Introduction to MSTest

Overview

MSTest implements hypothesis testing procedures for determining the number of regimes in Markov switching models. When fitting such a model the number of regimes must be specified by the researcher, yet the standard asymptotic theory for likelihood ratio tests does not apply, because regularity conditions are violated (unidentified nuisance parameters and parameters on the boundary of the parameter space under the null).

The central contribution of the package is a pair of Monte Carlo likelihood ratio tests introduced by Rodriguez-Rondon and Dufour: the local (LMCLRTest) and maximized (MMCLRTest) Monte Carlo LR tests. These apply to both univariate and multivariate models and to general hypotheses of $M_0$ versus $M_0 + m$ regimes; the maximized version additionally controls test size in finite samples and is robust to identification failures. The package also provides three further procedures for the $M_0 = m = 1$ case: the moment-based tests of Dufour and Luger (2017) (DLMCTest, DLMMCTest), the parameter stability test of Carrasco, Hu, and Ploberger (2014) (CHPTest), and Hansen’s (1992) standardized likelihood ratio test (HLRTest).

A typical workflow is to simulate (or load) a process, estimate a model, and test the number of regimes. The examples below use small Monte Carlo replication counts so the vignette builds quickly; in applications these would be larger.

Simulating a Markov switching process

We simulate a two-regime Markov switching autoregressive process in which both the mean and the variance switch across regimes.

set.seed(seed)
mdl <- list(n     = 200,
            mu    = c(0, 8),
            sigma = c(1, 1),
            phi   = c(0.5),
            k     = 2,
            P     = rbind(c(0.90, 0.10),
                          c(0.10, 0.90)))

sim <- simuMSAR(mdl)
plot(sim)

The returned object stores the simulated series in sim$y and the true latent regime path in sim$St.

Estimating a model

Markov switching models are estimated with the expectation–maximization (EM) algorithm by default. The summary method reports the parameter estimates with asymptotic standard errors, the log-likelihood, and information criteria.

set.seed(seed)
mdl_est <- MSARmdl(sim$y, p = 1, k = 2,
                   control = list(msmu = TRUE, msvar = TRUE, use_diff_init = 5))
summary(mdl_est)
#> Markov Switching Autoregressive Model
#>            coef     s.e.
#> mu_1  -0.159860 0.179960
#> mu_2   8.044200 0.177450
#> phi_1  0.476950 0.063037
#> sig_1  0.949480 0.146820
#> sig_2  1.205800 0.159260
#> p_11   0.880950 0.034346
#> p_12   0.119050 0.034346
#> p_21   0.087719 0.026710
#> p_22   0.912280 0.026710
#> 
#> log-likelihood =  -356.1409
#> AIC =  730.2818
#> BIC =  759.9216
#> 
#> Residuals:
#>        Min      1Q   Median      3Q    Max
#> Y1 -2.3155 -0.7503 0.063879 0.70229 3.2086

Testing the number of regimes

Local Monte Carlo likelihood ratio test

LMCLRTest is the local Monte Carlo LR test. For each of N Monte Carlo replications it simulates the model under the null and re-estimates under both hypotheses, and reports the p-value as the rank of the observed likelihood ratio statistic in the simulated null distribution. It is a finite-sample analogue of the parametric bootstrap. Here we test the null of a single regime ($M_0 = 1$) against a two-regime alternative on the data simulated above.

set.seed(seed)
lmc <- LMCLRTest(sim$y, p = 1, k0 = 1, k1 = 2,
                 control = list(N = 19,
                                mdl_h0_control = list(const = TRUE, getSE = FALSE),
                                mdl_h1_control = list(msmu = TRUE, msvar = TRUE,
                                                      getSE = FALSE, use_diff_init = 3)))
summary(lmc)
#> Restricted Model
#>         coef
#> mu    4.5726
#> phi_1 0.7473
#> sig   7.5851
#> 
#> log-likelihood =  -483.9741
#> AIC =  973.9482
#> BIC =  983.8281
#> 
#> Unrestricted Model
#>            coef
#> mu_1  -0.159860
#> mu_2   8.044200
#> phi_1  0.476950
#> sig_1  0.949480
#> sig_2  1.205800
#> p_11   0.880950
#> p_12   0.119050
#> p_21   0.087719
#> p_22   0.912280
#> 
#> log-likelihood =  -356.1409
#> AIC =  730.2818
#> BIC =  759.9216
#> 
#> Rodriguez-Rondon & Dufour (2026) Local Monte Carlo Likelihood Ratio Test
#>          LRT_0  0.90%  0.95%  0.99% p-value
#> LMC_LRT 255.67 13.971 14.213 26.179    0.05

A small p-value is evidence against a single regime in favour of two regimes.

