Yapay sinir ağlarını kullanarak 2019 kriz öngörüsü üzerine bir deneme

Abstract

Son yıllarda tüm dünyada birçok ekonomik kriz yaşanmıştır. Bu krizler sadece ekonomik olarak değil sosyal ve politik sonuçlar da doğurmuştur. Yaşanan krizlerin etkisi tek bir ülkeyle sınırlı kalmamış, domino etkisiyle diğer ülkeleri de etkisi altına almıştır. Özellikle son 20 yıldaki finansal serbestleşme ile birlikte, ekonomisi çok daha dışa bağımlı ve kırılgan olan gelişmekte olan ülkeler krizlerden en çok zararı görmüştür. Globalleşme fırsatları ve tehditleri beraberinde getirmiştir. Finansal serbestleşmenin adından ihracat ve ithalatın artmasıyla ülke ekonomileri birbirine bağlanmıştır. Bu gelişmelerin getirdiği refah dönemi bir süre beklentileri çok yükselterek, ülkeleri altından kalkamayacağı finansal koşullara sürüklemiş ve sonrasında, 1997'de Asya'da veya 2008'de Amerika'da yaşandığı gibi, büyük krizlerin tetikleyicisi olmuştur. Türkiye'de dünyada yaşanan bu gelişmelerden etkilenmiştir. Ayrıca 20. yüzyılda ülke içi politik ve ekonomik gelişmelerin de etkisiyle, 1990 sonrası dönem şiddetli krizlerle sahne olmuş ve krizlerin etkisi birçok yönden hissedilmiştir. Özellikle 1994 ve 2001 yıllarında yaşanan krizlerin süresi kısa olmasına rağmen etkisi çok ciddi boyutlarda hissedilmiştir.Krizlerin küresel çaptaki etkileri yüzünden krizlerin nedenlerinin tespit edilmesi birçok çalışmanın konusu olmuş, dönemlerine göre farklı kriz modelleri geliştirilmiştir. Krizlerin nedenleri her zaman ölçülebilir değildir. Bazen ülke içi sosyal ve politik nedenler, bazen de piyasaların kırılganlığından faydalanan kişi veya kurumlar ekonomilerin zor duruma düşmesine neden olabilmektedir. Bununla birlikte uluslararası rezervlerdeki azalma, enflasyondaki artış gibi ölçülebilir ekonomik parametrelerdeki değişimler de krizlerin tetikleyicisi olabilmektedir. Erken Uyarı Sistemleri krizlerin, oluşmadan önce öngörülmesi için yapılan modelleri içermektedir. Bu şekilde kriz önlenemese bile, ekonomideki zayıflıklar tespit edilerek buna yönelik önlemler alınabilir ve krizlerin olumsuz etkilerinin en aza indirilmesi sağlanabilir. Literatürde lineer, probit ve logit modeller ile erken uyarı sistemleri geliştirilmiştir.Yapay sinir ağları (YSA), insan sinir sistemi modellenerek geliştirilmiş yapılardır. YSA'larda öğrenme algoritmaları sayesinde girdiler ve çıktılar arasında mantıksal bir model oluşturulmakta ve bu modeller yeni veriler üzerinde kullanılmaktadır. Bu yapılar mevcutta sağlık, finans, eğitim gibi birçok alanda kullanılmaktadır. YSA yapıları öğrenme, adapte olma kabiliyetleri sayesinde, krizlerin öngörülmesinde önemli bir rol oynayabileceği düşünülmektedir.Bu çalışma Yapay Sinir Ağı kullanılarak Türkiye için erken uyarı modeli oluşturulmuştur. Modelde belirlenen 1992-2018 döneminde belirlenen ekonomik göstergeler üzerinden yapay sinir ağı eğitilmiştir. Oluşturulan bu model 2019 1. Çeyrek dönemi için kriz öngörüsü yapılmıştır. Sonuç olarak 2019 1. Çeyrek döneminde kriz beklendiği sonucuna varılmıştır.

