UK Nonlinear News, May 1996

Recent Thesis

MIT/TU Berlin

Connectionist Forecasting:
Modeling Financial Data With Neural Networks

This thesis is available online at
http://www.cs.tu-berlin.de/~ghoff/papers.html
A shortened version given as a paper is also available.

Thesis:
Connectionist Forecasting: Modeling Financial Data With Neural Networks
(MIT/TUB)
Paper:
Function Approximation in the Financial Field
with an Application to the Interest Rate Sector
(Conference on Computational Intelligence for Financial Engineering; New York 1995)

Abstract

Quantitative analysis in the financial markets has traditionally been dominated by linear, parametric modeling approaches. Recent theoretical and empirical results suggest that nonlinear, nonparametric, multivariable regression techniques might offer a chance to discover and capture nontrivial relationships between variables more effectively. In this work ways of improving models and thus forecasts are explored by adapting two different ways of specifying Connectionist Networks: Radial Basis Function Networks (RBF) and Multi Layer Perceptrons (MLP). By employing these techniques we gain the potential to model complex data more effectively while at the same time we largely avoid imposing any particular and possibly incorrect model assumptions. Evolution Strategy and a speeded up error backpropagation procedure are utilized to estimate model parameters. To illustrate the application potential nonlinear yield models for a particular bond sector (Bunds) are estimated. For comparison benchmark models using a linear multivariable and a random walk approach are also estimated. To address model reliability bootstrap estimates of the expected model errors are derived. Here RBF and MLP models are found which consistently outperform the benchmark models in an out-of-sample scenario. Furthermore the problem of optimal regressor selection is addressed. We find that a brute force approach yields better regressor combinations than the usual step wise approach based on Akaike's information criterion. To bridge the problem of incorporating monthly data in a model of daily changes a naive method to approximate financial data to a different time scale based on RBF is discussed.

Source: Guenther Hoffmann (gunho@hp832.informatik.hu-berlin.de)


<< Move to UK Nonlinear News Issue 4 Index Page (May 1996).

Last Updated: 1st May 1996.