Publication: Forecasting GDP growth of 10 euro area countries : LSTM vs. NARX neural network”
Abstract
Bu tezde, 10 Euro Bölgesi ülkesinin (Avusturya, Belçika, İspanya, Almanya, Fransa, Finlandiya, İrlanda, İtalya, Hollanda ve Portekiz) GSYİH büyümesi, yapay sinir ağı algoritmasının en güncel tahmin yöntemlerinden ikisi kullanılarak tahmin edilmektedir. Bu iki tahmin yöntemi, tekrarlayan dinamik sinir ağı literatüründe en popüler teknikler olan Uzun Kısa Süreli Bellek (LSTM) ağı ve dışsal girdilere sahip doğrusal olmayan otoregresif sinir ağdır (NARX). LSTM ve NARX'in tahmin performansı, testlerin ortalama hata karesi, ortalama mutlak hata ve r-kare değerleri ile karşılaştırılmıştır. LSTM ağı İspanya, Finlandiya, Fransa, İtalya ve Portekiz gibi bazı ülkeler için daha iyi performans gösterirken, NARX yöntemi Avusturya, Belçika, Almanya, İrlanda için LSTM'den daha iyi performans göstermiştir. Hollanda örneğinde, her iki teknik de hemen hemen aynı sonucu vermiştir. Her bir değişkenin model tahminine katkısı Shapley Additive Explanations (SHAP) değerleri ile değerlendirilmiştir. Bu tezde, SHAP değerlerine göre, her açıklayıcı değişkenin ekonomik büyümeye olumlu ve olumsuz etkisi, ülkelere göre değişmektedir ve etkinin gücü de değişkenlik göstermektedir. Bunun nedeninin ülkelerin ekonomik ve demografik özelliklerinden kaynaklandığı sonucuna varılmıştır. Ayrıca, döviz kurunun etkisi modelde en alt sırada yer almaktadır, buna göre seçilen 10 avro bölgesi ülkelerin GDP büyümesi tahmini için döviz kurunun katkısı çok küçüktür. Bunun nedeni, tüm EA ülkelerinin yerel para birimi olarak Euro kullanmasıdır. Küresel para piyasasında Euro'nun güçlü bir satın alma gücüne sahip olması, EA ülkelerinin para biriminde bağımsız olmalarını ve döviz kuru dalgalanmalarından fazla etkilenmemelerini sağlamıştır.
In this thesis, the GDP growth of 10 Euro Area countries (Austria, Belgium, Spain, Germany, France, Finland, Ireland, Italy, The Netherland, and Portugal) are predicted using two of the latest forecasting methods of artificial neural network algorithm. These two forecasting methods are Long Short-Term Memory (LSTM) network and the Nonlinear Autoregressive Exogenous Inputs (NARX) network models which are the most popular techniques in the recurrent dynamic neural network literature. The forecasting performance of LSTM and NARX is compared with the root mean square error, mean absolute error, and r-square values of the tests. While the LSTM network shows better performance for some countries such as Finland, France, Italy, and, Portugal, the NARX method outperforms LSTM for Austria, Belgium, Germany, Spain, and Ireland. In the Netherland case, both techniques produced the nearly same result. The contribution of each variable to the model prediction is evaluated by Shapley Additive Explanations (SHAP) values. In this thesis, according to the SHAP values, the positive and negative impact of each feature varies for countries and the strength of the impact shows variety. We conclude this result as the economic, geopolitical, and demographic features of countries. Besides, the contribution of the exchange rate to the prediction ranked in the lowest order means that the exchange rate slightly contributes to GDP growth of selected euro area countries. The reason lies in the fact that all EA countries are using Euro as local currency. In the global monetary market, Euro has a strong purchasing power, and this makes EA countries to be independent in currency and not get affected much by exchange rate fluctuations.
In this thesis, the GDP growth of 10 Euro Area countries (Austria, Belgium, Spain, Germany, France, Finland, Ireland, Italy, The Netherland, and Portugal) are predicted using two of the latest forecasting methods of artificial neural network algorithm. These two forecasting methods are Long Short-Term Memory (LSTM) network and the Nonlinear Autoregressive Exogenous Inputs (NARX) network models which are the most popular techniques in the recurrent dynamic neural network literature. The forecasting performance of LSTM and NARX is compared with the root mean square error, mean absolute error, and r-square values of the tests. While the LSTM network shows better performance for some countries such as Finland, France, Italy, and, Portugal, the NARX method outperforms LSTM for Austria, Belgium, Germany, Spain, and Ireland. In the Netherland case, both techniques produced the nearly same result. The contribution of each variable to the model prediction is evaluated by Shapley Additive Explanations (SHAP) values. In this thesis, according to the SHAP values, the positive and negative impact of each feature varies for countries and the strength of the impact shows variety. We conclude this result as the economic, geopolitical, and demographic features of countries. Besides, the contribution of the exchange rate to the prediction ranked in the lowest order means that the exchange rate slightly contributes to GDP growth of selected euro area countries. The reason lies in the fact that all EA countries are using Euro as local currency. In the global monetary market, Euro has a strong purchasing power, and this makes EA countries to be independent in currency and not get affected much by exchange rate fluctuations.