A Hybrid Forecasting Model for American Dollar/Turkish Lira Exchange Rate Using Time Series Analysis and Deep Learning Models

Exchange rate forecasting, as well as stock market forecasting, has been an important topic for investors, researchers and analysts. In this study, it is aimed to create a hybrid model that predicts the exchange rate trend by performing financial emotion analysis and time series analysis. For this purpose, the proposed hybrid model was constructed in three stages: obtaining and modelling text data for financial sentiment analysis, obtaining and modelling numerical data for time series analysis, and blending two models. It is thought that this is the first study in the literature that uses social media platforms as a source for financial emotion analysis and blends it with time series analysis methods using numerical data. Moreover, it is the first study in the literature that performs the US Dollar/Turkish Lira exchange rate prediction of the trend by performing a financial sentiment analysis and using a hybrid model. The contribution of the study to the literature can be summarized in five sections: In the section one, in order to perform financial sentiment analysis, the data collected from Twitter has been prepared and modelled by using a couple of pre-processing stages such as parsing the documents, lemmatization, stemming, removing the unused characters and the words, normalizing the words. The data set, which is ready for modelling, is classified using word embedding methods such as Word2vec, GloVe, fastText, and deep learning models such as CNN, RNN, LSTM, both individually and in combination. This study is the first attempt as it is understood from the literature in terms of performing the financial sentiment analysis by using the combinations of deep learning models with word embedding methods. In the second section, time series analysis was performed by using Simple Exponential Smoothing, Holt-Winters, Holt's Linear, and ARIMA models along with real US Dollar/Turkish Lira exchange rates. In the third section, two different prediction models were combined with a weighted majority algorithm to form a hybrid model. In the fourth section, in order to prove the performance of the proposed model, the real Twitter and the exchange rate data sets from January 1, 2018 to December 12, 2018 were used to achieve financial sentiment analysis and time series analysis respectively. In the fifth section, as a result, the performance of the proposed model is significantly superior to that of the literature.