Combining Models

Abstract

There is a large body of literature describing how best to combine models of various types of processes (Bunn, 1975; Bunn and Kappos, 1982; Bunn, 1985; Jacobs, 1995; Harrald and Kamstra, 1997; de Menezes et al, 1998; Sharkey, 1998). Two such groups of methods are ensemble methods (Freund and Shapire, 1995; Breiman, 1994), and mixture of experts methods (Jacobs et al., 1991; Jordan and Jacobs, 1994). Combining forecasts effectively is a non-trivial process in the case where high levels of noise exist, as can occur in the financial markets. The current trend is towards combining predictive models, rather than employing large monolithic predictors. The advantage of the former is that such a methodology could be more efficient in terms of training time (Lu and Ito, 1998). Further, it is also possible to achieve a lower generalisation error from the combiner (Krogh and Vedelsby, 1995), as well as to prevent overfitting. The models considered in this paper are of the prediction of bond returns over the next few time points, as for example for the next month for monthly data, based on the values of a range of inputs provided initially from economic analysis. We will consider how an effective solution be achieved for such a task, an important component in the overall prediction process for such time series with important financial implications