Türkı̇ye getı̇rı̇ eğrı̇sı̇nı̇n derı̇n öğrenme ve ekonometrı̇k modeller kullanılarak tahmı̇n edı̇lmesı̇: Karşılaştırmalı bı̇r analı̇z

This thesis investigates the modeling and forecasting of the Turkish Treasury bond yield curve using both traditional time series models and deep learning approaches. The study focuses on evaluating the performance of models such as Autoregressive (AR), Random Walk (RW), the Dynamic Nelson-Siegel (DLNS) model with xed and time-varying de cay parameters, and Long Short-Term Memory (LSTM) networks. Daily zero-coupon yield data between July 2012 and February 2025 were used, covering maturities rang ing from 3 months to 10 years. Forecasting performance was assessed using an out-of sample expanding window framework for different horizons. Root Mean Squared Error (RMSE) was employed as the main accuracy metric, and the Diebold-Mariano (DM) test was applied to determine whether performance differences across models were statisti cally signicant. The empirical ndings show that while simpler models such as AR and RW perform well in the short term, their forecasting accuracy deteriorates as the horizon increases. In contrast, LSTM models—particularly those trained on latent factors derived from the Nelson-Siegel framework—demonstrate superior generalization performance in medium- and long-term forecasts. Notably, the DLNS model with a time-varying λ also outperforms benchmark models at longer horizons. These results highlight the value of incorporating structured factor dynamics and deep learning techniques for forecasting complex nancial time series such as the yield curve.