In recent years, many economic crises have been experienced all over the world. These crises brought up not only economical but also political and social consequences. Also the effect of a crisis was not limited to one country but it affected other countries like domino effect. Especially developing countries suffered the most from the economic imbalances. Together with the financial liberalization in the last 20 years, the economies of the developing countries became more fragile and dependent to global economy. Therefore these economies respond to the ups and downs of the global economy rapidly. Globalization brought opportunities and threats together. As a result of the financial liberalization, import and export increased between countries. These interactions bound the economies to each other. The period of prosperity brought about by these developments increased expectations. As a consequence, economies were dragged into financial situation which are difficult to meet. These result in the devastating financial crises like in Southeast Asia in 1977 or in America in 2008.Turkey also affected from these conditions. In addition, due to the domestic political and economic developments in the country, Turkey witnessed destructive economic crises after 1990. Especially, in 1994 and 2001, although the crises was short in duration, the impact was felt in serious dimensions Because of the global effect of crises, identifying the causes of crises was chosen as the subject of many studies. The main purpose of these studies was to come up with models that define indicators of crises. However, due to the diverse nature of the crises and the fact that indicators changes in time as the global economic conditions change, the crisis models are divided into different generations.The first generation models, created by Krugman (1979), are successful in explaining the crises in 1970s. The model links financial crises to the macroeconomic weaknesses of a county. According to the model, a country can decide to close the budget deficit with money supply. As a result of the excess domestic currency in the market, inflation rises and the value of money decreases. In order to keep the fixed exchange rate, The Central Bank sells foreign currencies in the market. As a result international reserves decrease to an uncontrollable level and the government is forced to give up fixed exchange rate (Yüksel, 2015).Second generation models focus on the situations which macroeconomic conditions are stable and yet the country confronts a crisis. The model states that, in order to overcome speculative attacks to the currency, the governments could either increase the interest rate or the money supply. Increase in interest rate affects the domestic market adversely and leads to unemployment. On the other hand increasing money supply leads to currency crises. This model is successful in understanding 1992-93 ERM crises.Third generation models are created as a result of the 1997 Southeast Asian Crises. Model states the effect of the financial problems of the domestic banks over the domestic currency. Also model emphasizes the contagion effect of a crisis. As the interactions between countries increase, the effect of the financial problems in one country spread to the others. Initially the problems spread to the neighbor countries and to the countries with economic interaction. Then like in 1997 Southern Asian crisis and 2008 Global Financial crisis, the economic problems spread to the whole world.Early Warning Systems are created to predict a crisis before happening. Even crisis cannot be prevented, economic weaknesses can be detected and necessary measures can be taken. In this way the adverse effect of the crisis to the global economy can be minimized. According to the literature signal, probit and logit methods are used in early warning systems for financial crises. Signal model was created by Kaminski and Reinhart (1996). In the model, a threshold value is determined for each economic indicator like inflation, import, export. If the value of the indicator crosses the threshold a signal is produced. Also a signal period is determined. Signal period is the time period of the predicted crisis for example 24 months. If there is a crisis in the period the signal is a good signal, if not it is a false signal or noise. By using the model, Kaminski and Reinhart conclude that the behavior of international reserves, real exchange rate, domestic credits and domestic inflation are meaningful indicator to a predict crisis Logit and probit models are used when there is two or more level for a dependent variable. In early warning systems dependent variables are crisis and no crisis. On the other hand the independent variables are the economic indicators such as inflation and interest rates. The difference between logit and probit models is the function used in the models. In logit model logistic distribution function is used. However probit models use cumulative distribution function. Logit and probit models are nonlinear functions. Therefore 1 point increase in an independent variable does not affect the dependent variable in the same way. Also because of the nonlinearity, the result of the function is kept between 0 and 1. Kunt and Detragiache (1998) and Fratscher ve Bussiere (2002) used logit model to predict crisis. On the other hand Esquivel and Larrain (1998) and Oktar and Dalyancı (2010) used probit to evaluate economic indicators and anticipate crisis.Although many studies are conducted, they could not be successful in predicting 2008 Global Financial Crisis. New approaches are investigated for better understanding of the indicators of crisis and make better anticipations. Because of their learning and adapting capacity, Artificial Neural Networks are powerful tools in analyzing and making predictions. Therefore these tools can be used in an early warning system to predict an upcoming economic disaster. In this study, an early warning model for Turkey is created with ANN. The input values of the model are separated into two categories. First category is the macroeconomic parameters of Turkey. The parameters are import, export, inflation, M1 balance, USD exchange rate, credit profile of private banks, short term debt, and international reserves. The second group is used to evaluate the contagion effect. In order to evaluate the effect the countries Turkey has the highest volume of export and is chosen. Between those countries, the countries which have larger effect in the economy and easy to find economic data are selected. The selected countries are Germany, England and America. In order to evaluate economic conditions of these countries short term interest rate and inflation parameters are used in the model. Output of the model is 1 for the periods in which Turkey is in crisis and 0 for the others. First the model is trained with the input and output data. The output value is not just 1 and 0. Therefore the output values bigger than 0.5 is accepted as 1 (crisis) and the others are accepted as 0 (no crisis). The network is trained until all the outputs are evaluated correctly. After the training, the network is used to predict the upcoming period, 2019 first quarter. The result of the prediction is 0.9384. The value is bigger than 0.5 which means a crisis is expected in the first quarter of 2019.