| Composed of two empirical studies, this dissertation investigates the effectiveness and utility of neural networks in accurately forecasting the short-term behavior of a variety financial time series.; The first study employs a recurrent neural network as a nonlinear function approximator to forecast the out-of-sample return on two stock market indices: the Dow Jones Industrial Average and Standard and Poor's 500 Composite Index. The use of an extensive, multivariate information set and a global stochastic maximization algorithm distinguishes this study from prior work. The data set investigated encompasses daily observations from 1970 through 1993, with the following forecast exercise undertaken. For a variety of model sizes, the network task is to approximate the weekly, monthly or quarterly conditional mean return. These forecasts are conditioned on a daily information set containing a number of index-specific and market-wide variables, term structure and corporate bond yields, and calendar variables. Network performance is evaluated by out-of-sample normalized mean-squared error, sample statistics describing the joint distribution of forecasted and actual returns, and tests for market-timing ability and nonlinear independence. A further performance evaluation concerns the construction of trading portfolios with transaction costs. Bootstrapping techniques are also applied to construct surrogate distributions of the out-of-sample statistics. Finally, impulse-response and input ranking analysis are addressed to characterize the network solutions. It is found that neural networks perform more than adequately when compared with a benchmark linear model, and that is possible to generate large risk-adjusted returns over and above simple buy-and-hold strategies.; The second study again using this nonlinear model concentrates on accurately forecasting daily high and low Deutschmark futures prices. Results are presented for a number of model estimations which employ eight years of in-sample data to forecast two years of out-of-sample data. Two experiments are conducted: (1) a data shuffling exercise to determine the sensitivity of network performance with respect to partitioning of data into in-sample and out-of-sample data sets; and (2) a Monte Carlo analysis of network performance with respect to variation in parameter search initialization. Generally, evidence is found of substantial ability to conditionally forecast both market direction and magnitude. |
