MS Regress - The MATLAB Package for

Markov Regime Switching Models

Marcelo Perlin
[email protected]
First Version: November 23, 2010
This version: April 19, 2015

Abstract
This paper provides an overview of MS Regress, a MATLAB toolbox specially designed for the estimation, simulation and forecasting
of a general markov regime switching model. The package was written
in an intuitive manner so that the user have at its reach a large number of different markov switching specifications, without any change
in the original code. This document introduces the main functionality
of the package with the help of several empirical examples.

Assistant Professor of Finance, Federal University of Rio Grande do Sul (Porto Alegre,
Brazil)

Contents
1 Introduction

2 Overview of the package

3 Installation of the package

4 Introduction to markov regime switching models

4.1 Maximum likelihood estimation . . . . . . . . . . . . .

4
6

5 Interface of MS Regress for estimation

5.1 Interface to univariate modeling . . . . . . . . .
5.2 Interface to multivariate modeling . . . . . . .
5.3 Estimating autoregressive Models (MS - VAR)
5.4 The output from the estimation function . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

9
10
15
17
18

6 Interface of MS Regress for simulation

19

7 Using advanced options

7.1 Using the mex version of Hamiltons filter . . . . . . .
7.2 Using advOpt.constCoeff (constraining coefficients) . .

22
25
25

8 Comparing results against Hamilton [1989]

30

9 Advises for using the package

32

10 Possible error messages

33

11 Frequently asked questions

35

12 Reporting a bug

37

13 Citing the package

37

14 Final remarks

37

Introduction

The purpose of this document is to introduce the user to the functionality of MS Regress package1 . The document is organized as follows,
first I give a brief exposition on the topic of regime switching models
and the further sections of the paper are related to the introduction
to the features of the package along with illustrative examples.

Overview of the package

The MS Regress package was written for the estimation, simulation

and forecasting of a general markov switching model. The main functionality of the code is build around three functions:
MS Regress Fit - Function for estimating a MS model
MS Regress For - Function for forecasting a regime switching model
MS Regress Sim - Function for simulating a MS model
Each of these functions have a similar interface. Within the package there are several example scripts that show the functionality of
each. In this paper, special attention is given to the fitting function
(MS Regress Fit) since this is the one most likely to the used.

Installation of the package

The installation of the package is quite straightforward. The downloaded zip file contains several m files and a few .txt files with the
data used in the example scripts. In order to use the main functions,
all you need to do is to tell Matlab to place the files from the m Files
folder in the search path:
addpath('m Files');
Once Matlab recognizes the path of the package, the functions will be
1

The package is still under development to accommodate new features. The up to date
version can be downloaded from https://sites.google.com/site/marceloperlin/.
I also wrote a lighter version of the package in R. The code is available within the
Rmetrics project (https://r-forge.r-project.org/projects/rmetrics/, search for
fMarkovSwitching). Please be aware that the R version is no longer being maintained.

available to the user. More details about how to use each function can
be found in the description of the files2 . After the use of the package,
it is advised (but not necessary) to remove the files from the search
path. This is accomplished with:
rmpath('m Files');
These commands are already included in all example scripts. These
scripts were written so that they can run without any modification.
They are a good starting point in learning the interface of the package.
Next I give a brief introduction to markov regime switching models.

4 Introduction to markov regime switching models

Markov regime switching models are a type of specification in which
the selling point is the flexibility in handling processes driven by heterogeneous states of the world. In this section I give a brief exposition
on the subject. Technical details regarding markov regime switching
models can be found in Hamilton [1994], Kim and Nelson [1999], Wang
[2003]. For introductory material on the subject, see Hamilton [2005],
Brooks [2002], Alexander [2008] and Tsay [2002] among others.
Consider the following process given by:
yt = St + t

(1)

where St = 1..k and t follows a Normal distribution with zero

mean and variance given by S2 t . This is the simplest case of a model
with a switching dynamic. Note that for the model given in Equation
1, the intercept is switching states with respect to an indicator variable
St . This means that if there are k states, there will be k values for
and 2 . If there is only one state of the world (St = 1), formula 1
takes the shape of yt = 1 + t and it can be treated as a simple linear
regression model under general conditions.
Assuming now that the model in 1 has two states (k = 2). An
2

These are related to the first comments in each m file, which are available by opening
it in the editor. An alternative way to access the description of each file is to use the help
function (e.g. help('MS Regress Fit')).

alternative representation is:

yt = 1 + t

for State 1

(2)

yt = 2 + t

for State 2

(3)

t (0, 12 )

for State 1

(4)

(0, 22 )

for State 2

(5)

where:

t

This representation clearly implies two different processes for the

dependent variable yt . When the state of the world for time t is 1
(2), then the expectation of the dependent variable is 1 (2 ) and the
volatility of the innovations is 12 (22 ).
For an empirical example, yt can represent a vector of log returns
for a financial asset3 . The value of 1 is the expected return on a bull
market state, which implies a positive trend for financial prices and
consequently a positive log return for yt . The lower value 2 measures
the expected log return for the bear market state, which then implies
a negative trend in prices.
The different volatilities (12 and 22 ) in each state represent the
higher uncertainty regarding the predictive power of the model in each
state of the world. Going back to my example, one could expect that
the bear market state is more volatile than the bull market. This
implies that prices go down faster than they go up4 . This means
2
2
that we can expect Bear
to be higher than Bull
.Note that I do not
identify the states (e.g. bull market is state 1). In general, the St
variable simply index the states, where the interpretation is given by
looking at parameters values.
So far I havent said how exactly the switching from one state to the
other happens. For instance, how is that one should know which state
of the world is for each point in time. Suppose that we had assumed
a deterministic transition of states where state 1 is true for time t
3

The log return is the geometric change for the price of a particular asset (e.g. stock
price) between time t 1 and t. Formally, if Pt is the price of a particular asset for time
t, then log(Pt /Pt1 ) is the log return for time t.
4
The usual explanation for this effect is that traders react faster to bad news when
comparing to good news. This can also be explained by the presence of limit loss orders,
which will sell at market prices once a particular threshold in the prices has been breached.
When used by a significant amount of traders and at different threshold levels, these limit
loss orders will create a cascade effect, therefore accelerating the downfall of prices.

