An Expert System for the evaluation
of stocks & shares

Traversum AB, a company that specialises in share and stock dealing in the Swedish stock exchange, worked with Novacast AB - XpertRule Software's distributor in Scandinavia - to develop an expert system to advise on how to deal in stocks and shares. The system was developed using XpertRule and has been in daily use by Traversum AB since 1987.

According to Mr Krister Ahnelšv of Traversum, their "hit rate" of dealing correctly with stocks has increased from 60% the luxury of spending long periods of time examining and analysing to 90%. The system even predicted the stock market crash in October 1987 by advising to sell almost all shares during the ten day period beforehand.

Dealings in the stock market have dramatically changed in pace and volume compared to ten years ago. Today brokers do not have the market and assessing the best deals. Minutes can often be crucial.

There is an ever increasing demand for quick, consistent and exact analysis in share dealing. Conventional techniques, such as technical charts or fundamental analysis of economic factors often fail to provide quick, precise and consistent predictions of the market.

Mr Krister Ahnelšv wanted an analytical tool to combine the various techniques and to develop an automated system. This is when Traversum AB and Novacast AB decided to use XpertRule, which provides rule induction analysis, to build "Stockmaster". The data used to develop the system was extracted from actual movements of stocks and shares over a period of twenty years.

An expert from Traversum AB first identified around thirty factors or attributes that are related to the decision for selling, waiting or buying shares. However, the number of attributes identified as important meant that a very large number of share movements had to be considered. This was both impractical and beyond the available time and resources.

To solve this problem, the set of thirty attributes was divided into several meaningful subsets. The attributes of each subset were related to a specific methodology for dealing in shares. There was the "technical" group, the "fundamentalist" group, the "index" group, etc.

Data was then extracted according to these groups keeping the outcome of "sell, wait or buy" in the examples of share movements. XpertRule rule induction was then used to develop the rules from each group.

Each rule corresponded to a real life opinion in share dealing. The outcomes of the groups were then used as attribute values to build the master rules to provide the final decision.

The first version of Stockmaster had 10 modules and around 400 rules. As the system was used, new aspects became apparent and the system was extended gradually. The current version has 17 modules and over 3000 rules. The number of questions asked by Stockmaster to arrive at a decision is typically between 20 and 40.

It provides advice on share dealing according to the latest information supplied. This consists of two types:

(i) General information about the state of the market and other general technical information such as indices, point and figure diagrams, etc. This information is fed into the system once every day.

(ii) Specific information about the share or stock the dealer wants advice on.

Mr Krister Ahnelšv believes that the Stockmaster project has shown that it is quite feasible to build practical and valuable systems, even in domains difficult to specify, by being able to use structuring and data analysis techniques.

Stockmaster is delivered on standard PC machines. Currently it is used by 20 security dealers in six European countries.