Very Small Information Systems/SS2006/Group E Stock Monitor

The Stock Monitor by Nikolaj Ostri and Christian Christiansen

This page is the home of the Stock Monitor, we will during the spring 2006 keep this page updated with information concerning our project.

Project Desription

The scope of this paper is to develop software for data mining on real-time stock rates to support buying decisions for stock traders. The application is intended to run on a mobile device like a mobile phone or PDA.

Trading stocks has become something every person with a computer can do. Unfortunately trading stocks can be very risky and especially short-term trading (less than one year)is close to 50/50 betting.

The concrete implementation of the stock monitor will first happen on a normal PC and several experiments will be run to decide which algorithm is most accurate and able to deliver the best return on investment.

Given this is an application which is intended to be run on a very small processor we will after deciding on algorithm implement a prototype on a FPGA board.

Project Results

Our initial purpose for this project was to prove that it is possible to develop technical stock trading applications to be used on small processors, like the ones used in mobile devices. We hypothesized that even given the very limited resources we would be able to create extra-ordinary returns on investments. In order to do that we created a number of experiments which all used different classifiers, but the same data. We constructed our experiments with regard to the fact that stock prices are a time series, thereby not independent of each other. After the experiments it became clear that IB5 (5-nearest-neighbors) was the best-performing algorithm on our domain data, hence we decided to implement it on a small processor, to act as prototype. IB5 performed quite well on the small processor and were able to make the same predictions as WEKA’s IB5. During our evaluation we proved that IB5 on the small processor would be able to produce extra-ordinary returns, thus we can conclude that we fulfilled our goal with this project and proved that even given limited resources data mining on complex data is possible and development of future applications within this domain is possible. Our evaluations raised some questions concerning how reliable and generic our solution is and if it really is the success rate which is the best indicator for how suited an algorithm is for technical trading. Further studies should be conducted within this domain, but our initial findings claim that the number of wrong classifications leading to buying is a better indicator. Other studies which could be done in the future, was to try more advanced algorithms on the small processor and maybe derive even better results. In addition we could implement our solution in a portfolio of applications all aimed at predicting stock movements and thereby improve our success rate and in the end the return on investment.