when a exogenous time series zt is positive. This greatly simplifies

the model as each state is observable and, therefore, we can treat the
model given before as a regression with dummy variables. This would
take the shape of yt = Dt (1 + 1,t ) + (1 Dt )(2 + 1,t ), where Dt is
the dummy variable taking value of 1 if zt > 0 and 0 otherwise.
For a markov regime switching model, the transition of states is
stochastic (and not deterministic). This means that one is never sure
whether there will be a switch of state or not. But, the dynamics behind the switching process is know and driven by a transition matrix.
This matrix will control the probabilities of making a switch from one
state to the other. It can be represented as:
p11 . . . p1k
..
..
..
P = .
.
.
pk1 . . . pkk

(6)

For 6, the element5 in row i, column j (pij ) controls the probability

of a switch from state j to state i. For example, consider that for some
time t the state of the world is 2. This means that the probability of a
switch from state 2 to state 1 between time t and t + 1 will be given by
p12 . Likewise, a probability of staying in state 2 is determined by p22 .
This is one of the central points of the structure of a markov regime
switching model, that is, the switching of the states of the world is
a stochastic process itself6 . Usually these transition probabilities are
assumed constant, but it is also possible to allow it to vary over time.
This is called the TVTP (time varying transition probabilities) model,
which is not supported7 by MS Regress. See Wang [2003] for more
details in this type of specification.

4.1

Maximum likelihood estimation

A general MS model can be estimated with two different methods,

maximum likelihood or Bayesian inference (Gibbs-Sampling). In the
5

The ordering of the elements in the matrix is a matter of notation. It is not uncommon
to find different notations in the literature.
6
In fact, the probabilities of each regime over time can be represented as a AR process,
[Hamilton, 1994]
7
Zhuanxin Ding developed a matlab package for TVTP models based on MS Regress.
You can access it at: http://www.mathworks.com/matlabcentral/fileexchange/
37144.

matlab package all of the models are estimated using maximum likelihood and this procedure is described here. Consider the following
regime switching model:
yt = St + t
t

(7)

N (0, S2 t )

(8)

St = 1, 2

(9)

The log likelihood of this model is given by:

ln L =

T
X

ln q
2S2 t
t=1

yt S t
exp
2S2 t

(10)

For the previous specification, if all of the states of the world were
know, that is, the values of St are available, then estimating the model
by maximum likelihood is straightforward. All you need is to maximize
Equation 10 as a function of parameters 1 , 2 , 12 and 22 . Is should
be clear by now that this is not the case for a markov switching model,
where the states of the world are unknown.
In order to estimate a regime switching model where the states
are not know, it is necessary to change the notation for the likelihood function. Considering f (yt |St = j, ) as the likelihood function
for state j conditional on a set of parameters (), then the full log
likelihood function of the model is given by:
ln L =

T
X
t=1

ln

2
X

(f (yt |St = j, )P r(St = j))

(11)

j=1

which is just a weighted average of the likelihood function in each state,

where the weights are given by the states probabilities. When these
probabilities are not observed, we cannot apply equation 11 directly,
but we can make inferences on the probabilities based on the available
information. This is the main idea of Hamiltons filter, which is used
to calculate the filtered probabilities of each state based on the arrival
of new information.
Consider t1 as the matrix of available information at time t 1.
Using Hamiltons filter the estimates of P r(St = j) are available using
the following iterative algorithm:
1. Setting a guess for the starting probabilities (t = 0) of each state
P r(S0 = j) for j = 1, 2. Here one can use a naive guess, e.g.

P r(S0 = j) = 0.5 or the steady-state (unconditional) probabilities of St :

P r(S0 = 1|0 ) =

1 p11
2 p11 p22

(12)

P r(S0 = 2|0 ) =

1 p22
2 p22 p11

(13)

2. Set t = 1 and calculate the probabilities of each state given

information up to time t 1:
P r(St = j|t1 ) =

2
X

pji (P r(St1 = i|t1 ))

(14)

i=1

where pji are the transition probabilities from the markov chain
(see equation 6).
3. Update the probability of each state with the new information
from time t. This is accomplished by using the parameters of the
model in each state, in this case 1 ,2 ,1 ,2 , the transition probabilities p11 and p22 for the calculation of the likelihood function
in each state (f (yt |St = j, t1 )) for time t. After that, use the
following formula to update the probability of each state given
the new information:
f (yt |St = j, t1 )P r(St = j|t1 )
P r(St = j|t ) = P2
(15)
j=1 f (yt |St = j, t1 )P r(St = j|t1 )
4. Set t = t + 1 and repeat steps 2-3 until t = T , that is, you
reached all observations in your sample. This should provide a
set of filtered probabilities for each state, from t = 1 to t = T .
The previous set of steps provides the probabilities that one needs
for calculating the log likelihood of the model as a function of the set
of parameters:
ln L =

T
X
t=1

ln

2
X

(f (yt |St = j, )P r(St = j|t ))

(16)

j=1

The estimation of the model is obtained by finding the set of parameters that maximize the previous equation. While I used k = 2 in

the previous example, the formula for the generic case of k states is
also available, where matrix notation can greatly simplify the calculations. See Hamilton [1994] and Kim and Nelson [1999] for further
details on this topic.
One important point in the estimation of the regime switching
model is that the parameters in the transition matrix are not variation
free. Since they are probabilities, their values have to be between 0
and 1 and need to sum to 1 in each column of the transition matrix
P . For k = 2, these conditions are easy to implement by just applying
a numerical transformation8 for p11 and p22 and setting p12 = 1 p11
and p22 = 1 p21 . But, be aware that when k > 3 this procedure will
not work as the second condition of the model will not be guaranteed.
One solution to this problem is to use a constrained optimization
function, so that the desired properties in the parameter vector are
explicitly defined in the search of the log likelihood solution. This
was the approach used in the previous version of the code, with the
use of Matlabs fmincon function. Recently Zhuanxin Ding provided
an interesting numerical transformation9 that made possible for the
transition matrix to be estimated in a variation free framework. This
allowed for the use of different optimization functions such as fminsearch and fminunc in the estimation of the model.

