Expert Systems with Applications 38 (2011) 15009–15013

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

An approach to portfolio selection using an ARX predictor for securities’ risk

and return
D.D.D. Pinto a, J.G.M.S. Monteiro a,⇑, E.H. Nakao b
a
Federal University of Rio de Janeiro, Chem. Eng. Dep., Ilha do Fundão, s/n, Rio de Janeiro, Brazil
b
IRB Brasil Resseguros, Avenida Marechal Câmara, 171, Rio de Janeiro, Brazil

a r t i c l e i n f o a b s t r a c t

Keywords: It is well known that every investment carries a risk associated, and depending on the type of investment,
ARX it can be very risky; for instance, securities. However, Markowitz proposed a methodology to minimize
Markowitz the risk of a portfolio through securities diversification. The selection of the securities is a choice of
Portfolio selection the investor, who counts with several technical analyzes to estimate investment’s returns and risks. This
Risk and return
paper presents an autoregressive exogenous (ARX) predictor model to provide the risk and return of some
Brazilian securities – negotiated at the Brazilian stock market, BOVESPA – to select the best portfolio,
herein understood as the one with minimum expected risk. The ARX predictor succeeded in predicting
expected returns and risks of the securities, which resulted in an effective portfolio. Additionally the
Markowitz theory was confirmed, showing that diversification reduces the risk of a portfolio.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction securities’ monthly closing data were the output. With the data
available, the ARX predictor was trained, predicting the securities’
Portfolio selection is a critical key for investors’ return. As Sper- closing, in other words, the expected return of the securities. The
anza (1996) stated, portfolio optimization deals with the tradeoff confidence interval was used as a Value-at-Risk (VaR) measure-
between rate of return and risk. Markowitz (1952) presented a ment. Hence, the ARX predictor was able to provide every variable
work revealing that the portfolio risk decreases as it diversifies, necessary to use Markowitz mean-variance model and select the
which, in 1990, awarded him the Nobel Memorial Prize of Eco- best capital allocation for an effective portfolio.
nomic Sciences. His work incorporated, for the first time, the risk Two ARX models were used to estimate risk and return of secu-
and return relationship in a financial model and the concept of rities: (i) Single Input Multiple Output (SIMO), where outputs were
investor’s rational behavior was presented, starting the classical estimated simultaneously; and (ii) Single Input Single Output
theory of portfolio (Rambaud, Pérez, Granero, & Segovia, 2005). (SISO), where the outputs were estimated one by one. Portfolios
Portfolios with minimum risk for a given level of return, among with two to four securities were proposed.
all the possibilities, form the efficient frontier, which gives the best Continuing this introduction, Section 2 presents a brief intro-
investment alternatives (Fernández & Gómez, 2007). The accuracy duction to Markowitz’s work, while Section 3 describes the ARX
of the Markowitz model is strongly dependent on the expected model. In Section 4 the results of the securities risks and returns
return and risk accuracy (Ong, Huang, & Tzeng, 2005). Several risk prediction and the best portfolio selection are presented. Finally,
and return estimators are widely used, such as artificial neural net- Section 5 shows the conclusions and suggestions for further devel-
work, fuzzy logics, genetic algorithm and autoregressive and/or opment of the method herein presented.
moving average models, as reported, for example, in Bilbao-Terol,
Pérez-Gladish, and Antomil-Ibias (2006), Soleimani, Golmakani, 2. The Markowitz’s work
and Salimi (2009), Hlouskova, Schmidheiny, and Wagner (2009),
Yu, Wang, and Lai (2008) and Yoshida (2009). The publication, in Markowitz (1952) mean-variance model
In this paper, an autoregressive exogenous (ARX) predictor is was a milestone for capital allocation and portfolio selection. Based
used to provide the returns and risks of securities negotiated in in this model, several other theories and methodologies were cre-
the Brazilian Stocks and Exchange market, BOVESPA. BOVESPA ated to minimize risks and, therefore, select the best portfolio.
index (monthly closing, as well) was the model input, while Basically, Markowitz showed that the risk of a portfolio can be
minimized by diversification of the investment; in other words,
⇑ Corresponding author. Tel./fax: +55 21 2562 7535. adding securities to the portfolio. According to Xia, Liu, Wang,
E-mail address: [email protected] (J.G.M.S. Monteiro). and Lai (2000) ‘‘the core of the Markowitz mean-variance model is

