Comparing Estimation Methods for Non-Gaussian Affine Term Structure Models

This thesis investigates two methods used to estimate the yield curve using affine term structure models (ATSMs) with non-Gaussian factors. ATSMs describe the dynamics of the yield curve using affine transformations of latent factors. The factors behave according to Gaussian or non-Gaussian dynamics. The methods of AYt-Sahalia and Kimmel (2010) and Creal and Wu (2015) both promise quick and efficient estimation of ATSMs with non­Gaussian factors, but it is unclear which model performs better. The method of Ai:t-Sahalia and Kimmel (2010) approximates the likelihood of the latent state variables through Hermite expansions, while the method of Creal and Wu (2015) approximates the entire ATSM by a discrete-time version. Ex ante, it is not clear which of these approximations performs better. This research performs a sensitivity analysis to the amount of starting values and observations. A comparison between the two methods is done using an efficient amount of starting values and observations, as using too much of either increases computation time without signilicantly increasing performance. The Creal and Wu (2015) method results in a lower root-mean-square error (RMSE). This lower RMSE is mostly noticeable in the yields corresponding to the lowest maturity. The RMSE is comparable between the two methods in the yields corresponding to the higher maturities. The lower RMSE comes at the cost of an increased computation time. An empirical estimation of the parameters using real-world data is performed for both methods, which supports the conclusion that the CW method outperforms the ASK method for real-world data.