5 Interface of MS Regress for estimation

When thinking in computational terms, a (univariate10 ) markov switching model can be represented in a generalized notation. Consider the
following formula:
yt =

N
nS
X

i xnS
i,t +

i=1

NS
X

j,St xSj,t + t

(17)

j=1

t P (St )

(18)

This representation can nest a high variety of univariate markov

switching specifications. The terms NS and NnS simply counts the
8

See Hamilton [1994].

See Zhuanxin [2012].
10
The package also has support for multivariate models but, for sake of simplicity, I
restrict the introduction in the topic for univariate specifications, only.
9

number of switching (and non switching) coefficients, respectively.

The variable xnS
i,t is a subset of xi,t and contains all explanatory variables which dont have a switching effect. Likewise, the variable xSj,t
contains all the variables that have the switching effect. The term
P () is the assumed probability density function of the innovations,
with its own set of parameters (vector ).
When translating the structure in 17 and 18 into a computational
framework, the required information one would need in order to set
up its own specification is the explained (yt ) and the explanatory data
(xi,t ), the location (and number) of parameters that will switch states
and the shape of the probability density function for the innovations.
The MS Regress package has a very intuitive way of addressing such
structure, which makes it a handy package for this type of specifications.
The central point of this flexibility resides in the input argument
S, which controls for where to include markov switching effects. The
package can estimate univariate and multivariate markov switching
models but, the interface is slightly different between the cases. Lets
first begin with the univariate interface.

5.1

Interface to univariate modeling

When in MATLAB environment, the way to call the fitting function

of the package is11 :
1

Spec Out=MS Regress Fit(dep,indep,k,S,advOpt)

The first variable Spec Out is the output structure which contains
all the information regarding the estimated model. Variable dep is
the dependent variable and it can either be a vector (univariate) or a
matrix (multivariate model). The input indep represents the independent variables. For the case of a univariate model, it is represented
as a matrix with columns equal to the number of regressors. For the
case of a multivariate model, this variable is a cell array12 . The last
three inputs, k, S and advOpt are, respectively, the number of states
in the model, the locations of the switching parameters and advanced
11

Special thanks to Florian Knorm for providing a latex package for Matlab code (available at http://www.mathworks.com/matlabcentral/fileexchange/8015/).
12
Details on multivariate models are given later on the paper.

10

options fed to the algorithm. Further details on these are given next,
with the help of an example.
Consider the following case of a model with two explanatory variables (x1,t , x2,t ) where the innovations follow a Gaussian (Normal)
distribution and the input argument S, which is passed to the fitting function MS Regress Fit.m, is equal to S=[1 1 1]. Given this
configuration, the model is represented as:
yt = 1,St x1,t + 2,St x2,t + t

(19)

t N (0, S2 t )

(20)

Where:
St : The state at time t, that is,St = 1..k , where k is the number
of states.
S2 t : The variance of the innovation at state St .
i,St : Beta coefficient for explanatory variable i at state St where i
goes from 1 to 2.
t : Residual vector which follows a particular distribution (in this case
Normal).
Now, changing the input argument to S=[1 1 0], the model is:
yt = 1,St x1,t + 2,St x2,t + t

(21)

t N (0, 2 )

(22)

Notes that that with S=[1 1 0], the variance term is no longer
switching states. Now, with the switching argument as S=[0 1 1],
the model is:
yt = 1 x1,t + 2,St x2,t + t

(23)

t N (0, S2 t )

(24)

With this change in the input argument S, only the second coefficient and the models variance are switching according to the transition probabilities. That is, the first coefficient (1 ) does not change
states. Therefore, the logic is clear: the first elements of S control the
switching dynamic of the mean equation, while the last term controls
the switching dynamic of the residual vector, including distribution

11

parameters. For an example with extra distribution parameters, consider the following definition for a model with GED (generalized error
distribution) innovations and with input S=[1 1 1 1]. This configuration yields:
yt = 1,St x1,t + 2,St x2,t + t

(25)

t GED(0, S2 t , KSt )

(26)

In this setup, the new parameter K will also switch states. This
coefficient is part of the GED distribution and should not be confused with the k (number of states in the model). If we had set
S=[1 1 1 0], the model would switch in all coefficients, except in
the K parameter.
As an example for the markov switching fitting function, assume
that there is a variable called logRet in MATLABs workspace. This
represents the log returns of a particular asset over a particular frequency (e.g. daily). Consider the input of the following options at
MS Regress Fit():

1
2
3
4
5
6
7
8

% Defining inputs
dep=logRet;
constVec=ones(length(dep),1);
indep=constVec;
k=2;
S=[1 1];
advOpt.distrib='Normal';
advOpt.std method=1;

9
10
11

% Calling fitting function

Spec Out=MS Regress Fit(dep,indep,k,S,advOpt)

For the last piece of code, the vector dep is the dependent variable. The term constVec is a vector full of ones (the constant), k is
the number of states and S defines the location of the switching parameters. The model represented in computational terms in the last
piece of Matlab code is equivalent to the model with only a constant
term and two states given previously in the paper (see Equation 1).
The input structure advOpt determines advanced information of
the model, in this case, the distribution assumption and the method
for calculating the standard errors. More details regarding the use of
advOpt can be found in a later section of the paper.

12

The inputs given before are related to the estimation of the following equations:
State 1 (St = 1)

(27)

yt = 1 + t

(28)

t N (0, 12 )

(29)

State 2 (St = 2)

(30)

yt = 2 + t

(31)

t

N (0, 22 )

with:
P=

p1,1 p2,1
p1,2 p2,2

(32)
!

(33)

as the transition matrix, which controls the probability of a switch

from state j (column j) to state i (row i). The sum of each column in
P is equal to one, since they represent full probabilities of the process
for each state.
Another flexibility of the package is that, since I wrote it for dealing
with a generic type of regression, you can set any kind of explanatory
variable in the model as long as they are observed (you have it available for the whole time period studied). This includes autoregressive
components, constants or just plain regression on other variables.
If you run the script Example MS Regress Fit.m with MATLAB
version 7.10.0.499 (R2010a), this is the output you should be getting
if you have all the proper packages installed (optimization, statistics).