0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.05.046
15010 D.D.D. Pinto et al. / Expert Systems with Applications 38 (2011) 15009–15013

to take the expected return of a portfolio as the investment return and Table 1
the variance of the expected returns of a portfolio as the investment Securities’ correlation factor.

risk’’. Hence, using statistic principles, it is possible to calculate PETR4 VALE5 EMBR3 TLPP4
the risk of the portfolio, as defined by Eq. (1). PETR4 1.00 0.77 0.11 0.35
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi VALE5 0.77 1.00 0.60 0.14
uX
u ns X ns EMBR3 0.11 0.60 1.00 0.69
rp ¼t ri  rj  qi  qj  qij ð1Þ TLPP4 0.35 0.14 0.69 1.00
i¼1 j¼1

where ns is the number of securities in the portfolio, qi is the The website www.bolsapt.com provided the data necessary for
percentage of capital invested on stock i, ri is the standard devia- the identification (the closing values of BOVESPA index and the se-
tion of stock i return and qij is the correlation factor of securities lected securities’ prices). The identified ARX(3,3) model used 24
i and j. points, starting in April 2007, to estimate the parameters and a
Minimizing Eq. (1), an investor can find a portfolio with mini- 90% confidence level to construct the confidence interval. As a con-
mum risk. However, the portfolio optimization is not simple, straint, the portfolio was set to contain a minimum of 10% of each
becoming more complicated as the number of securities increase. security.
Gondzio and Grothey (2007), Rustem, Becker, and Marty (2000), A VBA (Visual Basic for Application) code was used to provide
Xia et al. (2000) and Ong et al. (2005), for instance, demonstrate the calculation for the identification and estimation and Excel sol-
methods to optimize portfolio selection. ver was used to minimize Eq. (1). Table 1 shows the securities’ cor-
Another important concern, as reported by Xia et al. (2000), is relation factor, for the identification period.
the transaction costs. Yoshimoto (1996, apud Xia et al., 2000) sta-
ted that inefficient portfolios can be generated if transactions costs 4.1. Portfolios using the ARX SIMO predictor
are neglected. Considering the transaction costs, Sanvicente and
Bellato (2003) identify the number of securities necessary to suffi- Table 2 brings the securities’ estimations for portfolios contain-
ciently diversify a portfolio. ing two to four securities, using the ARX SIMO predictor. In this
type of predictor, one model estimates all the outputs.
In general, in Table 2, it is possible to see that the confidence
3. ARX predictor interval (VaR) enlarges as the number of securities added to the
predictor model increases. The results of the best portfolio are
Ljung and Glad (1994) defined a mathematical model as the shown at Table 3. For the portfolio containing two securities it is
relation between quantities that can be measured in a system. possible to visualize, graphically, a combination where the portfo-
These models can be employed both as simulation instruments lio’s risk is lower than the risks of the individuals securities. That is
and as forecasters (Ljung, 1987). shown at Fig. 1.
Identification is the procedure of constructing mathematical
models, by submitting previously available input and output data 4.2. Portfolios using ARX SISO predictor
to specific analysis (Haykin, 1999; Ljung, 1987). Erdogan and Gülal
(2009) define system identification as ‘‘a matter of finding the The ARX SISO predictor, contrary to the SIMO model, estimated
numerical values of the model parameters that provide the best agree- the outputs one by one; thereby, there was a necessity to identify
ment between the computed and observed system outputs’’.
Usually, ARX prediction models try to minimize a positive func- Table 2
tion of the prediction errors. The outputs are calculated consider- Risk and return estimation using ARX SIMO predictor.
ing that system changes unhurriedly during the time, by a set of
# of securities Security Value (BRL) VaR (%)
parameters that are estimated through system identification
(Huang & Jane, 2009). Estimated Real

