J Math Model Algor

DOI 10.1007/s10852-012-9214-4

A Higher-Order Hidden Markov Chain-Modulated

Model for Asset Allocation
Xiaojing Xi Rogemar Mamon Matt Davison

Received: 21 January 2012 / Accepted: 30 October 2012

Springer Science+Business Media Dordrecht 2012

Abstract This paper presents an analysis of asset allocation strategies when the
asset returns are governed by a discrete-time higher-order hidden Markov model
(HOHMM), also called the weak hidden Markov model. We assume the drifts
and volatilities of the asset returns switch over time according to the state of the
HOHMM, in which the probability of the current state depends on the information
from previous time-steps. The switching and mixed strategies are studied. We
use a multivariate filtering technique in conjunction with the EM algorithm to
obtain estimates of model parameter at a given time. This, in turn, aids investors
in determining the optimal investment strategy for the next time step. Numerical
implementation is applied to data on Russell 3000 value and growth indices. We
benchmark the respective performances of portfolio using three classical investment
measures.
Keywords Markov chain Regime-switching Filtering Investment strategy
Portfolio performance

1 Introduction
It is well documented in the asset allocation literature that the inclusion of market
regime-switching dynamics has considerable impact on the optimal portfolio strategy

X. Xi R. Mamon M. Davison
Department of Applied Mathematics, University of Western Ontario, London, ON, Canada
R. Mamon (B) M. Davison
Department of Statistical and Actuarial Sciences, University of Western Ontario,
1151 Richmond Street, London, ON, Canada N6A 5B7
e-mail: [email protected]
M. Davison
Richard Ivey School of Business, University of Western Ontario, London, ON, Canada

J Math Model Algor

of individual investors. In financial portfolio management, the portfolio risk cannot

be entirely eliminated although it can be controlled with an optimal asset allocation
strategy combining different types and amount of investment. Each investment has
its own unique risk and return characteristics. A well-designed investment aims at
the maximisation of expected portfolio return whilst controlling the level of risk. A
fundamental example of the single-period mean-variance asset allocation problem is
given by the Nobel prize-winning work of Markowitz [24], in which the variance is
employed as a measure of risk and the efficient allocation of wealth amongst different
investment classes is provided. Since practical asset allocation problems involve
inter-temporal decisions, Samelson [30] and Merton [26] considered asset allocation
problem in a multi-period model and in a continuous-time model, respectively. In
particular, Merton utilised stochastic optimal control theory to derive a closedform solution for an asset allocation strategy under certain assumptions. A key
assumption of these early works in the literature is that the dynamics of asset returns
are linear processes with constant coefficients. However, the state of the economy
and the financial market varies randomly over time. Investors are concerned with
regime-switching uncertainty affecting the portfolio return. Such uncertainty affects
the future payoffs and therefore could alter the optimal asset allocation. Ang and
Bekaert [3] introduced a regime-switching model with time-varying correlations and
volatilities for asset allocation. They reported evidence of shifting regimes in the
US, UK and German equity markets. In a subsequent study, Ang and Bekaert [4]
expanded the lists of markets and assets in the investigation of optimal asset
allocation under a regime-switching framework. Their out-of-sample test shows that
the regime-switching strategy dominates a non-regime dependent strategy. Bauer
et al. [5] observed the tendency of changing correlation and volatility amongst assets,
and considered a regime-switching technique for portfolio optimisation. An out-ofsample backtesting was applied on a six-asset portfolio consisting of equities, bonds,
commodities and real estate. The results demonstrated a significant information gain
from using a regime-switching strategy. Guidolin and Timmermann [18] considered
asset allocation decisions under a regime-switching model for asset returns with four
separate regimes. It was found that the optimal allocations vary considerably across
these states and change over time as investors revise their estimate of the state
probabilities.
Various works that support model assumptions in which parameters change over
time in accordance with the evolution of an unobserved Markov chain have been
proposed. Elliott and van der Hoek [15] put forward a model for the rates of asset
returns driven by a Markov chain in discrete time. Their work features filtering and
prediction techniques in the model identification and outlines how their method
could be applied to the asset allocation problem using mean-variance type utility
criterion. Graflund and Nilsson [17] investigated dynamic portfolio selection within a
Markovian switching framework. Their results highlight the economic importance of
regimes and suggest that ignoring the regime will require significant compensation.
In the study of Ammann and Verhofen [2], Markov Chain Monte Carlo methods
were applied to estimate a multivariate regime-switching model. Two clearly separable regimes characterised by different mean returns, volatilities and correlations were
found. The results of their out-of-sample backtest suggests that the buy-and-hold
strategy based on regime-switching model can be profitable. Bulla et al. [7] focused
on daily stock market return series at five major regional markets over the last four

J Math Model Algor

decades. They presented an out-of sample performance analysis with transaction

costs taken into account and concluded that the strategy is improved by considering
Markovian switching model. Erlwein et al. [16] developed and compared investment
strategies in allocating funds to either growth or value stocks, whose price dynamics
are driven by a hidden Markov model (HMM). Their investigation shows that the
HMM-based strategies are more stable and outperform the pure growth strategy in
terms of higher Sharpe ratios and lower variance of the performance. Elliott et al. [14]
considered a mean-variance portfolio selection problem where the appreciation rate
of the risky asset is modulated by a continuous-time Markov chain. They employed
the gauge transformation technique to obtain robust filters and developed the filterbased Expectation Maximisation (EM) algorithm in calculating the estimates of the
unknown parameters. An explicit solution to the mean-variance portfolio problem is
derived using the filtering results.
The original form of HMM is now in widespread use because of its ability to
capture the switching of market or economic states. However, there is growing
evidence that many financial series have longer memories than may be captured
by original form HMM models. Some even suggest that financial time series are
of the long memory form pioneered by Hurst [20]. Lobato and Savin [21] and Ray
and Tsay [28] found evidence of longer-range dependence in the volatility of S&P
500 returns. McCarthy et al. [25] found evidence of long memory in percentage
changes on Treasury debt security yields. However, Couillard and Davison [10]
noted that caution must be taken in applying statistical tests for longer-range Hurstlike dependence. Some effort has been made to develop long memory financial
models. For instance, Dajcman [11] examined a time varying long memory parameter
for eight European stock market returns by using an auto-regressive fractionally
integrated moving average model. Others suggest that much of this behaviour can
be explained using shorter memory models, an approach that this paper also takes.
Maheu [23] suggested that GARCH models can in some circumstances account for
the long memory property found in financial market volatility.
There is a vast literature devoted to modelling longer-range dependence using
single-state stochastic models. Unfortunately, Rydn et al. [29] suggested that an
ordinary HMM cannot describe the stylized fact of very slow decay in return
autocorrelations. So some extension of the HMM literature is needed. One approach
is to use a hidden semi-Markov model. For example, Yu et al. [34] developed a
recursive formula to estimate the Hurst parameter corresponding to a second-order
HMM. They showed that this approach could capture longer-range dependence in
web server workload when the distribution of at least one state is heavy tailed. In
a more financial application, Bulla and Bulla [6] explored goodness of fit of two
hidden semi-Markov models to 18 pan-European sector return indices, obtaining
promising results. In this the current work, we take a related approach which,
although motivated by [6, 34], is slightly different. We choose to extend the singlestate HMM models to involve multiple time lags, much as a simple AR(1) model
can be extended to AR(q) by using multiple time lags. We call this approach the
higher-order hidden Markov model (HOHMM).
In this paper, we apply an HOHMM, in particular a second-order HMM, to an
asset allocation problem. Our financial motivation for this comes from the fact that
practitioners often like to divide a market not only into high and low volatility states,
but also into trending and choppy market states. A one-state model is unlikely