13

***** Numerical Optimization Converged *****

2
3
4
5
6
7
8

Final log Likelihood: 2487.232

Number of estimated parameters: 8
Optimizer: fminsearch
Type of Switching Model: Univariate
Distribution Assumption -> Normal
Method SE calculation -> 1

9
10

***** Final Parameters for Equation #1 *****

11
12

---> Non Switching Parameters <---

13
14
15
16
17
18
19

Non Switching Parameter for Eq #1, Indep column 2

Value:
0.4788
Std Error (p. value): 0.0269 (0.00)
Non Switching Parameter for Eq #1, Indep column 3
Value:
0.1555
Std Error (p. value): 0.0333 (0.00)

20
21

---> Switching Parameters (Dist Parameters)

<---

22
23
24
25
26
27
28

State 1
Model's Variance:
Std Error (p. value):
State 2
Model's Variance:
Std Error (p. value):

0.000266
0.0000 (0.00)
0.000837
0.0001 (0.00)

29
30

--->

Switching Parameters (Regressors)

<---

31
32

Switching Parameters for Eq #1 - Indep column 1

33
34
35
36
37
38
39

State 1
Value:
Std Error (p. value):
State 2
Value:
Std Error (p. value):

0.0003
0.0006 (0.63)
0.0005
0.0016 (0.75)

40
41

---> Transition Matrix (std. error, p-value) <---

42
43
44

0.99 (0.20,0.00)
0.01 ( NaN, NaN)

0.02 (0.21,0.91)
0.98 ( NaN, NaN)

45

14

46

---> Expected Duration of Regimes <---

47

Expected duration of Regime #1: 109.01

Expected duration of Regime #2: 43.73

48
49

The figures generated by the code are as follows:

Figure 1: Output from Example MS Regress Fit.m

5.2

Interface to multivariate modeling

By now it should be clear the use of input argument S. The multivariate interface I built for this package has the same intuition as
in the univariate case and consists of basically a group of inputs in
cell array notation, where the elements iterates over the equations in
the system. An example should make it clear. Consider the following
model:
y1,t = 1,1,St x1,t + 1,2,St x2,t + 1,t

15

(34)

y2,t = 2,1,St x3,t + 2,t

(35)

2
1,t N (0, 1,S
)
t

(36)

2
2,t N (0, 2,S
)
t

(37)

St = 1, 2

(38)

cov(1,t , 2,t ) = 0

(39)

with

This last system of 2 equations is translated in the packages notation as:

1
2
3
4
5
6
7

% Defining input variables

dep=y;
indep{1}=[x1 x2];
indep{2}=x3;
k=2;
S{1}=[1 1 1];
S{2}=[1 1];

8
9
10

% Calling fitting function

Spec Out=MS Regress Fit(dep,indep,k,S);

For the last piece of code, variable y is a matrix with two columns.
As one can see, the inputs have the same shape as for the univariate
case, but they are iterated over the equations of the system by using
cell arrays13 . Notice that in the previous code it is not allowed covariance between the residuals of the different equations (see 39). This is
a default option of the algorithm, but it can be changed so that a full
covariance matrix is estimated from the data.
All options to univariate modeling, including reduced estimation,
are also available for the multivariate case. For further details see
the script Example MS Regress Fit MultiVar.m, which can be
found in the packages zip file.
13

For those not familiarized with Matlab, cell arrays are a type of flexible structure that
can accommodate any kind of data. They are similar to the use of lists in R.

16

5.3 Estimating autoregressive Models (MS VAR)

The package also comes with a simple wrapper function for estimating
a general autoregressive markov switching model. In the current version of the package, moving average terms and error correction models
are not supported.
For example, consider a matrix consisting of two time series (Yt =
[y1,t y2,t ]) that follows an autoregressive system with BSt as the matrix
of parameters:
Yt = BSt Yt1 + t

(40)

with:
t N (0, St )
St =

2
1,S
t
1,2,St

1,2,St
2
2,S
t

(41)
!

(42)

This model translates into the packages notation as:

1
2
3
4
5
6

% Defining input
dep=logRet(:,1:2);
nLag=1;
k=2;
advOpt.diagCovMat=0;
doIntercept=0;

7
8
9

% Calling fitting function

Spec Out=MS VAR Fit(dep,nLag,k,doIntercept,advOpt);

As one can see, the input structure is similar to MS Regress Fit.

The new inputs, nLag, doIntercept and advOpt.diagCovMat
are respectively the number of lags in the system, the choice of using
intercept in the equations and the use of a full matrix as the covariance
matrix14 .
14

This implies that all elements in the covariance matrix are estimated from the data.
If advOpt.diagCovMat=1, then only the elements in the diagonal (the variances) are
estimated and the rest (the non diagonal elements, the covariances) are all set to zero.

17

5.4

The output from the estimation function

The estimation function MS Regress Fit returns a structure with all

the information regarding your model. This comprises of the following
fields:
Coeff: A Structure with the fields:
Coeff.p: The transition probability matrix
Coeff.nS Param: All non switching betas (iterated over equations (cells) and independent variables (rows))
Coeff.S Param: All switching betas (iterated over equations
(cells), over independent variables (rows) and states (columns))
Coeff.covMat: Covariance matrix (iterated over states (cell)).
filtProb: The filtered probabilities of regimes (iterated over the
states (columns))
LL: Final log likelihood of model
k: Number of states
param: All estimated parameters in vector notation
S: Switching flag control (iterating over equations (cell))
advOpt: advanced options fed to algorithm (see next section for
details)
condMean: Conditional mean calculated15 by the model
condStd: Conditional standard deviation calculated16 by the
model.
resid: Residuals from the model (iterating over equations (columns))
stateDur: Expected duration of each state
smoothProb: Smoothed probabilities of regimes (iterated over
the states (columns))
nObs: Number of observations (rows) in the model
Number Parameters: Number of estimated parameters
Coeff SE: A structure with all standard errors of coefficients
(same fields as Coeff)
15
16

These are calculated based only on filtered probabilities prior to time t.

These are also calculated with information prior to time t.

18

 Coeff pValues: A structure with all parameters p-values (same

fields as Coeff)
AIC: Akaike information criteria of the estimated model
BIC: Bayesian information criteria for estimated model
These fields should contain all the information you need for further
testing of the model. If youre missing something, please let me know.

