Library Header Image
LSE Theses Online London School of Economics web site

Topological optimisation of artificial neural networks for financial asset forecasting

He, Shiye (2015) Topological optimisation of artificial neural networks for financial asset forecasting. PhD thesis, London School of Economics and Political Science.

Text - Submitted Version
Download (1MB) | Preview


The classical Artificial Neural Network (ANN) has a complete feed-forward topology, which is useful in some contexts but is not suited to applications where both the inputs and targets have very low signal-to-noise ratios, e.g. financial forecasting problems. This is because this topology implies a very large number of parameters (i.e. the model contains too many degrees of freedom) that leads to over fitting of both signals and noise. This results in the ANN having very good in-sample performance on the data used for its training but poor performance outof-sample for forecasting. The main contribution of my research is to develop a new heuristic method called “ANN reduction” for optimising the topological structure of a feed-forward ANN in order to improve its out-of-sample performance (using an RMS measure). The research concentrated on the topological optimization of the graph representing an ANN, which reduces the effective degrees of freedom of the ANN whilst still maintaining its feed-forward (but incomplete) topology. Such reductions in the number of parameters have been attempted before in the literature, but our procedure is of a different (graph theoretic) nature and (in extremis) optimal for small-size ANNs. Two applications of the ANN reduction are also implemented and programmed for empirical simulations. For this purpose, two datasets generated from deterministic functions and three datasets derived from foreign exchange market prices are used for evaluating the ANN reduction applications. These applications generate new ANN topologies with some clear performance advantages over those obtained by the best complete ANNs, improving the generalization (out-of-sample) performance by up to 27.6% compared to the complete ANN on the function generated datasets and up to 14.1% on the financial forecasting problem for the FX data.

Item Type: Thesis (PhD)
Additional Information: © 2015 Shiye He
Library of Congress subject classification: H Social Sciences > HG Finance
Sets: Departments > Management
Supervisor: Christofides, Nicos

Actions (login required)

Record administration - authorised staff only Record administration - authorised staff only


Downloads per month over past year

View more statistics