Stock Prediction - A Neural Network Approach
K.G. Nygren · 2004
Predicting stock data with traditional time series analysis has proven to be difficult. An artificial neural network may be more suitable for the task. Primarily because no assumption about a suitable mathematical model has to be made prior to forecasting. Furthermore, a neural network has the ability to extract useful information from large sets of data, which often is required for a satisfying description of a financial time series. This thesis begins with a review of the theoretical background of neural networks. Subsequently an Error Correction Neural Network (ECNN) is defined and implemented for an empirical study. Technical as well as fundamental data are used as input to the network. One-step returns of the Swedish stock index and two major stocks of the Swedish stock exchange are predicted using two separate network structures. Daily predictions are performed on a standard ECNN whereas an extension of the ECNN is used for weekly predictions. In benchmark comparisons, the index prediction proves to be successful. The results on the stocks are less convincing, nevertheless the network outperforms the naive strategy. Sammanfattning Att prediktera borsdata med traditionell tidsserieanalys har visat sig vara svart. Ett artificiellt neuralt natverk kan vara mer passande for uppgiften. Framst darfor att inga antaganden om en passande matematisk modell maste goras innan prediktering. Vidare har ett neuralt natverk formagan att extrahera anvandbar information fran stora datamangder, vilket ofta ar nodvandigt for en tillfredsstallande beskrivning av en finansiell tidsserie. Det har examensarbetet borjar med en genomgang av teorin bakom neurala natverk. Darefter definieras och implementeras ett felkorrigerande neuralt natverk (ECNN) for en empirisk studie. Bade tekniskaoch fundamentala data anvands som indata till natverket. Enstegsavkastningar for Generalindex samt tva stora aktier pa Stockholmsborsen predikteras med tva separata natverksstrukturer. Dagliga prediktioner utfors pa en standard ECNN medan en utokad variant av ECNN anvands for veckoprediktioner. Vid jamforelser med andra strategier visar sig prediktionen av index vara framgangsrik. Resultaten for aktierna ar mindre overtygande, likval presterar natverket battre an den naiva strategin. Acknowledgements First and foremost I thank Prof. Kenneth Holmstrom at Tomlab Optimization AB for initialising this project. Tomlab Optimization AB has been supportive with financial data, computer and office facilities. Furthermore I am grateful to Dr. Hans-Georg Zimmermann at Siemens AG Corporate Technology Department for giving me an excellent introduction to the SENN software. I also acknowledge staff at Siemens AG for technical support. Finally, I am thankful to Dr. Thomas Hellstrom for useful comments.