Página 14 - ANAlitica6

Elsy Gómez-Ramos y Francisco Venegas-Martínez
Analíti ak
6
Revista de Análisis Estadístico
Journal of Statistical Analysis
Tabla 1. Application of different networks for the stock market and the exchange rate. Source: author elaboration.
Year Author(s)
ANNs
Resultsa
1 2 3 4 5
MLP RBF Others RCNs PLNs MDNs SVM
Exchange rate
2011 Dhamija & Bhalla × × RBF performs better
2010 Bildirici et al. × × × RBF performs better
2009 Ghazali et al. × × PLNs & MLP perform better
2008 Kiani & Kastens × × RCNs perform better
2008 Hussain et al. × × Same performance
2007 Ghazali et al. × × PLNs performs better
2007 Portela et al. × × MLP performs better
2006 Ince & Trafalis × × SVM performs better
2002 Kodogiannis & Lolis × × × RCNs perform better
2001 Zhang & Berardi × × MDNs perform better
2000 Leung et al. × × Others performs better
1995 Kuan & Liu × × Small differences
1994 Pham & Liu × × × MLP & others perform better
Stock market
2012 Chang et al. × × Others perform better
2011 Shen et al. × × × RBF performs better
2011 Lu & Wu × × × × Others perform better
2011 Kara et al. × × MLP performs better
2011 Carpinteiro et al. × × SVM performs better
2011 Guresen et al. × × MLP performs better
2010 Mostafa × × Others performs better
2005 Enke & Thawornwong × × MLP performs better
2006 Selvaratnam & Kirley × × Same performance
2005 Yumlu et al. × × RCNs perform better
2005 Pérez et al . × × MLP performs better
2004 Thawornwong & Enke × × MLP performs better
2002 Sitte × × Same performance
2000 Leung et al. × × Others perform better
Both
2011 Adeodato et al. × × MDNs perform better
2011 Ghazali et al. × × PLNs performs better
1994 Hutchinson, Lo, & Poggio × × MLP & RBF perform better
aThe results are based on the author’s criterion for the multilag forecast
The above results suggest that, in general, the basic
functions, the initialization weights, and the network ar-
chitecture will produce a path that concentrates great ef-
forts. These issues have already been established by other
authors who seek through intelligent methods (e.g. gene-
tic algorithms) to improve the performance of the MLP,
especially with the network architecture and initialization
weights (Hansen and Nelson, 2003; Karathanasopoulos et
al. 2010). With respect to the basic function, the literature
in finance does not discuss this task with respect to GA or
another methodology (or we did not find it). As a conclu-
sion, we should not suggest complicated networks, in some
cases, the simplicity is the best.
6 The IBM case
In this section we apply the MLP to the IBM case. To do
this, we will focus on the IBM daily common stock returns
from January 02, 2003 to May 05, 2013 (see Figure 2). Da-
ta consisting of 2592 days will be used for training and the
last 10 days for testing (or forecast period). The software to
be used is Mathematica 6.0.
12 Analítika, Revista de análisis estadístico , 3 (2013), Vol. 6(2): 7-15