J Math Model Algor

to pick this up, since on average returns are serially uncorrelated because some of
the time they have a positive serial correlation and at other times a negative serial
correlation. The hope is that a two-state model might better capture this behaviour.
Also, as mentioned by Solberg [32], the real significance of HOHMM is to establish
that the Markov chain assumption is not really as restrictive as it first appears. One
is not limited to a dependence on just one prior time epoch but can make the
dependency extend to any finite number of prior epochs, thereby capturing more
information from the past. This, in turns, widens the literature on models aiming to
reflect longer-range dependence in financial models.
In the HOHMM model the transition matrix between one state and the next state
is itself dependent on the information in the prior states. An nth-order Markov chain
is dependent on the prior n state. The higher the order, the more extended the
dependency, and therefore more information from the past can be captured. Xi and
Mamon [33] proposed an HOHMM for discrete-time continuous-range observations
and provided a detailed implementation of the model to a financial dataset. Hess [19]
considered conditional CAPM strategies based on regime forecasts from an autoregressive Markov regime-switching behaviour with lag two. The improvement of the
portfolio performance by using the proposed strategy is examined through in-sample
and out-of-sample analyses. An application of higher-order Markovian switching
model for risk measurement is presented by Siu et al. [31]. Other applications of
HOHMM, such as in option pricing, can be found in Ching et al. [9].
In this paper, we investigate optimal investment strategies for asset allocation
under a weak Markov-switching framework. In particular, we assume the log returns
of risky assets are modulated by a second-order multivariate Markov chain, whose
current behaviour depends on its behaviour at the previous two time steps. The
states of the higher-order Markov chain are interpreted as states of the economy.
Compared to the previous research conducted in [33], we extend the single-variate
HOHMM to a multivariate case by modifying the Radon-Nikodm derivative. This
extension allows us to investigate the application of HOHMM involving multivariate
financial series, such as those occurring in asset allocation. We use the same asset
allocation strategies from [16]. Compared to their research, we relax the Markov
assumption by increasing the first-order HMM to second-order HMM. The filtering
technique for HOHMM is implemented on updated market data, which includes the
period of the subprime crisis. The numerical results show how an HOHMM captures
information during a crisis period and the resulting impact on the strategy. The
HOHMM has the advantage that it can capture the longer-ranging dependence of
the states of the market, and therefore it is more appropriate when memories are
evident in financial series. From the investors view, tactical investment decisions
require the evaluation of the expected future payoff on risky assets. More economic
insights can be gained if relevant historical information can be incorporated into the
unobservable market state; this will be beneficial to investors from both the economic
and statistical perspectives. Although a higher-order Markov chain, more specifically
a Markov chain of order higher than two, leads to more information incorporated
in the HMM, the number of model parameters involved increases exponentially.
Ching et al. [8] apply a higher-order multivariate HMM to a sequence of multivariate categorical data and show that an nth-order, s-variate, N-state Markov chain
model requires ns2 N 2 parameters. To facilitate the dynamic estimation of this huge
number of parameters, we use a transformation that converts an HOHMM into a

J Math Model Algor

regular HMM thereby enabling the estimation algorithm to perform smoothly. The
transformation, which is essentially a mapping of states, is employed to eventually
recover the required number of parameters. Our asset allocation strategies rely
on the estimates of parameters and forecasted return through the mathematical
techniques of HOHMMs.
This paper is organised as follows. In Section 2, we present the multidimensional HOHMM filtering and estimation techniques. HOHMM filtering procedure
is applied to the Russell 3000 value and growth indices data, whose log-returns are
assumed to follow a normal distribution with regime-switching dependent on two
previous time epoch. The EM algorithm is then applied to obtain the online recursive
estimates of the model parameters. In Section 3, we utilise the optimal estimates
to forecast the two indices and conclude that a two-state HOHMM is sufficient to
capture the characteristics of our data based on four error metrics. We investigate
an investment strategy switching between the Russell 3000 value and growth indices.
The switching decision determined by the one-step ahead forecasts return of each
index. In Section 5, a mixed investment strategy is considered. The optimal weights
of investment between the two indices are obtained by solving a mean-variance
problem under the regime-switching setting. The estimation of the optimal weights
incorporates the parameter estimates as well as the states of a higher-order Markov
chain. Portfolio performance is investigated in Section 6, where we use three classical
measures for benchmarking. Furthermore, a bootstrap analysis is used to compare
the stability of portfolios with various level of transaction costs. Section 7 concludes
the paper.

2 Filtering and Parameter Estimation

Let (, F , P) be a complete probability space under which xk is a Markov chain
with finite-state space in discrete time (k = 0, 1, 2 . . .). To simplify the discussion and
present a complete characterization of the parameter estimation, we only consider a
second-order Markov chain. That is,
P (xk+1 = xk+1 |x0 = x0 , . . . , xk1 = xk1 , xk = xk )
= P(xk+1 = xk+1 |xk1 = xk1 , xk = xk ).
Without loss of generality, the N-state higher-order Markov chain takes value
from the canonical basis {e1 , e2 , . . . , e N } R N , where ei is the vector with unity
in the ith element and zero elsewhere. We interpret xk , ei  as the event that the
economy is in the state i at time k. Here ,  denotes the inner product in R N . The
element almv , l, m, v 1, . . . , N, of the transition probability matrix A refers to the
probability that the process enters state l given that the current states is the mth state
and the previous state was in v.
Instead of studying the higher-order Markov chain directly, we introduce a
mapping , to embed the second-order Markov chain into the first-order Markov
chain, and then apply the regular filtering method. This idea is analogous to the
embedding of higher-order ODEs into a system of first-order ODEs and solving the
system by regular methods, see Abell and Braslton [1]. The mapping is defined by
(er , es ) = ers , for 1 r, s N,

J Math Model Algor

2

where ers is an R N unit vector with unity in its ((r 1)N + s)th position. Note that
(xk , xk1 ), ers  = xk , er xk1 , es 
represents the identification of the new first-order Markov chain with the canonical
basis. We also define the new N 2 N 2 transition probability matrix  of the new
Markov chain by

almv if i = (l 1)N + m, j = (m 1)N + v
ij =
0
otherwise.
Note that at time k, each non-zero element ij represents the probability
ij = almv = P(xk = el |xk1 = em , xk2 = ev ),
and each zero represents an impossible transition. It is known [31] that the new
Markov chain (xk , xk1 ) has a semi-martingale representation
(xk , xk1 ) = (xk1 , xk2 ) + vk ,

(1)

where {vk }k1 is a sequence of R N martingale increments. In a comprehensive

monograph on HMM, MacDonald and Zucchini [22] devoted a section introducing
the HOHMM and give a detailed example of transforming a second-order two-state
Markov chain into a regular two-state Markov chain. An efficient recursive algorithm
for computing the likelihood from consecutive observations under a second-order
HMM is given. However, no application of the second-order HOHMM MLE is
given in this book. Du Preez et al. [27] developed a computing algorithm to reduce
any HOHMM to a corresponding first-order HMM. The algorithm is applied to
language recognition. In contrast to their research objectives, we focus on financial
time series applications. In particular, we obtain a reduced first-order HMM using a
transformation and estimate parameters of asset price logreturns.
g
Let yk = (y1k , y2k , . . . , ydk ) be a d-dimensional process. Each component yk , 1
g d, is the sequence of log returns of an asset price with the dynamics
g

yk+1 = f g (xk ) + g (xk )zk+1 .

Here {zk } is a sequence of N(0, 1) independent, identically distributed (IID)
random variables and independent of x. The function f g and g are determined by
g
g
g
g
g
g
the vectors fg = ( f1 , f2 , . . . , f N ) and g = (1 , 2 , . . . , N ) in R N , and f g (xk ) =
g
fg , xk  and g (xk ) =  g , xk  represent the mean and variance of yk , respectively
and  denotes the transpose of a matrix. Note that all components of the vector
observation process have the same underlying higher-order Markov chain.
It must be noted that we do not observe the underlying higher-order Markov chain
from the financial market directly. Under the real world measure P, the state xk is
contained in the noisy observations yk , k 1. We aim to filter the noise out of
the observations. By the Kolmogorov Extension Theorem, there exists a reference
under which the observation yk are N(0, 1) IID random
probability measure P,
variables and therefore P is deemed to be an easier measure to work with. The
filters are derived under the reference measure. We perform a measure change to
construct the real-world measure P from the ideal-world measure P by invoking a

J Math Model Algor

discrete-time version of Girsanovs theorem. Let (z) denote the probability density
function of a standard normal random variable z. For each component g, define
( g (xl1 )1 (yl f g (xl1 )))
,
g
g (xl1 )(yl )
g

l =

and the Radon-Nikodm derivative of P with respect to P,

k =

k
d 


(2)
dP
|
d P Yk

= k , is given by

l , k 1, 0 = 1.

(3)

g=1 l=1

To obtain the estimates of (xk , xk1 ) under the real world measure, we first
Then Bayes
perform all calculations under the reference probability measure P.
theorem is employed to relate the conditional expectation under two different
measures. Note that we can also consider another reference probability measure P
under which the yk s are N(0, 2 ), = 1. In such a case, we define
g
l =

( g (xl1 )1 (yl f g (xl1 )))

.
g
( g (xl1 )1 yl )
g

Based on our numerical experiment, since is much larger than , the speed of
convergence with is ten steps slower than using .
Let us derive the conditional expectation of (xk , xk1 ) given Yk under P. Write
ij

pk = P(xk = ei , xk1 = e j|Yk ) = E[(xk , xk1 ), eij|Yk ],

ij

(4)

NN
N
with pk = ( p11
k , . . . , pk , . . . , pk ) R . Using Bayes theorem, we have

pk = E[(xk , xk1 )|Yk ] =

k (xk , xk1 )|Yk ]

E[
.
k | Yk ]
E[

(5)

k (xk , xk1 )|Yk ] and 1 = (1, . . . , 1) R N2 , we see that

Defining qk = E[
N


(xk , xk1 ), eij = (xk , xk1 ), 1 = 1,

i, j

so that
k (xk , xk1 ), 1|Yk ] = E[
k |Yk ].
qk , 1 = E[
With Eqs. 5 and 6, we get the explicit form of the conditional distribution
pk =

qk
.
qk , 1

(6)

J Math Model Algor

Now, we need a recursive filter for the process qk in order to estimate the state
process (xk , xk1 ). Define the diagonal matrix Bk by

Bk =

b 1k

..