Eq. (2) shows the form of an ARX model. 2 PETR4 29.95 34.45 20.47
VALE5 33.61 32.50 18.09
X
n X
m
3 PETR4 33.69 34.45 25.12
yðtÞ þ ak  yðt  kÞ ¼ eðtÞ þ bk  uðt  kÞ ð2Þ
VALE5 30.82 32.50 26.40
k¼1 k¼1
EMBR3 7.75 9.70 50.17
Where y(t) is the output, u(t) is the input and e(t) is the white 4 PETR4 39.44 34.45 19.17
noise term, m and n are the model order for input and output, VALE5 30.86 32.50 31.32
respectively, and ak and bk are the model parameters. EMBR3 8.90 9.70 53.86
TLPP4 53.94 45.59 14.50
The parameters were obtained maximizing the likelihood func-
tion, as explained in Ljung (1987).

4. Portfolio selection – case study Table 3

Portfolio selection using ARX SIMO predictor.

This section presents a case study in which four securities, # of securities Security % Invested Expected return (%) Risk (%)
negotiated at BOVESPA, were selected to comprise portfolios. 2 PETR4 23.64 1.39 17.81
Two predictors were used, namely ARX SIMO and ARX SISO, to se- VALE5 76.36 10.28
lect portfolios varying from two to four securities. 3 PETR4 75.64 14.05 23.48
For both models, BOVESPA index was the input and the closing VALE5 10.00 1.10
prices of securities Petrobras (PETR4), Vale (VALE5), Embraer EMBR3 14.36 12.38
(EMBR3) and Telesp (TLPP4) were the outputs. The portfolio with 4 PETR4 32.11 33.53 14.47
two securities was formed by PETR4 and VALE5. EMBR3 was added VALE5 10.00 1.24
EMBR3 10.00 0.59
to form the portfolio with three securities, and, finally, the portfolio
TLPP4 47.89 12.89
with four securities was formed by adding TLPP4.
D.D.D. Pinto et al. / Expert Systems with Applications 38 (2011) 15009–15013 15011

Fig. 1. Risk and return relation for portfolio with two securities.

Fig. 2. SISO model for PETR4.

Fig. 3. SISO model for VALE5.

15012 D.D.D. Pinto et al. / Expert Systems with Applications 38 (2011) 15009–15013

Fig. 4. SISO model for EMBR3.

Fig. 5. SISO model for TLPP4.

four ARX SISO models, one for each security. The identifications re- 4.3. Comparison
sults for all four securities can be found in Figs. 2–5.
It is possible to observe in Figs. 2–5 that the confidence interval From Tables 3 and 4 it is possible to see that the ARX SISO pre-
for all securities was tight enough, during the identification and dictor led to portfolios with lower risks; besides, from Table 3, it is
prediction periods, ensuring a good estimation. Moreover, the esti- seen that the portfolio with three securities has a risk higher than
mations were quite accurate. the portfolio with two securities, betraying Markowitz theory,
The best portfolio selections are found in Table 4. which stated that diversifying the investment the portfolio’s risk
must decrease. This fact occurred only for portfolios selected with
Table 4 ARX SIMO predictor, as Table 4 shows.
Portfolio selection using ARX SISO predictor.
ARX SIMO predictor was used for user’s convenience, since it al-
# of securities Security % Invested Expected return (%) Risk (%) lows all the outputs to be calculated by estimating only one model.
2 PETR4 69.80 14.00 13.07 However, as the number of securities increases, the number of
VALE5 30.20 3.47 parameters also increases. It is possible to see in Table 5 that the
3 PETR4 75.52 14.00 12.55 number of parameters necessary to estimate in a SIMO model
VALE5 10.00 3.47 (n = m = 3 and nu = 1) is always higher than SISO ones. For instance,
EMBR3 14.48 18.10 for 4 outputs, SIMO models would have to estimate 60 parameters,
4 PETR4 17.71 14.00 7.15 while SISO models would have to estimate just 24–6 for each mod-
VALE5 10.00 3.47 el. This excess of parameters can lead to an error propagation
EMBR3 10.00 18.10
which could be one of the reasons why the SIMO model generates
TLPP4 62.29 0.09
worst estimations than the SISO one.
 D.D.D. Pinto et al. / Expert Systems with Applications 38 (2011) 15009–15013 15013