6 Interface of MS Regress for simulation

While the most likely use of MS Regress is for the estimation of a
regime switching model based on a time series data, I also provide
functions for simulation of a model, which can be used in a teaching
environment.
The structure of the interface of the simulation function MS Regress Sim
is very similar to the fitting function MS Regress Sim, but there are
some differences. Here I make some remarks in the use of MS Regress Sim:
In the use of MS Regress Sim, the vector S is related only to
the switching dynamic of the conditional mean. This is different
than the notation in MS Regress Fit, where the last elements
of S defined the switching or not of the innovations parameters.
In the simulation function it is assumed that the conditional
variance is always switching states.
The independent variable used in the simulation are created
within the code as normal random variables with user defined
moments (see input Coeff.indepMean and Coeff.indepStd).
The input variable Coeff.nS param must always be defined,
even when it is not switching states.
An example should make it clear. Consider the following model:

yt = 0.5 + t

State 1

yt = 0.5 + t
2

t N (0, 0.5 )
t N (0, 1)

19

State 2
State 1
State 2

(43)
(44)
(45)
(46)

"

0.95 0.1
0.05 0.9

(47)

In order to simulate the previous set of equations, one need to

build the innovations in each state and also the states of the world
over time, which are defined according to the transition matrix. The
simulation of the previous model is accomplished with the following
code17 in MS Regress Sim:
1

% Example Script for MS Regress Simul.m

2
3
4

% add 'm Files' folder to the search path

addpath('m Files');

5
6

clear; clc;

7
8
9
10
11

nr=500; % Number of observations in simulation

advOpt.distrib='Normal';
% Distribution assumption
% Transition matrix (this also defines the value
% of k)

12
13
14

Coeff.p=[.95 .1; ...

.05 .9];

15
16
17
18

% Setting up which variables at indep will have

%switching effect
Coeff.S=[0 1];

19
20
21
22

% Setting up the coefficients at non switching

%parameters (each row is each % variables
%coefficient). The order is the same as Coeff.S

23
24
25
26

% Setting up the coefficients at non switching

%parameters
Coeff.nS param=0;

27
28
29
30
31

%
%
%
%

Setting up the coefficients at non switching

parameters (each row is each variables coefficient
and each collum is each state). This example has
1 switching parameter and 2 states

32
33

17

Coeff.S param(1,1)= .5;

This script is available in the zip file as Example MS Regress Sim.

20

34

Coeff.S param(1,2)=-.5;

35
36
37
38
39

% Setting up the standard deviavion of the

%model at each state
Coeff.Std(1,1)=0.5;
Coeff.Std(1,2)=0.1;

40
41
42
43

% The explanatory variables used in the simulation

% are always random normal, with
% specific mean and standard deviation

44
45
46

Coeff.indepMean=[1 0];
Coeff.indepStd= [0 0];

47
48
49

% getting the value of k, according to Coeff.p

k=size(Coeff.p,1);

50
51
52

% calling simulation function

[Simul Out]=MS Regress Sim(nr,Coeff,k,advOpt.distrib);

53
54

rmpath('m Files');

The inputs to the simulation function MS Regress Sim are explained within the codes comments. When running the previous
script, a figure with the time series plot of the simulated model will
be created:

21

Figure 2: Output from Example MS Regress Sim.m

This plot will be different for each run of the code. For further
instruction on how to use the simulation function in the case of more
complex models, including the multivariate case, have a look in the
examples scripts provided within the packages zip file.

Using advanced options

When I was writing this markov switching package I always had in

mind that it should be simple and intuitive to use but, at the same
time, be complete in providing different options to the user. In the use
of the fitting function MS Regress Fit.m, all of the different options
of the algorithm are controlled by the input advOpt. The possible
fields are:
advOpt.distrib: Defines the distribution to be used in the maximum likelihood calculation. If advOpt.distrib='Normal', then
the distribution used is the Gaussian, if advOpt.distrib='t', it is
used the t distribution, with one extra parameter (degrees of freedom).
If this field is equal to 'GED' then the distribution for the error is the

22

Generalized Error distribution with one extra parameter (K).

advOpt.std method - Defines the method to be used for the
calculation of the standard errors of the estimated coefficients. The
options are:
advOpt.std method=1: Calculation of the covariance matrix is
performed using the second partial derivatives of log likelihood
function (a.k.a. the Hessian matrix).
advOpt.std method=2: Calculation of the covariance matrix is
performed using the first partial derivatives of log likelihood,
that is, the outer product matrix. This methods approximates
the Hessian matrix with the gradient vector and is a robust choice
for when std method=1 fails to give reasonable values18 . See
Hamilton [1994] for more details in the calculations of standard
errors.
advOpt.useMex: Defines whether to use the mex19 version of
Hamiltons filter in the calculation of likelihood function. The advantage in using the mex version is the increase of speed in the fitting
function. But, in order to use it, youll need a proper C++ compiler.
Please check next topic for more details.
advOpt.constCoeff: Defines the use of constrained/reduced estimation of the model. This is particularly useful when the desired
model has a very particular representation which cannot be addressed
in the usual notation. See the following topic for instructions on how
to use this feature.
advOpt.diagCovMat: Defines the use of a diagonal matrix for
sigma (covariance matrix) in a multivariate estimation, only. If
advOpt.diagCovMat=0 then the whole covariance matrix is estimated from the data.
18

The calculation of the Hessian by numerical differentiation is not guaranteed to provide

a real number. When this happens, the function MS Regress Fit will output NaN (Not
a Number) values for the standard errors.
19
Mex files stands for matlabs executable file and is a way to interface MATLAB with
low level programming languages such as C++ and Fortran.

23

advOpt.optimizer: Defines which Matlabs optimizer to use in

the maximum likelihood estimation of the model.20
advOpt.printOut: Flag for printing out to screen the model in
the end of estimation.
advOpt.printIter: Flag for printing out numerical iterations of
maximum likelihood estimation.
advOpt.doPlots: Flag for plotting fitted conditional standard
deviations and smoothed probabilities in the end of estimation.
Next in Table 1 I show the possible values for each input in advOpt
and also the default values (if no value is assigned).

