correspond to the ones defined when producing the data generating process.

The package also provides a function for filtered transition probabilities $P(S_t = i | \Psi_t)$, as well as smoothed ones (Kim, 1994) $P(S_t = i | \Psi_T)$. Essentially, the difference is that, in order to calculate the smoothed probabilities the whole sample is used.

using Plots

plot(filtered_probs(model),
     label     = ["Regime 1" "Regime 2"],
     title     = "Regime probabilities", 
     linewidth = 2)

plot(smoothed_probs(model),
     label     = ["Regime 1" "Regime 2"],
     title     = "Smoothed regime probabilities",
     linewidth = 2)  

Every exported function have a docstring, which can be accessed by ? in REPL.

The function for estimating the Markov switching model is:

MSModel(y::VecOrMat{V},                    # vector of dependent variable
        k::Int64,                          # number of regimes
        ;intercept::String = "switching",  # "switching" (default), "non-switching" or "no" intercept
        exog_vars::VecOrMat{V},            # optional matrix of exogenous variables
        exog_switching_vars::VecOrMat{V},  # optional matrix of exogenous variables with regime switching
        switching_var::Bool = true,        # is variance state-dependent?
        exog_tvtp::VecOrMat{V},            # optional matrix of exogenous variables for time-varying transition matrix
        x0::Vector{V},                     # optional initial values of parameters for optimization
        algorithm::Symbol = :LN_SBPLX,     # optional algorithm for NLopt.jl
        maxtime::Int64 = -1,               # optional maximum time for optimization
        random_search::Int64 = 0           # Number of random search iterations (model estimations with random disturbance to the x0)
        ) where V <: AbstractFloat  

The function returns MSM type object:

struct MSM{V <: AbstractFloat}
    β::Vector{Vector{V}}  # β[state][i] vector of β for each state
    σ::Vector{V}          # error variance
    P::Matrix{V}          # transition matrix
    δ::Vector{V}          # tvtp parameters
    k::Int64              # number of regimes 
    n_β::Int64            # number of β parameters
    n_β_ns::Int64         # number of non-switching β parameters
    intercept::String     # "switching", "non-switching" or "no"
    switching_var::Bool   # is variance state dependent?
    x::Matrix{V}          # data matrix
    T::Int64              # number of observations
    Likelihood::Float64   # vector of parameters for optimization
    raw_params::Vector{V} # raw parameters used directly in the Likelihood function
    nlopt_msg::Symbol     # message from NLopt.jl solver
end

Filtered transition probabilities can be calculated from estimated model:

filtered_probs(model::MSM,                           # estimated model
               y::Vector{Float64},                   # optional vector of dependent variables
               exog_vars::Matrix{Float64}            # optional matrix of exogenous variables
               exog_switching_vars::Matrix{Float64}, # optional matrix of exogenous variables with regime switching
               exog_tvtp::Matrix{Float64})           # optional matrix of exogenous variables for time-varying transition matrix
                

Similarly, smoothed transition probabilities can be also calculated from estimated model:

smoothed_probs(model::MSM,                           # estimated model
               y::Vector{Float64},                   # optional vector of dependent variables
               exog_vars::Matrix{Float64}            # optional matrix of exogenous variables
               exog_switching_vars::Matrix{Float64}, # optional matrix of exogenous variables with regime switching
               exog_tvtp::Matrix{Float64})           # optional matrix of exogenous variables for time-varying transition matrix

The predict() function can be used to calculate instantaneous or one step ahead predictions from estimated model:

predict(model::MSM,                            # estimated model
        instantaneous::Bool = false;             # instanteous or one-step ahead prediction
        y::Vector{Float64},                    # optional vector of dependent variables
        exog_vars::Matrix{Float64},            # optional matrix of exogenous variables
        exog_switching_vars::Matrix{Float64},  # optional matrix of exogenous variables with regime switching
        exog_tvtp::Matrix{Float64})            # optional matrix of exogenous variables for time-varying transition matrix
    

Which is the probability weighted average of predictions of each state equation:

And for one step ahead, the state probabilities have to be predicted themselves:

The one step ahead prediction will return a vector of size $(T-1) \times 1$, as the observation $t-1$ is used to forecast state probability ($P\hat{\xi}_{i,t}$)

The provided new data needs to match the data used for estimation (with except of observations size). If not provided the prediction is done on the data used for estimation. For one step ahead forecast, there is no look-ahead bias, the y vector needs to be provided in order to calculate the state probabilities at time $t$.

The summary_msm(model::MSM; digits::Int64=3) function outputs a summary table that is built from 4 functions:

• transition_mat(model::MSM; digits::Int64=2) - prints transition matrix
• coeftable_tvtp(model::MSM; digits::Int64=3) - prints coefficient table for time-varying transition matrix
• state_coeftable(model::MSM, state::Int64; digits::Int64=3) - prints coefficient table for given state
• expected_duration(model::MSM; digits::Int64=2) - prints expected duration of each state, or a time series of expected duration for TVTP model

It is also possible to simulate data from a given set of parameters:

generate_msm(μ::Vector{Float64},    # vector of intercepts for each state
             σ::Vector{Float64},    # vector of error variances for each state
             P::Matrix{Float64},    # transition matrix
             T::Int64;              # number of observations
             β::Vector{Float64},    # vector of coefficients for each state
             β_ns::Vector{Float64}, # vector of non-switching coefficients
             δ::Vector{Float64},    # vector of coefficients for time-varying transition matrix
             tvtp_intercept::Bool)  # should TVTP have an intercept?

or thanks to multiple dispatch, simulate data from estimated model (as in example):

generate_msm(model::MSM, T::Int64=model.T) 

The function returns a tuple of 3 elements, respectively:

• y: vector of dependent variables
• s_t: vector of states
• X: matrix of exogenous variables (generated from standard normal distribution), where last variables are for TVTP

Function add_lags(y::Vector{Float64}, p::Int64) adds p lags to the matrix of dependent variables. The function returns a matrix of dependent variables with p lags.

• PRs with fixed bugs or new methods are highly appreciated. Especially the ones described in functionality section.
• Open an issue if the PR changes current code substanitally.
• If unsure, check the ColPrac guide on collaborative practices for Packages.

• Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica: Journal of the Econometric Society, 357-384.

• Kim, Chang Jin (1994). Dynamic Linear Models with Markov-Switching. Journal of Econometrics 60, 1-22.

• Filardo, Andrew J. (1994). Business cycle phases and their transitional dynamics. Journal of Business & Economic Statistics, 12(3), 299-308.

• Guidolin, Massimo & Pedio, Manuela (2018). Essentials of Time Series for Financial Applications. Academic Press.