.
b kN

..

.
b 1k

..

(7)

b kN
where
b ik =

d
g
g
g

((yk fi )/i )
.
g
g
i (yk )
g=1

(8)

k = E[Xk |Yk ] and (X)k =

Notation For any Yk -adapted process Xk , write X
k Xk |Yk ]. Again invoking Bayes theorem, we have
E[
k =
X

(X)k
.
k | Yk ]
E[

(9)

To estimate the parameters of the model, recursive filters shall be derived for the
following processes:
J rst ,the number of jumps from (es , et ) to state er up to time k.
Ors
k , the occupation time of the higher-order Markov chain spent in state (er , es )
up to time k,
3. Ork , the occupation time spent by the higher-order Markov chain in state er up to
time k,
4. Tkr (h), the level sum for the state er , where h is a function with the form h(y) = y
or h(y) = y2 .

1.
2.

To obtain on-line estimates for the quantities of the above four related process,
we shall take advantage of the semi-martingale representation in Eqs. 1 and 9.
We can then obtain recursive equations for the vector quantities Jkrst (xk , xk1 ),
r
r
Ors
k (xk , xk1 ), Ok (xk , xk1 ) and Tk (h)(xk , xk1 ). The recursive relation of these
vector processes and qk under a multi-dimensional observation set-up are given in
the following proposition.
Proposition 1 Let Vr , 1 r N be an N 2 N 2 matrix such that the ((i 1)N + r)th
column of Vr is eir for i = 1 . . . N and zero elsewhere. If B is the diagonal matrix
def ined in Eq. 7, then
qk+1 = Bk+1 qk

(10)

J Math Model Algor

and
(J rst (xk+1 , xk ))k+1 = Bk+1  (J rst (xk , xk1 ))k
+ b rk+1 est , ers qk , est ers ,

(11)

(Ors (xk+1 , xk ))k+1 = Bk+1  (Ors (xk , xk1 ))k

+ b rk+1 qk , ers ers ,

(12)

(Or (xk+1 , xk ))k+1 = Bk+1  (Or (xk , xk1 ))k

+ b rk+1 Vr qk ,

(13)

(T (h)(xk+1 , xk ))k+1 = Bk+1  (T (h)(xk , xk1 ))k

r

+ h(yk+1 )b rk+1 Vr qk .

(14)

Proof See [33] for an analogous proof of each filter under the single observation
setting.


Similar to Eq. 6, by summing the components, Eqs. 1114 give expressions for
(J rst )k , (Ors )k , (Or )k and (T r (g))k .
Now we make use of the expectation maximization (EM) algorithm to estimate
the optimal parameters. The calculation is similar to the technique as in single
observation set-up. The estimates are expressed in terms of the recursions in Eqs. 11
14, which are provided in the following proposition.
Proposition 2 Suppose the observation is d-dimensional and the set of parameters
g
g
g
{arst , fr , r } determines the dynamics of yk , k 1, 1 g d, then the EM estimates
for these parameters are given by
a rst =


Jkrst
(J rst )k
=
, pairs (r, s), r = s,
st
(Ost )k
O
k

r
T
(T r (yg ))k
frg = kr =
,

(Or )k
O
k


T
r
r ((yg )2 )k 2 frg T
r (yg )k + ( frg )2 O
k
g
.
r = 
r
O

(15)

(16)

(17)

Proof See [33] for an analogous proof of each estimate under the single observation
setting.


g
g
Given the observation up to time k, new parameters a rst (k), fr (k), r (k), 1
r, s, t N are given by Eqs. 1517. The recursive filters for the unobserved Markov
chain and the related process in Proposition 1 can be re-evaluated using the new estimates. Consequently, it allows the algorithm to update the parameters automatically.

J Math Model Algor

3 Forecasting Indices
Suppose an investor wants to choose a portfolio with two investments to diversify
his/her risk. In order to have such diversification the two assets should act differently
during different periods in the economic cycle. For example, growth and value stocks
tend to perform well at different times of the economic cycle, so switching between
the classes at appropriate times may add value. We apply the iterative procedure
derived in the previous section to two weekly datasets of stock indices: Russell 3000
growth and Russell 3000 value indices. The data were recorded from June 1995 to
December 2010; thus there are 783 data points in each dataset. Both indices are
constructed based on the Russell 3000 index, in which the underlying companies are
all incorporated in the U.S. and representing approximately 98 % of the investable
U.S. equity market. Companies within the Russell 3000 that exhibit higher price-tobook and forecasted earnings are used to form the Russell 3000 growth index. This
subindex therefore measures the performance of the broad growth segment of the
US equity market. The Russell 3000 value index includes Russell 3000 companies
with lower price-to-book value and lower forecasted growth values. Therefore the
Russell 3000 value index measures the performance of the value stocks in the US
equity market.
The regime-switching models are developed to capture particular behaviour of
the evolution of an asset price. We segregate the observation data into four intervals
to investigate the index values and returns. Tables 1, 2 and 3 provide descriptive
statistics of the Russell 3000 Index return together with the growth- and valuesubindex returns for the entire period as well as the subperiods. The descriptive
statistics demonstrate the possible segregation of the actual data into different states
according to the levels of mean and volatility. We find the subperiods characterised
by different levels of mean and volatility. For example, we can see that the log return
yk has a higher volatility when the mean is negative, and vice versa. If the data has
only one state, the model will collapse to one regime. As a result, the estimated
parameters of each state will be close to each other.
We consider the two indices as a two-dimensional observation process. The
dynamics of the log returns are given by
RV
RV
yk+1
= log RValue(k+1)
= f RV (xk ) + RV (xk )wk+1
RValue(k)
RG
RG
yk+1
= log RGrowth(k+1)
= f RG (xk ) + RG (xk )wk+1
RGrowth(k)

Table 1 Summary statistics of Russell 3000 growth returns

Max
Min
Median
Mean
Std
Skewness
Kurtosis

Entire data

06/9507/98

07/9809/03

09/0308/08

09/05/0812/31/10

0.1659
0.1806
0.0026
0.0010
0.0299
0.4478
5.0859

0.0669
0.0462
0.0052
0.0048
0.0196
0.0281
0.1581

0.1659
0.1683
0.0007
0.0008
0.0379
0.0721
2.6943

0.0429
0.0522
0.0012
0.0009
0.0184
0.3730
0.2936

0.1090
0.1806
0.0034
0.0005
0.0399
0.8401
3.7011

J Math Model Algor

Table 2 Summary statistics of Russell 3000 value return
Max
Min
Median
Mean
Std
Skewness
Kurtosis

Entire data

06/9507/98

07/9809/03

09/0308/08

09/05/0812/31/10

0.1381
0.2167
0.0025
0.0011
0.0266
0.8031
8.1906

0.0554
0.0551
0.0052
0.0042
0.0168
0.2649
0.5851

0.0719
0.1162
0.0007
0.0001
0.0258
0.2585
1.6751

0.0615
0.0609
0.0028
0.0011
0.0187
0.4757
0.7746

0.1381
0.2167
0.0031
0.0005
0.0465
0.7317
4.0000

where
f RV = ( f1RV , . . . , f NRV ) R N , f RG = ( f1RG , . . . , f NRG ) R N ,
RV = (1RV , . . . , NRV ) R N , RG = (1RG , . . . , NRG ) R N ,
are governed by the same HOHMM x. Here, wkRV and wkRG are N(0, 1) IID random
variables independent of each other. The data are processed in batches of ten
observation points. At the end of each pass through the data, f, , A and  are
updated with new estimates using the formulas given in the previous section. These
new estimates are in turn used as initial parameters for the next pass. This means
the parameters are updated roughly every two and a half months. We process
the data in batches in order to lower computational expenses. Furthermore, the
use of data batches is consistent with the idea of suboptimal schemes; see page
15 of Elliott et al. [13]. In our case, such choice of ten data points provides the
best fitting performance of the forecasts to the actual data. Investors can choose
any length of a batch to update their information according to their needs. In our
numerical experiment, we find that updating parameters every two and half month is
sufficient to capture market information. Whilst utilising batches with fewer data
points improves forecasting errors slightly, it does not lead to a better portfolio
performance. Figure 1 displays the plot of the evolution of f RV , f RG , RV , RG and
the transition matrix A under the two-state HOHMM setting.
The optimal investment strategy is developed based on the forecasts of index
returns. To assess the predictive performance of the model, we calculate the onestep ahead forecasts for both indices through the following equations
E[RValuek+1 |Yk ] = RValuek