Maximized Monte Carlo likelihood ratio test

MMCLRTest is the maximized version. Rather than fixing the nuisance parameters at a point estimate, it maximizes the Monte Carlo p-value over a consistent set of nuisance parameter values. This is what gives the test exact control of size in finite samples and robustness to identification problems, at the cost of greater computation. The threshold_stop control lets the search terminate as soon as it finds a nuisance value for which the test fails to reject, which is useful when the null is true.

To illustrate the size-control property we apply the test to a series with no regime switching (a single-regime Gaussian process); the test should fail to reject. We use p = 0 (no autoregressive dynamics) and threshold_stop = 0.05 so the search stops as soon as it finds a nuisance value for which the p-value exceeds the 5% level. We set N = 19, the smallest number of Monte Carlo replications that permits a test at the 5% level, since $(N + 1)\alpha$ must be an integer (here $20 \times 0.05 = 1$); see Dufour (2006).

Because the maximization over the nuisance parameter space is computationally intensive, this example is shown with its pre-computed output rather than evaluated when the vignette is built.

set.seed(seed)
y0 <- simuNorm(list(n = 150, q = 1, mu = 0, sigma = as.matrix(1)))$y
mmc <- MMCLRTest(y0, p = 0, k0 = 1, k1 = 2,
                 control = list(N = 19, eps = 0.1, CI_union = FALSE,
                                type = "GenSA", threshold_stop = 0.05,
                                maxit = 10, silence = TRUE,
                                mdl_h0_control = list(getSE = FALSE),
                                mdl_h1_control = list(msmu = TRUE, msvar = TRUE,
                                                      getSE = FALSE, use_diff_init = 1)))
summary(mmc)

#> Rodriguez-Rondon & Dufour (2026) Maximized Monte Carlo Likelihood Ratio Test
#>         LRT_0 p-value
#> MMC_LRT 1.671     0.9

As expected, the test does not reject the null of a single regime (p-value $= 0.9$). In applications N would be larger, and the search can be parallelized with the workers control. The Monte Carlo LR tests also apply to multivariate models (MSVARmdl) and to hypotheses with $M_0 > 1$; see ?LMCLRTest and ?MMCLRTest.

Faster alternatives for the one-versus-two regime case

For the special case of testing one regime against two ($M_0 = m = 1$), the package provides computationally lighter procedures. The moment-based local Monte Carlo test of Dufour and Luger (2017) only requires estimation under the null and is nearly instantaneous:

set.seed(seed)
mom <- DLMCTest(sim$y, p = 1, control = list(N = 99, simdist_N = 10000))
summary(mom)
#> Restricted Model
#>         coef     s.e.
#> mu    4.5726 0.772590
#> phi_1 0.7473 0.047119
#> sig   7.5851 0.760410
#> 
#> log-likelihood =  -483.9741
#> AIC =  973.9482
#> BIC =  983.8281
#> 
#> Dufour & Luger (2017) Moment-Based Local Monte Carlo Test
#>           phi_1   M(ε)   V(ε)    S(ε)   K(ε) F(ε)   0.90%   0.95%   0.99%
#> LMC_min  0.7473 1.2322 25.802 0.11162 3.5487    1 0.96760 0.98268 0.99189
#> LMC_prod 0.7473 1.2322 25.802 0.11162 3.5487    1 0.99861 0.99918 0.99996
#>          p-value
#> LMC_min     0.01
#> LMC_prod    0.01

The parameter stability test of Carrasco, Hu, and Ploberger (2014) is another option that requires estimation only under the null:

set.seed(seed)
chp <- CHPTest(sim$y, p = 1, control = list(N = 99, rho_b = 0.7))
summary(chp)
#> Restricted Model
#>         coef     s.e.
#> mu    4.5726 0.772590
#> phi_1 0.7473 0.047119
#> sig   7.5851 0.760410
#> 
#> log-likelihood =  -483.9741
#> AIC =  973.9482
#> BIC =  983.8281
#> 
#> Carrasco, Hu, & Ploberger (2014) Parameter Stability Test 
#> 
#> - Switch in Mean only
#>       test-stat  0.90%  0.95%  0.99% p-value
#> supTS   0.31733 1.0150 1.4064 2.1762 0.22222
#> expTS   0.97589 1.0661 1.3407 1.5089 0.26263

References

Carrasco, M., Hu, L., and Ploberger, W. (2014). Optimal test for Markov switching parameters. Econometrica, 82(2), 765-784.

Dufour, J.-M., and Luger, R. (2017). Identification-robust moment-based tests for Markov switching in autoregressive models. Econometric Reviews, 36(6-9), 713-727.

Hansen, B. E. (1992). The likelihood ratio test under nonstandard conditions: Testing the Markov switching model of GNP. Journal of Applied Econometrics, 7(S1), S61-S82.

Rodriguez-Rondon, G., and Dufour, J.-M. (2026a). Monte Carlo likelihood-ratio tests for Markov switching models. Bank of Canada Staff Working Paper, No. 2026-23. doi: 10.34989/swp-2026-23.