Modeling Early Warning System Using Combination Of Logit Model And Nearest Neighbour Tree For Predicting Currency Crisis
Loading...
Date
2014-11
Authors
Ramli, Nor Azuana
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
A number of studies and researches were conducted in finding a suitable method
for early warning systems in order to acquire accurate prediction results. These are not
limited to generate forecast for the currency crisis, but in other applications as well.
Varieties of methodologies from different fields have been introduced in modeling the
currency crisis but the main problem highlighted in this thesis is some of these methods
are actually contributing to more false alarms than tangible predictions. Due to poor
prediction results, the economists themselves questioned the possibilities of financial
crisis. Hence, one of the objectives in this study is to find a suitable method that can
provide better accuracy in predicting currency crisis as it is the best answer to improvise
the predictability of an early warning system.
An early warning system is modelled by using methodologies based on statistical
pattern recognition. Initially, four single classifiers are tested and the best one is chosen
as a base classifier in order to combine two or more different type of classifiers. By
creating a novel ensemble of classifiers, we then compared its performance with other
three ensembles of classifiers. Thirteen macroeconomic indicators were selected for this
study which is based on the previous literatures, economic point of views and
availability of the data. Since we believe there are other factors that affect currency crisis
besides economical factors, we took variables from the International Country Risk
Guide as well which comprising of risk from both financial and political factors. Both of
the data sets were collected from 27 countries starting the first quarter 1984 until the
fourth quarter 2011. Results from all of these experiments showed that these four ensembles have
nearly the same figures in terms of error rates, with none being bigger than 0.15. Based
on comparison between the two data sets, we found that different types of data do affect
the prediction accuracy, but it depends on respective countries itself. In summary, we
did achieve our main objective in this study, which is to find a suitable method in
predicting currency crisis. Nearest neighbour tree is considered a suitable ensemble of
classifiers that can be used to predict currency crisis in the future, with a 95.87% average
percentage of accuracy. Although it took 0.03 seconds longer on average CPU time
compared to ensembles of Support Vector Machine, it has an advantage in terms of
interpretability.
Description
Keywords
Modeling Early Warning System