Forecasting Economic Time Series using Locally Stationary Processes: A New Approach with Applications (Volkswirtschaftliche Analysen)

Stationarity has always played an important part in forecasting theory. However, some economic time series show time-varying autocovariances. The question arises whether forecasts can be improved using models that capture such a time-varying second-order structure. One possibility is given by autoregressive models with time-varying parameters. The author focuses on the development of a forecasting procedure for these processes and compares this approach to classical forecasting methods by means of Monte Carlo simulations. An evaluation of the proposed procedure is given by its application to futures prices and the Dow Jones index. The approach turns out to be superior to the classical methods if the sample sizes are large and the forecasting horizons do not range too far into the future.-- Provided by Publisher Content: 1 Introduction; 2 From stationarity to local stationarity; 2.1 Stationary stochastic processes; 2.1.1 A short introduction to stationarity; 2.1.2 Spectral representation of stationary processes; 2.1.3 Stationary ARMA processes; 2.1.4 Asymptotical properties of the sample partial autocorrelations of a stationary AR(p) process; 2.2 Locally stationary processes; 2.2.1 Evolutionary spectrum; 2.2.2 Definition of local stationarity; 2.2.3 Local covariance estimation; 2.2.4 Local partial autocorrelation; 2.2.5 TVAR; 3 Estimation. 3.1 Maximum likelihood estimation with the Kullback-Leibler information divergence3.2 Sieve estimation; 4 Forecasting; 4.1 Prediction in the case of stationarity; 4.2 Approaches to forecast time series using TVAR processes; 4.3 Iterative stages in the selection of a model; 4.4 Simulations; 5 Application; 5.1 Motivation; 5.2 Futures data; 5.2.1 Course of action; 5.2.2 Practical evaluation of TVAR processes on futures series; 5.3 Dow Jones index data; 6 Conclusion; 6.1 Contributions; 6.2 Possible directions for future research; References; Notations and abbreviations; List of tables. List of figuresA Appendix; B GAUSS source code; B.1 Fitting time-varying autoregressive models to non-stationaryprocesses; B.2 Procedures for computing the coefficient functions.