N


pk , eij exp( fiRV + (iRV )2 /2)

(18)

i, j=1

Table 3 Summary statistics of Russell 3000 return

Max
Min
Median
Mean
Std
Skewness
Kurtosis

Entire data

06/9507/98

07/9809/03

09/0308/08

09/05/0812/31/10

0.1204
0.1986
0.0024
0.0012
0.0272
0.6741
5.8949

0.0611
0.0507
0.0048
0.0045
0.0177
0.1051
0.2693

0.0995
0.1267
0.0008
0.0002
0.0302
0.2204
2.0180

0.1204
0.1986
0.0038
0.0015
0.0465
0.6765
2.9595

0.1204
0.1986
0.0042
0.0000
0.0426
0.8040
3.9150

J Math Model Algor

Estimates for fgrowth

Estimates for growth

0.035

0.16
f1

0.03

0.14

f2

0.12

0.02

growth

fgrowth

0.025

0.015
0.01

0.06

0.005
0

0.1
0.08

0.04
0

10

20

30
40
50
Number of steps

60

70

80

10

20

Estimates for f

60

70

80

Estimates for value

value

0.03

0.16
f1

0.025

0.14

f2

0.12
value

0.02
fvalue

30
40
50
Number of steps

0.015

0.1
0.08

0.01

0.06

0.005
0

0.04
0.02
0

10

20

30
40
50
Number of steps

60

70

80

10

20

30
40
50
Number of steps

60

70

80

Weak HMM transition probability

1
a111

0.8

Probability

211

a112

0.6

a212
a

121

0.4

a221
a122

0.2

222

10

20

30
40
50
Number of steps

60

70

80

Fig. 1 Evolution of parameter estimates under the 2-state setting

E[RGrowthk+1 |Yk ] = RGrowthk

N


pk , eij exp( fiRG + (iRG )2 /2).

(19)

i, j=1

In this paper, we do not model the correlation amongst assets explicitly. However,
the two indices are governed by the same hidden higher-order Markov chain, and

Table 4 Error measures: root

mean square error (RMSE),
absolute mean error (AME),
relative absolute error (RAE)
and absolute percentage error
(APE) for one-step ahead
forecasts under 1-, 2- and
3-state HOHMM set-ups

1-state
HOHMM

2-state
HOHMM

3-state
HOHMM

RAE value
APE value
MAE value
RMSE value

0.0966
0.0188
38.6423
54.4720

0.0976
0.0192
39.0510
54.6768

0.1044
0.0216
41.7675
56.7974

RAE growth
APE growth
MAE growth
RMSE growth

0.1159
0.0215
43.0295
64.4588

0.1170
0.0218
43.4276
64.7098

0.1241
0.0241
46.0558
66.5664

J Math Model Algor

Table 5 Error measures
restricted to financial crisis
period: root mean square error
(RMSE), absolute mean error
(AME), relative absolute error
(RAE) and absolute
percentage error (APE) for
one-step ahead forecasts under
1-, 2- and 3-state HOHMM
set-ups

1-state
HOHMM

2-state
HOHMM

3-state
HOHMM

RAE value
APE value
MAE value
RMSE value

1.0365
0.0191
13.5314
17.8370

0.9836
0.0127
15.8644
16.4198

0.9828
0.0142
16.4198
20.0430

RAE growth
APE growth
MAE growth
RMSE growth

0.9530
0.0088
11.9003
13.5887

0.4213
0.0043
5.7440
8.3321

0.6690
0.0075
10.0917
12.0772

thus, they are correlated implicitly. Actual filters with correct correlation structure
between Brownian motions driving the logreturns of growth and value indices will
presumably be better. So, one may view that this study is as a lower bound for the
validity of a larger study.
Our goal in this exercise is not to obtain a better forecast of either growth or
value returns over a short time horizon: the relative size of the random to the
deterministic terms in the stock price model makes this impossible. Instead, our goal
is to determine whether HOHMM can improve our management of a given portfolio.
However, for completeness, we do present an assessment of the goodness of fit of the
one-step ahead forecasts. To make this comparison we use four criteria: root mean
square error (RMSE), absolute mean error (AME), relative absolute error (RAE)
and absolute percentage error (APE) for N = 1, N = 2, and N = 3. The results of
this error analysis are given in Table 4.
The results show that the two-state model tends to outperform the three-state
model in all forecasting metrics. Although the one-state model has a slight improvement, Table 4 shows that the performance of the 2-state model is statistically nearly
indistinguishable with that of the one-state model. However, Table 5 depicts a similar
error analysis restricted to the time window surrounding the financial crisis of 2007
2008. Table 5 clearly shows the advantages of a 2-state model over the one-state
model especially in the modelling of the growth index. These results on forecast
performance are interesting.
But the true test of a model like this is in its application to a trading strategy. We
want to decide if these filter results enable us to better manage a portfolio. That is
the topic of this papers next section.

4 A Switching Investment Strategy

There are various asset allocation strategies that one can devise. Here we focus on
a dynamic asset allocation which assumes active changes to an investment based on
short-term market forecasts for returns. The switching strategy utilise the forecasted
risk-adjusted returns of the indices as signals to switch investments between the
Russell 3000 growth and Russell 3000 value index. The forecasted risk-adjusted
return is calculated by dividing the forecasted returns by the realised volatility of
each index covering the previous 20 data points.

J Math Model Algor

We implement the switching strategy on a 15-year dataset recorded from June

1995 to December 2010. The observation data is divided into 15 intervals and
each interval covers roughly one-year dataset. We suppose a hypothetical starting
investment of $100 and then apply the forecasting method on the period considered.
At the beginning of each interval, the signals from the forecasted risk-adjusted return
on both indices are compared. The full amount is invested in the index with the higher
forecasted risk-adjusted return. We also assume that the transaction cost is a fixed
percentage of total investment. When asset allocation changes, this transaction cost
is subtracted from total investment.
The overall performance of the switching strategy is compared with that of the
pure investment strategy on the basis of the log-return of the terminal wealth. Let
X RG and X RV denote the differences of log-returns from the switching and pure
strategies, which are defined by
RGi
SWi
log
100
100
RVi
SWi
log
,
= log
100
100

XiRG = log

(20)

XiRV

(21)

for i = 1, 2, . . . , 15. Here, SWi denotes the terminal wealth of the portfolio with
switching strategy at the end of the ith interval. RGi and RVi denote the terminal
wealth of the investment, holding 100 % of Russell 3000 growth index and Russell
3000 value index, respectively, at the end of the ith interval. The portfolio performance under varying transaction costs from five basis points (1 bp=0.01 %) to
70 bps is presented in Table 6. In addition, we present the performance of both
switching and pure indices strategies using the usual HMM forecasts. Our study
shows that the HOHMM-based switching strategy has higher values in Mean(X RV )
and Mean(X RG ) than those from HMM-based switching strategy yielding negative
values. HOHMM-based strategy shows higher std(X RV ) and lower std(X RG ) than
those based on HMM strategy. As we can observe, HOHMM-based Mean(X RV )
and Mean(X RG ) are positive and slightly decrease as transaction cost increases. This
means that on average the log return from the switching strategy is higher than
that from the pure index investments. Compared with the pure growth and value
strategies, the HOHMM switching strategy has either the highest or the second
highest terminal value in 15 intervals. Figure 2 displays number of the intervals in
which switching strategy has the highest and the second highest terminal values for
transaction cost varying from 1 bp to 80 bps. We observe, however, high values
of std(X RV ) and std(X RG ), which indicate a high risk of employing the switching
strategy. We next introduce a mixed strategy to address the diversification of risks.

