STUDY OF DISCRETE CHOICE MODELS AND FUZZY RULE BASED SYSTEMS IN THE PREDICTION OF ECONOMIC CRISIS PERIODS IN USA

Abstract

This paper studies the economic recessions and the financial crisis in US economy, as these crisis periods affect not only USA but the rest of the world. The wrong government policies and the regulations in bond market among others lead to the longest and deepest financial crisis since the Great depression of 1929. In this paper we examine three models in order to predict the economic recession or expansion periods in USA. The first one is the Logit model, the second is the Probit model and the last one is a fuzzy rule based system binary regression with sigmoid membership function. We examine the in-sample period 19132005 and we test the models in the out-of sample period 2006-2009. The estimation results indicate that the fuzzy regression outperforms the Logit and Probit models, especially in the out-of sample period. This indicates that fuzzy regressions provide a better and more reliable signal on whether or not a financial crisis will take place. Furthermore, based on the estimated values for the period 1913-2009 we estimate the forecasts to investigate if the economic recession will be continued or not during 2010. The conclusion is that Logit model presents a signal that the economic recession will be continued during the whole period 2010, while based on Probit and fuzzy regressions the economic recovery might begin in the second half of 2010.Keywords: financial crisis, discrete choice models, fuzzy rules, fuzzy regression, sigmoid membership functionJEL Classification: C53, E17(ProQuest: ... denotes formulae omitted.)1. INTRODUCTIONThe economic facts and the financial crisis in the last three years lead to doubts among the economic society and governments of which methodology must be followed and what regulations should be set up. The subprime mortgage crisis which took place in USA and became apparent in 2007 led to great weakness in the financial system and the financial industry regulation. Furthermore, even if economists and policy makers nowadays have in their disposable management portfolio, improved estimation methods, access in long databases and mainly the use of very fast computers and advanced algorithms and computational methods, where they are able to get the estimating results in a few minutes they didn't manage to forecast the crisis or even if they predicted the crisis they were unable to take the appropriate measures in order to eliminate it or to confront the situation. This indicates that models are not always enough to take decisions as equivalently significant is the correct and efficient government policies and the optimal set up of the financial regulations. In this paper we examine the methodology and the models which can be applied in order to predict very successfully the crisis periods, rather than to comment or propose the formulation of regulations in financial and market industry.Various approaches have been developed and applied in financial, banking and currency crisis. One of these approaches is the application of Logit and Probit models (Eichengreen and Rose,1998; Demirguc-Kunt and Detragrache ,1998; Frankel and Rose, 1996; Glick and Rose, 1998; Glick and Moreno, 1999). In this approach the dependent variable is a dummy variable taking two values, the value 1 for crisis periods and the value 0 for no crisis periods. More specifically Demirguc-Kunt and Detragrache (1998) found that low GDP growth high inflation and high interest rates might lead to economic downturns, which are the most significant variables among other macroeconomic factors. Eichengreen and Rose (1998), found that the Northern interest rates are strongly connected with the onset of banking crises in the developing countries. The methodology of Probit and Logit models is that allows for statistical testing, identifying the sign, the magnitude and the marginal distributions of the explanatory variables to the onset of crisis. On the other hand this approach confronts the problem of misspecification errors and serial correlation. …