Neural networks, financial trading and the efficient markets hypothesis

Abstract

The efficient markets hypothesis asserts that the price of an asset reflects all of the information that can be obtained from past prices of the asset. A direct corollary of this hypothesis is that stock prices follow a random walk, and that any profits derived from timing the market are due entirely to chance. In the absence of any ability to predict the market, the most appropriate strategy---according to proponents of the efficient markets hypothesis---is to buy and hold. In this paper we describe a methodology by which neural networks can be trained indirectly, using a genetic algorithm based weight optimisation procedure, to determine buy and sell points for financial commodities traded on a stock exchange. In order to test the significance of the returns achieved using this methodology, we compare the returns on four financial price series with returns achieved on random walk data derived from each of these series using a bootstrapping procedure. These bootstrapped samples contain exactly the same distribution of daily returns as the original series, but lack any serial dependence present in the original. Our results indicate that on some price series the return achieved is significantly greater than that which can be achieved on the bootstrapped samples. This lends support to the claim that some financial time series are not entirely random, and that---contrary to the predictions of the efficient markets hypothesis---a trading strategy based solely on historical price data can be used to achieve returns better than those achieved using a buy-and-hold strategy.