Input Argument
advOpt.distrib
advOpt.std method
advOpt.useMex

Possible Values
Normal,t or GED
1 or 2
1 or 0

advOpt.constCoeff

Any number or string

e (see next topic for
details)

advOpt.diagCovMat
advOpt.optimizer
advOpt.printOut
advOpt.printIter
advOpt.doPlots

1 or 0
fminsearch,fmincon
or fminunc
1 or 0
1 or 0
1 or 0

Default Value
Normal
1
0
A structure with
fields containing
string e (all
coefficients are
estimated from the
data)
1
fminsearch
1
1
1

Table 1: Default Values for input arguments in advOpt.

20

Special thanks for Zhuanxin Ding for providing a clever numerical transformation in
the transition matrix that allows for the model to be estimated without the need of equality
constraints in the optimization, that is, using functions fminsearch and fmincon. See
Zhuanxin [2012] for details.

24

7.1

Using the mex version of Hamiltons filter

The filter behind the calculation of the filtered probabilities of the

model is computationally intensive and this may be a pain when dealing with large amounts of data. The MS Regress package has a mex
version of the likelihood function that performs the heavy duty of the
model, Hamiltons filter, in a cpp mex file. Depending on the number
of observations in the system, the gain in speed can be higher than
50%, meaning that the mex version of the filter can decrease in half
the time needed to estimate the model.
In order to use it, first youll have to compile the C version of
the filter. The file in question is mex MS Filter.cpp, which is located
at the folder m Files. The cpp file was compiled and tested using
MATLAB 2010a and MS VC 2008. You can get the last one freely on
the internet21 . After installing it, type in MATLAB the command:
1

mex -setup;

This will take you to a set of steps for configuring your default mex
compiler to MS VC 2008. After that, just use the command:
1

mex mex MS Filter.cpp;

in the directory m Files and try running the script

Example MS Regress Fit with MEX.m.
Please note that the .cpp function will NOT compile under MATLABs
native LCC. For others C++ compilers, I havent tested it, hopefully it
will work. If you successively compiled the code with other compiler,
please let me know.
If youre having problems compiling it, please email me and Ill
send you the compiled mex file which should work for Matlab 32bit
version.

7.2 Using advOpt.constCoeff (constraining coefficients)

The MS Regress package also supports constrained/reduced estimation of the model. This means that, for all the coefficients, you can
21

http://www.microsoft.com/express/

25

choose whether you want to estimate the parameter from the data or
if you want to fix it to a specific value. This feature also holds for
multivariate models.
Consider the following case: you know for sure (or at least have a
theory) that one (or more) of the switching coefficients (say column 2
of indep) has the value of 0.5 when the model is in state 1 and value
of -0.5 when the model is in state 2. You now want to know what are
the maximum likelihood estimates for the other coefficients given this
restriction. The package allow for this particular specification (or any
other as matter of fact) to be set. The interface for using constrained
estimation has the following principles:
1. The parameters are grouped with the notation:
nS P aram: All non switching coefficients at indep matrix
(which were chosen with argument S)
S P aram: All switching coefficients at indep matrix (also
chosen with input S)
p: Transition matrix
covM at: Covariance matrix of innovations
df : Degrees of freedom for t distribution (if applicable)
K: Parameter for GED distribution (if applicable)
2. The size of each of these fields are set according to the number of
switching/non switching parameters and the number of states in
the model (value of k). For all of them (except covM at), the cells
iterates over the equations in the system, the columns iterate
the states and the rows iterate the coefficients. For instance, for
nS Param{iEq}, the element (1,1) of this matrix is the first non
switching parameters of equation iEq, the element (2,1) is the
second non switching parameter and so on. For S Param{iEq},
the element (1,2) is the parameter of the first switching variable,
at state 2, equation iEq. The sizes of covMat, df and K also
have to respect the choices made at input S. For instance if the
model is switching in the standard deviation of the innovations,
then covMat should have size {1,k}, otherwise it should be a
{1,1} cell. For the case of multivariate models, covMat should
be sized according to the number of equations in the system.
3. The rule for building the fields is, use the string 'e' (as in estimate) if you want the respective parameter to be estimated from

26

data or a numeric value if you want the respective parameter to

take that value. For instance, using:
1
2

advOpt.constCoeff.S Param{1}={0.5 , 'e' ; ...

'e' , 0.1};

youre telling MS Regress Fit to set the coefficients of the first

switching parameter of equation 1, in state 1, equal to 0.5, the
second switching parameter, state 2, equal to 0.1 and estimate
the rest of them.
Examples should make the structure on the use of advOpt.constCoeff
clear. Consider the estimation of the following model:
State 1 (St = 1)

(48)

yt = 1 + 0.x1,t + 2,1 x2,t + t

(49)

t N (0, 12 )

(50)

State 2 (St = 2)

(51)

yt = 1 + 1,2 x1,t + 0x2,t + t

(52)

t N (0, 22 )

(53)

with:
P=

p1,1 p2,1
p1,2 p2,2

(54)

For the previous specification, the restrictions are:

1,1 = 0
2,2 = 0
In order to estimate this model, these are the options fed to the fitting function Example MS Regress Fit using constCoeff.m:
1
2
3
4
5

% Defining Inputs
k=2;
S=[0 1 1 1];
advOpt.distrib=`Normal';
advOpt.std method=1;

27

6
7
8
9

% Defining constrained parameters

advOpt.constCoeff.nS Param{1}={'e'};
advOpt.constCoeff.S Param{1}={ 0 ,'e' ; 'e', 0}

10
11
12

advOpt.constCoeff.covMat{1}(1,1)={'e'};
advOpt.constCoeff.covMat{2}(1,1)={'e'};

13

advOpt.constCoeff.p={'e','e' ; 'e','e' };

14
15
16
17

% Estimate resulting model

Spec Out=MS Regress Fit(dep,indep,k,S,advOpt);

For the previous code, the command

advOpt.constCoeff.nS Param{1}={'e'};
is telling the function MS Regress Fit to estimate the non switching parameter of the first (and only) equation in the system. The input
advOpt.constCoeff.S Param{1}={ 0 ,'e'; 'e', 0};
fix the switching parameter of indep column 2 to 0 at state 1, indep column 3 equal to 0 at state 2 and estimate the rest of it. The
commands:
advOpt.constCoeff.covMat{1}(1,1)={'e'};
and
advOpt.constCoeff.covMat{2}(1,1)={'e'};
defines the estimation of the covariance matrix for both states. The
last input:
advOpt.constCoeff.p={'e','e'; 'e','e'};
directs the fitting function to estimate all of the transition probabilities.
Consider now the next example, where the model is:
State 1 (St = 1)

