The comparison of the Affine Term Structure Models, the PCA and the Diebold-Li Model in connection with forecasting ability

Absztrakt (kivonat)

The purpose of this paper is to compare the forecasting ability in connection with the yield curve of the Diebold-Li model (DL model), the principal component analysis (PCA) and the family of the affine termstructure models (ATSM). This comparison will take place at a theoretical field and at an empirical field using a specific dataset and then checking the results on other datasets. The estimation and prediction of the yield curve are extremely important problems that are connected to the risk management and the derivatives pricing really closely. The shape of the yield curve define the profitability of many business and investment strategies, also regulators and large financial institutions develop macroeconomic scenarios on the yield curve. Accordingly it is really substantial to estimate and forecast the yield curve with the least possible error, and there are a lot of different models capable of this. The three main type of these methods are the economic, the statistic and the stochastic group of models. In this paper I am presenting the theoretical properties of the modeltypes and then I describe the three special model that I mentioned above in a detailed way. These methods cover all of the three groups, which make it easy to demonstrate the differences between the modeltypes with them. I describe how to estimate and how to use in connection with forecasting the DL model, the PCA based model and the ATSM models, which are the simplest in their categories. Finally I am introducing the forecasting ability of the models on the U.S. yield curve with statistical analysis, and I am making a robustness check using datasets on the British and Japanese yield curves. In the analysis I have found that the DL model performs the best on the original U.S. dataset used by Diebold and Li, while the other two methods cannot outperform the random walk in forecasting. However the robustness check showed that the forecasting performance of the DL model is worse on other datasets, especially during more volatile periods, while the results suggest that the ATSM method is better during these times. In my way of thinking it would worth a further research to combine these two models, for instance to switch between them depending on the volatility of the period. Also the DL model’s nice achievement during less volatile periods can suggest that it could be useful in other fields of finance and economics, but this needs more examination, too.