Table 6 Performance comparison for HOHMM- and HMM-based switching strategies with varying
transaction costs
Transaction
cost

5 bps
HOHMM HMM

Mean (X RG ) %
Std (X RG ) %
Mean (X RV ) %
Std (X RV ) %

3.9106
2.4102
11.2393
18.4701

20 bps
HOHMM HMM

0.6895 3.5163
6.7138 2.4701
9.5292 10.8450
15.8552 18.2859

0.9281
6.5882
9.7677
16.0045

50 bps
HOHMM HMM
2.7261
2.7086
10.0547
17.9291

1.4062
6.3598
10.2458
16.3117

70 bps
HOHMM HMM
2.1979
2.9403
9.5266
17.7003

1.7258
6.2259
10.5654
16.5226

J Math Model Algor

Fig. 2 Numbers of the
intervals switching strategy has
the highest and the second
highest terminal values for
varying transaction cost

13
12
11
10
9
8
7
6
5
4
no. of the highest terminal value
no. of the second highest terminal value

3
2
0

10

20

30

40
Basis points

50

60

70

80

5 A Mixed Investment Strategy

Selecting an investment strategy is similar to the asset allocation decision problem
in that one tries to maximise strategy return whilst controlling portfolio risk. The
risk is evaluated in terms of the variance of the portfolios return. The meanvariance problem entails maximising the expected portfolios return and minimising
the variance of the portfolios return. In this section, we investigate a mixed asset
allocation strategy on two assets whose dynamics are modulated by HOHMM. With
this strategy investors determine the optimal weight of each asset to allocate based
on the estimated parameters and the state of the higher-order Markov chain. The
development here follows the applications of results in [15, 16]. Suppose an investor
is guided by an optimization equation MV, which is a linear combination of the
expected portfolios return and variance of the portfolios return. Let w = (wg , wv )
denote the weight of Russell 3000 growth and value indices, respectively. To solve
this mean-variance problem, we wish to estimate the optimal w which maximises the
function
RG
RV
RG
RV
MV(w) = v E[wg yk+1
+ wv yk+1
|Yk ] Var[wg yk+1
+ wv yk+1
|Yk ],

where y RG and y RV are the returns of Russell 3000 growth and value indices, and v
is a nonnegative risk aversion factor. The optimal weights are given in the following
proposition.
Proposition 3 Let v > 0 be the risk aversion factor. Suppose that neither short selling
nor borrowing is allowed. The optimal weight wg is given by

wg =

v(f RG ,xk f RV ,xk )+2 RV ,xk 2

2( RG ,xk 2 + RV ,xk 2 )

1
0

when 2 RV , x k 2 < v(f RG , x k  f RV , x k )

< 2 RG , x k 2
,
when v(f RG , x k  f RV , x k ) > 2 RG , x k 2
when v(f RG , x k  f RV , x k ) < 2 RV , x k 2

and the optimal weight wv is given by wv = 1 wg .

Proof See [16] for proof.

J Math Model Algor

Fig. 3 Optimal weights for
Russell 3000 value and growth
indices in the HOHMM-based
mixed strategy with v = 0.08

0.56

wGrowth

0.54

wValue

Weight

0.52
0.5
0.48
0.46
0.44
Jul97

Jan00

Jul02

Jan05

Jul07

Jan10

Note that
xk , er  =

N


E[(xk , xk1 ), eri |Yk ], for r = 1 . . . N,

i=1

where eri = (er , ei ) indicating that indeed, the optimal weight depends on the state
of the embedded MC (xk , xk1 ).
Note that the weights belong to the interval [0, 1] since neither short selling nor
borrowing is allowed. Similar to the previous section, we divide the observation
data into 15 intervals. For each interval, the optimal weights are calculated for
each time k utilising the optimal parameters and the estimated states of the weak
Markov chain. Investors can allocate investment using different parameter updating
frequency depending on their goal. To achieve consistency in comparison with the
switching and pure index strategies, the weights employed for each index is updated
at the beginning of each interval; transaction cost will also be considered. To gauge
the strategy performance, we shall focus on the terminal value of the portfolio.
Figure 3 exhibits a plot of optimal weights for Russell 3000 growth and value
indices. The risk aversion factor v is a scaling constant which is chosen by the investor.
Here, we allow this factor to vary from v = 0 (totally avoiding risk) to v = 5 (seeking
some risk). The evolution of optimal weights for Russell 3000 growth index with
different values of v is shown in Fig. 4. When v is small, the investor is relatively
conservative. The switching of markets regime has less impact on his/her asset

Fig. 4 Evolution of optimal

weights for Russell 3000
growth index in the
HOHMM-based mixed
strategy with varying vs

v=0.08
v=1
v=2
v=3
v=5

0.9
0.8
0.7

0.5

growth

0.6

0.4
0.3
0.2
0.1
0
Jul97

Jan00

Jul02

Jan05

Jul07

Jan10

J Math Model Algor

Table 7 Performance comparison between HOHMM- and HMM-based mixed strategies with
varying transaction costs
Transaction
cost

5 bps
HOHMM HMM

Mean (X RG ) % 4.5606
10.664
Std (X RG ) %
Mean (X RV ) % 2.7680
Std (X RV ) %
8.5098

20 bps
HOHMM HMM

2.9287 3.5750
11.8866 10.3087
5.9108 3.7536
7.6077 8.9849

50 bps
HOHMM HMM

70 bps
HOHMM HMM

0.2564 0.2789
10.9647 9.4039
9.0960 7.0497
9.4298 10.7537

1.8686 1.5993
11.5399 9.7089
6.9710 5.7292
8.1914 10.0169

1.6767
10.6797
10.5163
10.2950

wealth

250
200
150
100

Jan96

Jan97

Jan98

Jan99

350

wealth

300
250
200
150
Jan00

Jan01

Jan02

Jan03

400
350
wealth

300
250
200
150
Jan04

Jan05

Jan06

Jan07

450
400
wealth

350
300
250
200
150
Jan08
pure growth

Jan09
pure value

Jan10
switching

mixed

Fig. 5 Switching, mixed, pure growth and pure value strategies comparison between 1995 and 2010

J Math Model Algor

allocation as can be viewed from the stable variation of weights for the Russell 3000
growth index. The investor with higher v appears to aggressively react to market
regime switching. The Russell 3000 growth index has higher risk than the value index
before September 2001 and it has lower risk after that. Consequently, the weight to
allocate in growth index is higher than 0.5 before this time and it drops below 0.5
when the index has less uncertainty.
Table 7 shows the overall performance of the mixed strategy, which is compared
with the pure growth and pure value strategies with v = 0.08, obtained though the
analogue formulae of Eqs. 20 and 21. The standard deviations of the differences of
returns, std(X RG ) and std(X RV ), are lower than that of using switching strategy as
we expected. Mean(X RG ) and Mean(X RV ) decrease as transaction cost increases.
Compared to the mixed strategy based on the forecasts under the usual HMM
framework, the HOHMM-based mixed strategy produces higher values in both mean
and standard deviation. The HOHMM setting certainly carries more opportunities
to explore the trade off between expected return and risk, which means higher risk
may lead to higher return.
Figure 5 presents the evolution of investment under the switching, mixed and pure
index strategies. Each subplot covers three years of data. We can see that based
on the value of the investments, the mixed strategy does not always outperform
other strategies. It is not straightforward to establish from the plots which strategy
is the best. We shall then evaluate the portfolio performance through some classical
measures in investment.

6 Evaluating the Portfolio Performance

In this section, we carry out performance comparisons amongst portfolio allocation
strategies developed in the previous sections using historical data and simulated data.
We begin by discussing and evaluating the simple performance of the portfolio. We
evaluate the portfolio performance through a benchmark. In this case, the Russell
3000 index is a natural benchmark since both the Russell 3000 growth and value
indices are its subindices. The comparison of four portfolios with the benchmark is
made using three classical measures on returns; these are the Sharpe ratio, Jensens
alpha and the appraisal ratio.
Table 8 demonstrates the performance of both the switching and the mixing
strategies as compared to the performance of the simple all in the value or all in

Table 8 Unconditional and conditional performance of both switching strategy and mixed strategy

Switching strategy
Switching invest in value index
Switching invest in growth index
Pure value index
Pure growth index
Mixed strategy
Weight of value index > 0.5
Weight of growth index > 0.5

Mean return
(%)

Std error of
mean (%)

Volatility
(%)

Std error of
volatility (%)

