Nonstationarity, nonlinear dependence, and prediction: An application to the Treasury Bill futures market

This study describes the time series properties of U.S. Treasury Bill futures prices with special emphasis on unit root nonstationarity, nonlinear dependence, and prediction. Although most research of financial markets assumes that market prices follow a specific martingale process, namely the random walk, recently researchers have begun to question this assumption. This assumption implies futures prices must contain a unit root, yet many studies are inconclusive or contradictory on this point. In chapter 2 several tests for nonstationarity are applied and it is shown that futures prices undoubtedly contain a unit root.;A more formal analysis of the random walk hypothesis is conducted in chapter 3 by looking at both linear and nonlinear dependence of first differences of prices. Nonparametric and parametric tests of linear dependence are conducted and the results indicate that the data contains no significant linear dependence. However, when tests for nonlinear dependence were conducted, the results from every test indicated the nonlinear dependence.;Based on the results from chapter 3, chapter 4 estimates nonlinear models and uses them for prediction. In this chapter much is learned. First, some nonlinear models are excluded simply by their poor estimation performance. Second, when comparing the models' predictive performance to the random walk, it becomes clear that the nonlinearities of the data are exploitable. Two of four models are able to perform better than the random walk especially in shorter horizons. Third, the best nonlinear model is chosen after comparing the predictions of all the nonlinear models against each other. It is shown that the bilinear model is the best of the nonlinear models. Finally, it is shown that the bilinear model outperforms the popular autoregressive, conditional, heteroskedastic (ARCH) model.