Neural network ensembles for financial time-series prediction and risk management

Abstract

Neural Network Ensembles for Financial Time-Series Prediction and Risk Management. Recently, neural networks have become popular tools for modelling financial markets. Much of this popularity can be attributed to the fact that neural networks are universal approximators i.e. they can (in theory at least) approximate any complex non-linear function to arbitrary accuracy. Given the complexity of modern financial markets, and the non-linearity that is widely accepted as driving the relationships between related financial variables, neural networks are potentially very powerful. This was recognised by financial market practitioners and researchers very early on. However, when systems were developed and tested, performance was typically poor. This is because non-parametric universal approximators such as neural networks can have serious limitations, especially when applied to model noisy, real-world systems such as financial markets. One of the most serious is high-variance or instability i.e. small changes in training set and/or parameter selection can cause significant changes in generalisation (prediction) performance. Another problem (closely related to instability) is the tendency of neural networks to over-fit, essentially “memorize” their training sets, which also causes poor generalisation performance. The final, but probably most serious limitation of neural networks in the context of financial modelling, is the absence of model transparency - neural networks are "black-box" estimators. Given the regulatory pressures today on financial institutions and traders to manage and limit their exposure to risk, such black-box models are just not good enough - the trader cannot determine what factors drive the model or more importantly how much confidence he can have in specific predictions. In this thesis we attempt to (at least partly) solve these problems.