28

(55)

yt = 1 + 0.5x1,t + 2,1 x2,t + t

(56)

t N (0, 0.0004)

(57)

State 2 (St = 2)

(58)

yt = 1 + 1,2 x1,t 0.8x2,t + t

(59)

t N (0, 22 )

(60)

with:
P=

0.95 p2,1
0.05 p2,2

(61)

Note that this model has the following restrictions:

1,1 = 0.5
2,2 = 0.8
12 = 0.0004
p1,1 = 0.95
p1,2 = 0.05
The input arguments for advOpt.constCoeff that represent this
model in MS Regress Fit are:
1
2
3
4
5

% Defining inputs
k=2;
S=[0 1 1 1];
advOpt.distrib='Normal';
advOpt.std method=1;

6
7
8
9
10
11
12

% Defining restrictions
advOpt.constCoeff.nS Param{1}={'e'};
advOpt.constCoeff.S Param{1}={0.5, 'e' ; 'e',-0.8};
advOpt.constCoeff.covMat{1}(1,1)={0.0004};
advOpt.constCoeff.covMat{2}(1,1)={'e'};
advOpt.constCoeff.p={0.95,'e' ; 0.05,'e'};

As one can see, any parameter of the model can be constrained to

a particular value, including the transition matrix22 .
22

Be aware that when constraining the transition matrix, the probabilities have to sum
to one in each column.

29

8 Comparing results against Hamilton

[1989]
The markov switching specification of Hamilton [1989] is naturally
a benchmark in this class of models and I get an unusual amount
of emails regarding matching the results of the paper by using my
package. This section will shed some lights in this issue.
The markov switching model of Hamilton [1989] is specified as:
yt S t =

4
X

i (yti Sti ) + t

(62)

i=1

t N (0, 2 )

(63)

As you can see from previous equation, the independent variables

are conditional on the states, an unobserved process. That means
P
that the regressors, 4i=1 i (yti ti ), are also unobserved prior
to estimation. My package was not built to deal with this type of
setup but a two step estimation process is identifiable. Consider the
following notation for Hamiltons Model:
z t = yt S t

(64)

This is equivalent to:

zt =

4
X

i zti + t

(65)

i=1

By rearranging 62 we get:
yt = St + zt

(66)

where zt are the residuals from this reduced model. The two steps
for the approximation which can be used to estimate Hamiltons model
are:
1. Estimate using MS Regress Fit:
yt = St + t
2

(67)

t N (0, )

(68)

St = 1, 2

(69)

30

2. Retrieve the t vector and regress it on four lags:

t =

4
X

i ti + t

(70)

i=1

t N (0, 2t )

(71)

Note that the parameter 2t from last equation will approximate

of Hamiltons model. Next, I show the estimated parameters23
from Hamilton [1989] and the parameters from my approximation:
2

23

The GNP data is available at http://weber.ucsd.edu/~jhamilto/software.htm

31

Parameter
1
2
p1,1
p2,2
2

Hamilton [1989]
1.16
-0.36
0.9
0.75
0.866

Two-step MS Regress Fit

1.101
-0.48
0.906
0.682
0.988

1
2
3
4

0.01
-0.06
-0.25
-0.21

0.126
0.104
-0.133
-0.104

Table 2: Comparison of parameters estimates from Hamilton [1989]. The

same model and data is fitted by a two step procedure with function
MS Regress Fit. The estimates of the coefficients are then compared to
the ones in the paper.

From Table 2, one can see that the parameters are fairly comparable for the markov regime switching part of the model. For the autoregressive part, they are not as comparable as for the first part but,
since most of the i are not statistically significant, the approximation
performance is still good. When looking at the regimes probabilities
(Hamilton [1994], page 697), it is also clear that they are very similar.
The code for this estimation is available within the packages zip file
(Script Hamilton Comparison.m).

Advises for using the package

You probably want to apply the package to your own time series. This
topic will set some advices for using the fitting function.
1. For a better convergence, always check the scale of your dependent and independent variables. For instance, lets say the
explained variable varies from 0.1 to 0.1, while one of the independent varies from 1000 to 1000. If you estimate with this
setup (without any correction) the algorithm may not converge
well24 . In this case just divide the corresponding independent
24

This is usually associated with the algorithm converging to -Inf in the log likelihood

32

variable by 1000 so that they are correctly scaled. This gentlemans (linear) transformation will help the optimization functions (fminsearch, fminunc or fmincon) to find the set of maximum likelihood parameters.
2. Always try to estimate simple models. For instance, dont try
to estimate any model with k > 3 and number of explanatory
variables higher than 4. The models size (number of parameters) grows exponentially as n and k grows. For instance, in a
univariate framework, if k = 5, n = 4 (4 explanatory variables)
and youre switching in all coefficients, then the model has 50
parameters to be estimated from data, which is definitely too
much for a gradient descent method (fmincon function). Dont
get me wrong, the package will try to estimate it, but the solution is probably a local maximum and you cant really trust the
output you get.
3. If using constrained estimation, make reasonable choices for the
restricted coefficients, otherwise the function will have a hard
time in finding the maximum likelihood estimates.
4. With the last update in the package, the choice for the optimizer
to be used in the maximum likelihood was made available. If
you are having problems in the estimation, try to change the
optimizer and check whether the problem still persists.
If after those steps youre still having problems converging to a
solution, send me an email with a nice personal introduction and an
attached zip file containing:
The scripts youre running (the main .m file). This should be
send with all auxiliary m-files and data. The main program
should run without any problem except the one youre reporting.
The error message (if there is one, could be in .txt or just in the
email scope).

10

Possible error messages

The code in MS Regress Fit is not guaranteed to run flawlessly for

all versions of Matlab and the quality of the input data can also affect
the estimation. With time (and lots of emails!) I was able to map out
function.

33

the error messages that people can get when running the code. Next,
Table 3 you can find these messages, along with respective causes and
solutions. If you have an error message which is not there, please let
me know and Ill add it to the table.