Weeks

8.80
10.40
7.80
5.70
5.20
6.80
5.70
25.50

4.90
5.90
7.80
5.00
5.70
5.00
5.20
9.80

19.04
14.49
24.01
17.96
17.02
19.38
19.86
9.66

1.08
0.65
2.31
1.01
1.08
0.98
1.02
0.65

784
375
409
784
784
784
735
49

J Math Model Algor

the growth subindices. These strategies either switch between the Russell value and
growth indices (whose properties are summarised in the middle two rows) or create
a mixed portfolio of the two units, blended between them, as described in this paper.
All quantities are given in annualised units. The unconditional average performance
of the switching strategy is given in row 1, the unconditional average performance of
the mixing strategy is given in row 6, and the unconditional average performance of
the constituent subindices are given in rows 4 and 5. All of the unconditional averages
are computed for 784 weeks. Conditional average performances are also presented,
in rows 2 and 3 for the switching strategy and in the final two rows (rows 7 and 8)
for the mixing strategy. The number of weeks during which the various conditions
presented applied are given in the rightmost column of the table.
From Table 8, ignoring the standard error measurements for the time being, both
the new strategies appear to yield a higher return than their simple counterpart. The
high standard errors of these measurements highlight the ever present difficulty of
estimating strategy returns. However, both of these strategies obtain these higher
returns at the cost of increased strategy risk from the 17 % range to the 19 %
range. The conditional performances summarised here explain this. In the switching
strategy, higher returns during the times at which the strategy recommends the
growth index come at a significantly increased cost in risk. In contrast, the mixed
strategy is able to obtain huge returns during the small fraction of the time that it
allocates more than half the wealth to the growth portfolio at low risk; this only
occurs because the riskiest days of the market were all experienced when the strategy
opted for a majority in the value index.
The mixed strategy results displayed in Table 7 appear counterintuitive, because
it appears that the usual risk-reduction effect of diversification are not in evidence.
To understand this, one has to recall that the growth and value index returns are very
strongly correlated to one another reducing the benefits of diversification in this case.
We now go on to evaluate a number of risk-adjusted portfolio measures. The first
measure is the Sharpe ratio, denoted by SR, and
E[Rportfolio Rriskfree ]
SR = 
,
Var(Rportfolio Rriskfree )
where Rriskfree is the risk-free interest rate. This is used to characterise how well the
return of an asset compensates the investors for the risk taken. The higher the Sharpe
ratio the higher is the return with the same level of risk. In Table 9, we tabulate the
Sharpe ratio of five investment strategies using the dataset divided into 15 intervals.
Note that the switching strategy has the same Sharpe ratio with that of one of either
pure value or pure growth strategy. At the beginning of each interval, the switching
strategy allocates to one of the subindices with a number of shares depending on
the value of the switching investment and the chosen index at previous time step.
Hence, the switching portfolio and the chosen subindex have the same return in one
interval. Such similarity is also true in other measures since all the calculations are
based on the returns. The differences amongst the strategies are small. Out of the 15
intervals, the switching strategy shows a better performance than the benchmark in
11 intervals and the mixed strategy outperforms the benchmark in six intervals. Both
the HOHMM switching and HOHMM mixed strategies have higher risk-adjusted
mean than the benchmark. In particular, the switching strategy shows the highest
risk-adjusted mean in all of the strategies.

J Math Model Algor

Table 9 Sharpe ratio for five investment strategies using 15 intervals
Period

Switching
strategy

Mixed
strategy

Pure Russell
3000 value

Pure Russell
3000 growth

Pure Russell
3000 index

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

0.2164
0.2254
0.0083
0.0999
0.0141
0.0751
0.1488
0.1252
0.0206
0.0701
0.0976
0.0839
0.1293
0.0259
0.6620

0.2016
0.2178
0.0298
0.1485
0.0095
0.1347
0.1930
0.1304
0.0393
0.0630
0.0701
0.1471
0.1294
0.0243
0.5511

0.1645
0.2254
0.0523
0.0999
0.0141
0.0751
0.1488
0.1252
0.0206
0.0555
0.0976
0.1904
0.1293
0.0259
0.4190

0.2164
0.2020
0.0083
0.1699
0.0058
0.1652
0.2394
0.1331
0.1104
0.0701
0.0346
0.0839
0.1264
0.0219
0.6620

0.2020
0.2177
0.0292
0.1491
0.0184
0.1360
0.1984
0.1311
0.0475
0.0633
0.0653
0.1405
0.1292
0.0241
0.5581

Mean

0.0741
(4.96 104 )

0.0495
(4.74 104 )

0.0435
(3.92 104 )

0.0521
(5.42 104 )

(4.7 104 )

Std

0.1976
(5.99 104 )

0.1890
(4.46 104 )

0.1577
(3.10 104 )

0.2168
(5.82 104 )

(4.50 104 )

Mean/Std

0.3751
(2.24 103 )

0.2622
(2.55 103 )

0.2755
(2.66 103 )

0.2405
(2.53 103 )

0.2618
(2.50 103 )

0.0499

0.1906

Numbers inside the parentheses are standard errors

We calculate Jensens alpha, which is often used to measure the abnormal return
of a portfolio over the expected return. This is denoted by J and it is the constant in
the regression model,
J = Rportfolio [Rriskfree portfolio (Rbenchmark Rriskfree )].
A positive alpha indicates the portfolio has a higher marginal return. Table 10
shows the Jensens alpha for four allocation strategies. Although the differences
amongst the values of are very small, there are 11 and five positive s out of 15 for
the switching and mixed strategies, respectively. It indicates the marginal returns in
these periods are higher than that of the benchmark.
Finally, we consider the Treynor and Blacks appraisal ratio (AR), also known
as the information ratio. It is defined as the ratio between relative return and the
relative risk and is given by
E[Rportfolio Rbenchmark ]
AR = 
.
Var(Rportfolio Rbenchmark )
The formula is very similar to the Sharpe ratio. Whereas the Sharpe ratio measures
return relative to a riskless asset, the AR looks at returns relative to a risky
benchmark. The higher the AR, the higher is the active return of the portfolio given
the same risk level. Table 11 reports the AR of four investment strategies. The
switching strategy outperforms the mixed strategy in 11 intervals. In particular, we
have the highest mean and lowest standard deviation under this measure. We observe

J Math Model Algor

Table 10 Jensens alpha for
four investment strategies
using 15 intervals

Period

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

Switching
strategy
(104 )
1.2306
3.6729
4.0153
5.6901
6.1203
17.7209
12.7347
0.1632
11.3998
1.8036
6.3168
13.8661
2.1714
0.4850
18.0944

Mixed
strategy
(104 )
0.0417
0.0300
0.1148
0.0249
2.5431
0.8786
1.5509
0.0380
1.4556
0.1018
0.9435
1.7868
0.2851
0.0225
0.9831

Pure Russell Pure Russell

3000 value
3000 growth
(104 )
(104 )
1.230
3.7262
4.0153
5.0587
10.2510
21.5255
14.0599
0.1686
12.270
1.8036
6.3031
13.8661
1.9350
0.5308
18.0944

5.8977
0.0722
1.4289
(0.0183)
(0.0028)
(0.0261)
Std
7.3348
1.0969
10.3850
(0.0099)
(0.0022)
(0.0182)
Mean/Std 8040.7842 658.2883 1375.9094
(26.1277)
(28.2625)
(28.1958)
Mean

Numbers inside the

parentheses are standard
errors

1.3637 6
3.6729
4.3166
5.6901
6.1203
17.7209
12.7347
0.1632
11.3998 7
1.5748
6.3168
14.7742
2.1714
0.4850
18.4459
0.5987
(0.0245)
9.6750
(0.0169)
618.8129
(28.4250)

Table 11 Appraisal ratio (AR) for four investment strategies using 15 intervals
Period

Switching
strategy

Mixed
strategy

Pure Russell
3000 value

Pure Russell
3000 growth

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

0.1539
0.0239
0.0863
0.1729
0.0130
0.1810
0.2562
0.0528
0.3399
0.0251
0.1213
0.3355
0.0718
0.0260
0.3435

0.1761
0.0252
0.1354
0.0946
0.2037
0.1199
0.1838
0.0547
0.3356
0.0042
0.1122
0.3366
0.0686
0.0307
0.3505

0.1571
0.0239
0.0921
0.1729
0.0130
0.1810
0.2562
0.0528
0.3399
0.0206
0.1213
0.3338
0.0718
0.0260
0.3388

0.1539
0.0233
0.0863
0.1781
0.0227
0.1746
0.2658
0.0529
0.3478
0.0251
0.1222
0.3355
0.0703
0.0250
0.3435

Mean

0.1023
(4.03 104 )
0.1633
(2.14 104 )
0.6263
(27.1 104 )

0.0445
(4.66 104 )
0.1869
(3.19 104 )
0.2381
(29.41 104 )

0.0235
(4.88 104 )
0.1926
(3.17 104 )
0.1220
(29.01 104 )

0.0207
(4.90 104 )
0.1953
(3.27 104 )
0.1060
(28.63 104 )

Std
Mean/Std

Numbers inside the parentheses are standard errors

J Math Model Algor

Table 12 p-values for the Jarque-Bera test of normality on data given in Tables 911

