Abstract

Neural networks are one of the most widely used artificial intelligence methods for financial time series analysis. In this paper we describe the standard application of neural networks and suggest that it has two shortcomings. First, backpropagation search takes place in sum of squared errors space instead of risk-adjusted return space. Second, the standard neural network has difficulty ignoring noise and focusing in on discoverable regularities. Both problems are illustrated with simple examples. We suggest ways of overcoming these problems. 1 Introduction Ever since McCulloch and Pitts [8] published their landmark paper describing a neural calculus, researchers have explored how biologically inspired networks might be employed to solve a wide class of problems. In recent years, Wall Street, in its never ending search for new ways to beat the market, has turned to neural networks as a possible answer. The attention focused on neural networks for financial time series analysis stems ...

Citations