A neural network for stock selection

Verfasser: Schaar, Roland

Sachtitel: A neural network for stock selection

Abstract
In this paper I summarize the results of an experiment to determine wheter it is possible to use neural network methods to extract useful information from financial time series. The results of this experiment with a limited amount of data suggest that the answer is yes and demonstrate empirically that stock markets are not efficient. A method is presented showing that there is information in past prices of stocks which can be exploited for forecasting future price changes and therefore for beating the market. This method can be used for managing an investment fund. It is not applicable to small investors because it requires high frequency trading. We make predictions with a relatively short time horizon that demand low transaction costs. But though there is enough information in past prices that can be used for beating the market after all transaction costs. The treatise gives evidence that modern neural network technologies can be used to gain additional information out of past prices that are not accessible to investors using conventional methods.

Betreuer

Haase V.