Sharpe ratio
Jensens alpha
Appraisal ratio

Switching
strategy

Mixed
strategy

Pure Russell
value

Pure Russell
growth

Pure Russell
index

0.0026
0.4092
0.4929

0.0350
0.3608
0.5000

0.3136
0.5000
0.5000

0.0111
0.5000
0.5000

0.0322

higher mean for switching strategy in each performance measure. A t-test is carried
out to assess whether the means of portfolio under various performance measures
are statistically different. In order to run a t-test, each of the datasets (i.e., columns of
observations in Tables 9 through 11) being compared must be checked for normality.
Table 12 presents the results of Jarque-Bera normality tests for all three portfolio
measures applied to the five different portfolio selection approaches. The p-values
for Jensens alpha and Appraisal ratio of all portfolios are high which suggests there is
no sufficient evidence to indicate that these data sets are coming from a non-normal
distribution. Moreover, at 0.05 significance level, we can reject the null hypothesis
that under the Sharpe ratio, the data on switching, mixed, pure growth and pure
index strategies are from normal distribution. We test the difference between each
pair of portfolios for the two measures. It has to be noted that the measures or criteria
for comparison make use of the same experiment data set. The inherent problem
with multiple comparison is the increase in type I error (i.e., probability of falsely
rejecting the null hypothesis of no significance) that occurs when statistical tests are
used repeatedly. To control this familywise error rate, which is the probability of
making one or more false discoveries associated with multiple hypotheses tests, p
values are adjusted. The adjustment employed the Dunn-Sidk
procedure, which is
less conservative and more powerful than the Bonferroni method. Table 13 shows

the Dunn-Sidk-corrected
p-values for a one-tailed paired t-tests of significance
assuming unequal variances. Comparing the switching and mixed strategies, the pvalues in the first column are very small. This tells us that the difference in means
under these performance measures of these two strategies is highly significant. The
same can be said for the comparison of switching and pure growth strategies under
the Jensens alpha criterion. We recall that the switching strategy has the best performance for the period considered. For the switching versus pure value strategies,
the p-values are no longer small so that we cannot reject the null hypothesis, i.e., we
cannot reject that the two means are equal. Similar conclusion can be made when we
compare the mixed and pure value strategies as well as the mixed and pure growth
strategies where the p-values are extremely large. In addition to the t-test, we use
the Wilcoxon rank sum test in assessing the significance of the differences. The p
values, also corrected based on Dunn-Sidks
method for multiple comparison, are

Table 13 Dunn-Sidk-adjusted
p-values for a one-tailed significance test on the performance results
shown in Tables 1011

Jensens alpha
Appraisal ratio

Switching vs
Mixed

Switching vs
Pure growth

Switching vs
Pure value

Mixed vs
Pure growth

Mixed vs
Pure value

0.0070
0.0296

0.0345
0.2127

0.1004
0.0632

0.5257
0.3261

0.6360
0.6182

J Math Model Algor

Table 14 Dunn-Sidk-adjusted
p-values for a Wilcoxon rank sum test on the performance results
shown in Tables 911

Sharpe ratio
Jensens alpha
Appraisal ratio

Switching vs
Mixed

Switching vs
Pure growth

Switching vs
Pure value

Mixed vs
Pure growth

Mixed vs
Pure value

0.9712
0.0475
0.0997

0.9712
0.1821
0.6419

0.9977
0.2840
0.2065

1.0000
0.9446
0.6567

1.0000
0.9667
0.9667

reported in Table 14. Wilcoxon test does not rely on the normality assumption and so
it complements our use of the t-test. For switching strategy versus mixed strategies,
the result suggests that the Jensens alpha are significantly different but not for the
appraisal ratio. Note that it may appear counter intuitive that the Jensens alpha and
the appraisal ratios are actually better for the switching than for the mixing strategy.
Keep in mind, however, that we did not optimise the strategies for either ratio but
rather for raw return; these tests were done after the fact to compare return-optimal
strategies. Next, we give a simulation analysis in conjunction with the three portfolio
measures.
We are interested in the statistical inference of the above portfolio measures for
each portfolio strategy. The bootstrap is a way of finding sampling distribution from
one sample path. Introduced by Efron and Tibshirani [12], it is a technique allowing
estimation of the sample distribution of almost any statistics. This method can be
implemented when the sample could be assumed to be drawn from an independent

Table 15 Performance evaluation for 10,000 bootstrapped datasets with 5 bps transaction cost
Sharpe ratio
Switching
Mixed
Pure value
Pure growth
Pure index

Mean (102 )
3.1617
2.8455
2.8137
1.7502
2.2978

Std (102 )
0.1711
0.1321
0.1430
0.0829
0.1056

95 % Con.Int. (102 )
[3.1498 3.1737]
[2.8363 2.8548]
[2.8038 2.8238]
[1.7445 1.7561]
[2.2904 2.3052]

Jensens
Switching
Mixed
Pure Value
Pure growth

Mean (103 )
0.4037
0.2259
0.2673
0.0430

Std (103 )
0.0607
0.0483
0.0370
0.0506

95 % Con.Int. (103 )
[0.3995 0.4080]
[0.2225 0.2293]
[0.2647 0.2699]
[0.0465 0.0395]

AR
Switching
Mixed
Pure value
Pure growth

Mean (103 )
1.4270
2.3559
7.0522
6.2734

Std (103 )
0.4287
0.3149
0.5095
0.3956

95 % Con.Int. (103 )
[1.3970 1.4570]
[2.3779 2.3338]
[7.0879 7.0165]
[6.3010 6.2457]

Mean & Std return

Switching
Mixed
Pure value
Pure growth
Pure index

Mean (103 )
0.9250
0.8682
0.7770
0.7829
0.8967

Std (103 )
0.0429
0.0244
0.0065
0.0074
0.0051

95 % Con.Int. (103 )
[0.9220 0.9280]
[0.8665 0.8699]
[0.7765 0.7774]
[0.7824 0.7834]
[0.8964 0.8971]

J Math Model Algor

and identically distributed population. Bootstrap method constructs a number of

resamples of the observation datasets with equal size by random sampling with
replacement from the original dataset. The datasets used for the bootstrapping are
the Russell 3000 index, the growth subindex and the value subindex. Each of the
original datasets contains 784 data points. The procedure of the simulation is as
follows:
1. We divide the datasets into 16 intervals and each interval contains 49 weeks.
2. A resample is created by repeatedly sampling with replacement from these 16
intervals. This means that we randomly pick one interval for the resample path
and put the interval back for drawing again. As a result, any interval can be drawn
more than once, or not at all. The resample has the same size as the original data.
3. The three measures are calculated for the new resample path.
The construction is repeated 10,000 times and the statistics of the three classical
measures are obtained. Table 15 shows the statistics of the portfolios, Sharpe ratio,
Jensens alpha and AR with a transaction cost amounting to 5 bps. Table 16 presents
the same analysis with transaction cost of 30 bps. When 5 bps transaction costs
are introduced, the HOHMM switching and mixed strategies generate higher mean
Sharpe ratio and AR from the 10,000 bootstrap sample paths than those from other
strategies. Only the pure growth strategy leads to a negative mean in the Jensens
alpha measure. The standard deviation of switching strategy from the bootstrapped
samples is higher than that in the mixed strategy in all cases. The estimated 95 %

Table 16 Performance evaluation for 10,000 bootstrapped datasets with 30 bps transaction cost
Sharpe ratio
Switching
Mixed
Pure value
Pure growth
Pure index

Mean (102 )
3.7352
2.4876
2.8781
1.5679
2.1326

Std (102 )
0.3169
0.1220
0.1380
0.1173
0.1176

95 % Con.Int. (102 )
[3.7130 3.7574]
[2.4791 2.4962]
[2.8684 2.8877]
[1.5597 1.5761]
[2.1244 2.1408]

Jensens
Switching
Mixed
Pure value
Pure growth

Mean (103 )
5.0273
1.6992
2.6856
0.2362

Std (103 )
0.7980
0.5069
0.4149
0.5838

95 % Con.Int. (103 )
[4.9714 5.0832]
[1.6637 1.7347]
[2.6565 2.7146]
[0.2770 0.1953]

AR
Switching
Mixed
Pure value
Pure growth

Mean (103 )
7.5903
1.3412
3.4708
5.8516

Std (103 )
0.8554
0.3479
0.2721
0.5834

95 % Con.Int. (103 )
[7.5304 7.6501]
[1.3169 1.3656]
[3.4517 3.4898]
[5.8925 5.8108]

Mean & Std return

Switching
Mixed
Pure value
Pure growth
Pure index

Mean (103 )
1.0447
0.7382
0.7669
0.6304
0.7236

Std (103 )
0.0673
0.0193
0.0058
0.0108
0.0070

95 % Con.Int. (103 )
[1.0400 1.0494]
[0.7369 0.7396]
[0.7665 0.7674]
[0.6297 0.6312]
[0.7231 0.7241]

J Math Model Algor

confidence intervals for the mixed strategy are smaller than those in the switching
strategy. Apparently, the HOHMM mixed strategy is more stable than the HOHMM
switching strategy in terms of the standard deviation and 95 % confidence interval
under the 5 bps transaction cost. A comparison of the mean and variance of
the portfolio returns is also presented. The switching strategy outperforms other
strategies with the highest variance nonetheless. On the other hand, the pure Russell
3000 index strategy has the lowest variance, but the 95 % confidence interval is
bigger than those in both HOHMM strategies. It indicates that both HOHMM
strategies are more stable than the benchmark in terms of the confidence interval.
When transaction costs are set to 30 bps, the switching strategy produces the highest
mean in all cases. Since the mixed strategy is the most costly strategy, it has a lower
mean than pure value strategy under all measures. However, it still has positive
means in both Jensens alpha and AR measures. Both HOHMM-based strategies
outperform the benchmark with 30 bps transaction cost in terms of higher Sharpe
ratio, positive Jensens alpha and positive AR. The switching strategy is the most
risky judging the standard deviation of the measures. The mixed strategy shows
a lower variance than the pure sub-indices strategies in Sharpe ratio and Jensens
alpha, and a lower variance than the pure growth strategy in the AR. Furthermore,
the smaller 95 % confidence interval of mixed strategy indicates that it is more stable
than the switching strategy in the case of 30 bps transaction cost. Therefore, we could
conclude that the HOHMM switching strategy gives a higher mean return, however
the HOHMM mixed strategy is less risky and more stable.