Error Message
FMINCON/FMINSEARCH
cannot continue because
user supplied objective
function failed with the
following error: Variable
indep nS is used as a
command function.

Error using svd.

Input
to SVD must not contain
NaN or Inf.

Cause

Solution

This is an issue with your

matlab version, which is
not accepting the notation
used in the newer versions.

Get a newer version of

Matlab.

This is an issue that usually appears when working with optimizer fmincon. Basically the optimizing function got stuck
in a particular value of the
log likelihood function and
was not able to get out
on its own. Usually this
means that the input data
is badly shaped (e.g. outliers) or the case of bad
starting coefficients.

One simple solution which

I have seen to work quite a
lot is to divide the dependent variable by 100. This
simple linear transformation will make the log likelihood function smoother
and will help the optimizer
in finding the maximum
likelihood solution. Another solution which I also
have seen to work is to look
for outliers and removing
(or smoothing) then from
the data. At last, if none
of them worked, try changing the optimizer to fminsearch or fminunc.

Table 3: Possible error messages, causes and solutions when using function
MS Regress Fit.

34

11

Frequently asked questions

The first version of MS Regress was published in August of 2007. Since

then I have been receiving a significant amount of questions by email.
Some of these questions have a tendency to reappear over time. Here
I document the most common cases.
Q: Can I use your code for teaching and researching?
A: Yes! Fell free to use it or change it in any way you find best. My
only request is that you provide the original source, citing the paper
available at SSRN (see next topic, Citing the Package).
Q: I am trying to replicate the results of a particular paper and the parameters Im finding with MS Regress do not match the ones reported
in the article. Whats the problem?
A: First, make sure that the data and the specification are matching. For example, if the paper uses a dataset from 2010-2011 and
you use it from 2008-2011, the estimates will naturally be different.
This is also true if the underlying model is different. But, if you do
match data and the model then you should be aware that the estimation of the MS model is related to a optimizing algorithm finding a
maximum likelihood estimate. For the same data and specification,
different algorithms (or different software) will find different solutions.
The divergence should be small, but it will still be there.
Q: I need to apply a particular statistical test in the post estimation of
my model. Is there a function for such?
A: The package does not provide functions for hypothesis testing. But
in the output of the estimation function I made sure that the user gets
a lot of information so that he can apply any test on its own.
Q: The example scripts are not running in my computer. Can you
check whats wrong?
A: It depends. Before you email me, please make sure that this is not
something in the use of Matlab itself. The help files of Matlab are a
good source of material for learning to run scripts and functions. If

35

you still cant figure it out, perhaps it is a version issue. Try the code
in a newer version of Matlab and check how it goes. If still there are
problems, fell free to email me.
Q: Im rewriting some parts of MS Regress but Im having problems
with the programming part. Can you help me with the codes?
A: Sorry but no. Helping people with programming in Matlab turned
out to have a high cost in my side. If I helped everyone I would not
have time to do my normal job. While I wont help you directly with
the coding, I am willing to give you with some general guidelines of
how (and where) to change the code.
Q: Im estimating my model with MS Regress and and the optimization converges. However, I noticed that some of the standard error
and p-values are NaN (not a number). Why is that?
A: If you are finding NaN values for standard error or p-value, then
it is because the numerical algorithm that calculates the first and/or
second derivative has failed at that particular part. My advice is to either use an alternative optimization function (e.g. fmincon) or change
the method that calculates the standard errors.
But, there a specific case where NaN are expected. If you are estimating the model with fminsearch or fminunc then the last row of
the transition matrix is not estimated, but calculated indirectly from
the other probabilities. Therefore it is not possible to calculate the
first/second derivative and the standard error and p-value are not
available.
Q: Are there other versions of the package written in different programming languages?
A: Yes, there are. Some time ago I wrote a version of MS Regress for
R. The package can be downloaded here:
https://r-forge.r-project.org/projects/rmetrics/
or installed with the following command in R:

36

install.packages(fMarkovSwitching, repos=http://R-Forge.R-project.org)
But be aware that given time constraints, Im no longer maintaining
this package.
Q: Are you interested in working in a consultancy job?
A: Yes. Fell free to drop me an email and we will discuss it.

12

Reporting a bug

Im very happy to hear about any sort of bug in the program. If you
have found one please report it to my email containing:
Name of functions and lines of alleged bug (if applicable).
Reason why you think it is a bug.
Zip file with codes youre running (including data and scripts).
MATLAB s error message (if applicable).
And Ill happily look into it.

13

Citing the package

If you have used the package for research or teaching, make sure you
cite the code, not just for acknowledgment but also for replication of
results. My suggested citation is to use the paper from SSRN25 :
Perlin, M. (2014) MS Regress - The MATLAB Package for Markov
Regime Switching Models. Available at SSRN: http://ssrn.com/
abstract=1714016 or http://dx.doi.org/10.2139/ssrn.1714016

14

Final remarks

The interest of this paper was in presenting the main features of the
MATLAB package MS Regress. As one can see, the interface of the
25

Available in http://ssrn.com/abstract=1714016.

37

software is quite intuitive and it should be flexible enough to handle personalized markov switching specifications without any change
in the original code. For any doubts which are not clear from this
document, fell free to contact me at [email protected].

38

References
Carol Alexander. Market Risk Analysis: Practical Financial Econometrics. Wiley, 2008.
Chris Brooks. Introduction to Econometrics. Cambridge University
Press, 2002.
James Hamilton. A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57
(2):35784, March 1989. URL http://ideas.repec.org/a/ecm/
emetrp/v57y1989i2p357-84.html.
James Hamilton. Time Series Analysis. Princeton University Press,
1994.
James Hamilton. Regime switching models. Palgrave Dictionary of
Economics, 2005.
J. Kim and R. Nelson. State Space Model with Regime Switching:
Classical and Gibbs-Sampling Approaches with Applications. The
MIT Press, 1999.
Ruey Tsay. Analysis of Financial Time Series. John Wiley and Sons,
2002.
Peijie Wang. Financial Econometrics. Taylor and Francis, 2003.
Ding Zhuanxin. An implementation of markov regime switching model
with time varying transition probabilities in matlab. SSRN elibrary,
2012. URL http://ssrn.com/abstract=2083332.

39