City Research Online

Abstract

In this thesis we investigate the question of asset price predictability. The two major themes that we focus on are firstly; whether machine learning and statistical modelling techniques, which impose less restrictive assumptions on asset price dynamics than do classical linear methods, can be used to forecast and trade financial markets to a degree greater than that which traditional asset pricing models would lead us to expect and secondly; to what extent model combination/ensemble strategies can add value in this pursuit. The approaches used include support vector regression (SVR), k-nearest neighbours (KNN), trading rules, linear regression (LR) and the random subspace ensemble method.

We investigate these two themes using inherently data-driven models across datasets of sufficient size to render statistically meaningful results in three self-contained contexts. The first piece of empirical work compares the relative forecasting performance of SVR, KNN and LR models when applied to predicting daily returns of 58 UK stocks in the FTSE 100 over 4000 days. Bootstrap simulations are used to shed further statistical light on model performance.

Secondly, we investigate the extent to which model combinations can improve forecasting performance with the use of the random subspace ensemble method for constructing ensembles of linear regression models to predict the returns of a portfolio of FTSE 100 stocks. The primary ensemble consists of 62500 component models estimated by randomly sampling subsets of the feature set and the final result combined via a majority vote.

Lastly, we conduct an in-depth study of the channel break-out trading rule over a portfolio of 37 futures markets. We borrow a page from the book of modern portfolio theory where it is the performance of individual markets in the context of a portfolio that is ultimately of interest rather than on an individual basis. This approach is rarely used in the literature but is able to shed more light on the question of trading rule efficacy. Bootstrap resampling is employed to derive robust performance statistics. Our results show the Sharpe Ratio of the portfolio to be three times greater than of individual markets as a result of diversification in addition to being greater than that of S&P500 benchmark.

We did not set out in an attempt to refute the weak form of Fama's (1970) classic taxonomy of information sets or, colloquially, "to beat the market"; nonetheless, some of our results suggest economically significant returns.

Publication Type:	Thesis (Doctoral)
Subjects:	H Social Sciences > HG Finance
Departments:	Bayes Business School > Finance Doctoral Theses Bayes Business School > Bayes Business School Doctoral Theses

Export

Downloads