Table 5 Erdogan, H., & Gülal, E. (2009). Identification of dynamic systems using multiple
Parameters required for SIMO and SISO models. input–single output (MISO) models. Nonlinear Analysis: Real World Applications,
10, 1183–1196.
# of outputs # of parameters Fernández, A., & Gómez, S. (2007). Portfolio selection using neural networks.
Computers and Operations Research, 34, 117–1191.
SIMO SISO
Gondzio, J., & Grothey, A. (2007). Solving non-linear portfolio optimization
1 – 6 problems with the primal-dual interior point method. European Journal of
2 18 12 Operational Research, 181, 1019–1029.
3 36 18 Haykin, S. (1999). Neural networks: A comprehensive foundation. New Jersey:
4 60 24 Prentice Hall.
5 90 30 Hlouskova, J., Schmidheiny, K., & Wagner, M. (2009). Multistep predictions for
multivariate garch models: Closed form solution and the value for portfolio
management. Journal of Empirical Finance, 16, 330–336.
Huang, K.-Y., & Jane, C.-J. (2009). A hybrid model for stock market forecasting and
portfolio selection based on ARX, grey system and RS theories. Expert Systems
Another interesting comparison is that the securities’ risks esti- with Applications, 36, 5387–5392.
mated with the ARX SISO predictor are lower than those selected Ljung, L. (1987). System identification: Theory for the user. Prentice Hall.
by the ARX SIMO predictor. Ljung, L., & Glad, T. (1994). Modeling of dynamic systems. Prentice Hall.
Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7(1), 77–91.
Ong, C. S., Huang, J. J., & Tzeng, G. H. (2005). A novel hybrid model for portfolio
selection. Applied Mathematics and Computation, 169, 1195–1210.
5. Conclusions Rambaud, S. C., Pérez, J. G., Granero, M. A. S., & Segovia, J. E. T. (2005). Theory of
portfolios: New considerations on classic models and the capital market line.
European Journal of Operational Research, 163, 276–283.
The results of the SISO models were more accurate than the Rustem, B., Becker, R. G., & Marty, W. (2000). Robust min–max portfolio strategies
SIMO ones; furthermore, the portfolios’ risk using SISO models for rival forecast and risk scenarios. Journal of Economic Dynamics and Control,
estimations decreases as the number of securities increases, con- 24, 1591–1621.
Sanvicente, A. Z., & Bellato, L. L. N. (2003). Determinação do grau necessário de
firming Markowitz theory. The ARX SISO predictors proved to be diversificação de uma carteira de ações no mercado de capitais brasileiro. São
a useful risk and return estimators. Nonetheless, Excel solver Paulo: FAPESP.
showed not to be a good optimization tool. Some other optimiza- Soleimani, H., Golmakani, H. R., & Salimi, M. H. (2009). Markowitz-based portfolio
selection with minimum transaction lots, cardinality constraints and regarding
tion techniques, or even alternative objective functions, should sector capitalization using genetic algorithm. Expert Systems with Applications,
be tested having the ARX SISO models as the return and risk 36, 5058–5063.
estimator. Speranza, M. G. (1996). A heuristic algorithm for a portfolio optimization model
applied to the milan stock market. Computers and Operations Research, 23(5),
433–441.
Xia, Y., Liu, B., Wang, S., & Lai, K. K. (2000). A model for portfolio selection with order
References of expected returns. Computers and Operations Research, 27, 409–422.
Yoshida, Y. (2009). An estimation model of value-at-risk portfolio under
Bilbao-Terol, A., Pérez-Gladish, B., & Antomil-Ibias, J. (2006). Selecting the optimum uncertainty. Fuzzy Sets and Systems, 160, 3250–3262.
portfolio using fuzzy compromise programming and sharpe’s single-index Yu, L., Wang, S., & Lai, K. K. (2008). Neural network-based mean-variance-skewness
model. Applied Mathematics and Computation, 182, 644–664. model for portfolio selection. Computers and Operations Research, 35, 34–46.