Uluslararası tahvil piyasalarında vade yapısının ve sabit getirili menkul kıymet getirilerinin makine öğrenimi teknikleri kullanılarak tahmin edilmesi

In this study, I focus on predicting bond risk premia in Turkish Eurobonds market using machine learning methods. Machine learning uses statistical learning techniques to gather useful structures of a data set without being explicitly programmed. In recent years machine learning has become a very popular topic and shown very good results in a wide variety of fields, but there is a lack of research in the field of term structure modeling. In order to predict Turkish Eurobond returns, I implemented several machine learning models such as OLS, PCA, Ridge, Lasso, Elastic net and neural networks. The raw data set I used comprises of Turkey Government Eurobond yields between 2005 and 2020, inclusive. Both monthly and yearly returns are estimated separately. Zero-coupon rates and forward rates are calculated from the raw data and used as left-hand site elements for machine learning predictions. Macroeconomic variables are also added to forward rates as factors. I compared the out-of-sample performance of the models and I found that Penalized linear regression yields the best results for excess bond return prediction, providing nearly 10% out-of-sample R2. Neural networks are the second-best performer yielding around 3-4% out-of-sample R2. Plus, adding macroeconomic variables to the models slightly improved the results by 2-3%. Also, yearly returns estimation performed better than monthly returns for OLS, Ridge, Lasso and Elastic net regressions, but not for neural networks.