7 Conclusion
This paper examined asset allocation strategies for growth and value stocks under
a weak hidden Markovian regime-switching setting. We suppose that the mean and
volatilities of the price indices returns are modulated by a discrete-time multivariate
HOHMM process. Recursive optimal estimates by filtering multidimensional observations are given for the state and various processes related to the underlying
second-order Markov chain. The parameters of the model, including the transition
probabilities, the drift and the variance parameters in the multidimensional observations, can be re-estimated and the forecasts can be obtained using the estimates.
We investigated two investment strategies: a switching strategy and a mixed strategy,
using the weekly Russell 3000 growth and value indices data from 1995 to 2010. The
switching strategy made use of the one-step ahead forecasted return for both indices
and invested into the index with higher risk-adjusted forecasted return for each time
interval. The mixed strategy lead to a mean-variance optimisation problem, in which
the optimal weights for each index were calculated using the estimated drifts and
variance.
We compared both HOHMM strategies with the HMM-based approach. For
certain levels of transaction costs, the HOHMM-based strategies outperform the
HMM-based strategies in terms of the higher differences of log-return between the
tested strategy and the pure strategies. The HOHMM switching strategy never gives
the worst performance in the time interval considered. The evolution of the optimal
weights represents the investors reaction to regime-switching in the market. And
thus the mixed strategy has a lower variance of the return. Performance comparisons

J Math Model Algor

of the four portfolio strategies with the benchmark using Sharpe ratio, Jensens
alpha and AR were presented. When compared to the benchmark, which is the
pure Russell 3000 index strategy, both HOHMM strategies have higher risk-adjusted
return. The switching strategy has higher marginal and relative returns than the
benchmark. Furthermore, the bootstrap analysis with different transaction costs
demonstrates that with 5 bps transaction cost, both HOHMM-based strategies have
higher return and are more stable than the benchmark in terms of higher values in
the performance measures and smaller confidence intervals. In the case of 30 bps
transaction costs, the HOHMM strategies still have higher returns, but the switching
strategy is less stable and the mixed strategy is more stable than the benchmark.

References
1. Abell, M.L., Braslton, J.P.: Introductory Differential Equations: With Boundary Value Equations. Academic Press, California (2009)
2. Ammann, M., Verhofen, M.: The effect of market regimes on style allocation. Financ. Markets
Portf. Mgmt 20, 309337 (2006)
3. Ang, A., Bekaert, G.: International asset allocation with regime shifts. Rev. Financ. Stud. 15,
11371187 (2002)
4. Ang, A., Bekaert, G.: Timing and diversification: a state-dependent asset allocation approach.
Financ. Anal. J. 60, 11371187 (2004)
5. Bauer, R., Haerden, R., Molenaar, R.: Timing and diversification: a state-dependent asset allocation approach. J. Invest. 13, 7280 (2004)
6. Bulla, J., Bulla, I.: Stylised facts of financial time series and hidden semi-Markov models. Comput. Stat. Data Anal. 51, 21922209 (2006)
7. Bulla, J., Mergner, S., Bulla, I., Sesbo, A., Chesneau, C.: Markov switching asset allocation: do
profitable strategies exist? Journal of Asset Management 12, 310321 (2011)
8. Ching, W., Ng, M.K., Fung, E.S.: Higher-order multivariate Markov chains and their applications.
Linear Algebra Appl. 428, 492507 (2008)
9. Ching, W.K., Siu, T.K., Li, L.M.: Pricing exotic options under a higher-order hidden Markov
model. J. Appl. Math. Decis. Sci. 2007, 115 (2007). Article ID 18014
10. Couillard, M., Davison, M.: A comment on measuring the Hurst exponent in financial time series.
Physica. A, Statistical Mechanics and Its Applications 348, 404418 (2005)
11. Dajcman, S.: Time-varying long-range dependence in stock market returns and financial market
disruptionsa case of eight European countries. Appl. Econ. Lett. 19, 953957 (2012)
12. Efron, B., Tibshirani, R.J.: An Introduction to the Bootstrap. CRC Press, Boca Raton, FL (1994)
13. Elliott, R.J., Aggoun, L., Moore, J.B.: Hidden Markov Models: Estimation and Control. Springer,
New York (1995)
14. Elliott, R.J., Siu, T.K., Badescu, A.: On mean-variance portfolio selection under a hidden
Markovian regime-switching model. Econ. Model. 27, 678686 (2010)
15. Elliott, R.J., vander Hoek, J.: An application of hidden markov models to asset allocation
problems. Finance Stoch. 1, 229238 (1997)
16. Erlwein, C., Mamon, R., Davison, M.: An examination of HMM-based investment strategies for
asset allocation. Appl. Stoch. Models Bus. Ind. 27, 204221 (2011)
17. Graflund, A., Nilsson, B.: Dynamic portfolio selection: the relevance of switching regimes and
investment horizon. Eur. Financ. Manag. 9, 17900 (2003)
18. Guidolin, M., Timmermann, A.: Asset allocation under multivariate regime switching. J. Econ.
Dyn. Control 31, 35033544 (2007)
19. Hess, M.K.: Timing and diversification: a state-dependent asset allocation approach. Eur. J.
Financ. 12, 189204 (2006)
20. Hurst, H.E.: Long-term storage capacity of reservoirs. Trans. Am. Soc. Civ. Eng. 116, 770770
(1951)
21. Lobato, I.N., Savin, N.E.: Real and spurious long memory in stock market data. J. Bus. Econ.
Stat. 16, 261266 (1988)
22. MacDonald, I.L., Zucchini, W.: Hidden Markov Model and other Models for Discrete-Valued
Time Series. Chapman & Hall, London (1997)

J Math Model Algor

23. Maheu, J.: Can GARCH models capture long-range dependence? Stud. Nonlinear Dyn.
Econom. 9, 1 (2005)
24. Markowitz, H.M.: Portfolio selection. J. Finance 7, 7791 (1952)
25. McCarthy, J., DiSario, R., Saraoglu, H., Li, H.C.: Tests of long-range dependence in interest rates
using wavelets. Q. Rev. Econ. Finance 44, 180189 (2004)
26. Merton, R.C.: Optimum consumption and portfolio rules in a continuous-time model. J. Econ.
Theory 3, 373413 (1971)
27. Du Preez, J.A., Barnard, E., Weber, D.M.: Efficient high-order hidden Markov modeling. In:
Proceedings of the International Conference on Spoken Language Processing, pp. 29112914
(1998)
28. Ray, B., Tsay, R.: Long range dependence in daily stock volatilities. J. Bus. Econ. Stat. 18, 254
262 (2000)

29. Rydn, T., Tersvirta, T., Asbrink,

S.: Stylised facts of daily return series and the hidden Markov
model. J. Appl. Econ. 13, 217244 (1998)
30. Samuelson, P.: Lifetime portfolio selection by dynamic stochastic programming. Rev. Econ. Stat.
51, 23946 (1969)
31. Siu, T.K., Ching, W.K., Fung, E., Ng, M., Li, X.: A high-order Markov-switching model for risk
measurement. Comput. Math. Appl. 58, 110 (2009)
32. Solberg, J.: Modelling Random Processes for Engineers and Managers. Wiley & Sons, Inc.,
New Jersey (2009)
33. Xi, X., Mamon, R.: Parameter estimation of an asset price model driven by a weak hidden
Markov chain. Econ. Model. 28, 3646 (2011)
34. Yu, S., Liu, Z., Squillante, M.S., Xia, C.H., Zhang, L.: A hidden semi-Markov model for web
workload self-similarity. In: 21st IEEE International Performance, Computing, and Communications Conference, pp. 6572 (2002)