Finance Applications with the Web Mining Software Agent PISA

Abstract

Financial market websites contain valuable financial market information from banks, brokers, exchanges and financial service providers, e. g. stock quotes, option prices, exchange rates in particular and other information concerning exchange markets. The Internet offers these data in many cases real time or near time. Comparable up-to-dateness usually is only offered by commercial finance databases that must be paid like Reuters (www.reuters.com) or Bloomberg (www.bloomberg.com). Subscription fees usually are quite high. The Internet often offers the same information for free. In addition the Internet offers data not provided by any commercial service provider. Usually only data are provided for which a certain demand exists. If the demand is too low the data are not provided because provision is not cost efficient. Provision costs in the Internet are much lower. Here, up-to-dateness, cost-freeness and diversity of the offered data are exploited to build financial databases for free. Database quality is nearly as good as of commercial financial databases. Financial market websites usually do not contain static content but dynamic content. The presented data change frequently and usually are stored in support databases. When a webpage is requested the needed information is queried from the support database and the results are put into a template. This results in highly structured HTML or XML webpages. Usually only few different templates are used. Once a scheme is recognized, it can be used to identify the demanded information on other pages of the same website. Here, schemes are used to extract specified information to generate databases and dense time series. A time series is defined as a series or function of a variable over time. This means that a particular variable takes a particular discrete value at a sequence of (often equidistant) points in time. Here, quotes are used as variables. Time series can be used to train artificial neural networks. For example the FAUN (Fast Approximation with Universal Neural networks) neurosimulator can use neural networks to predict real market option prices, see [6], or to make forecasts, see [5] and [10]. Those neural networks need input data with highest data quality, i. e. patterns for the supervised learning, here. Outliers caused by wrong inconsistent data sets reduce the quality of the trained neural networks significantly.