Brought to you by EarthWeb
IT Library Logo

Click Here!
Click Here!

Search the site:
 
EXPERT SEARCH -----
Programming Languages
Databases
Security
Web Services
Network Services
Middleware
Components
Operating Systems
User Interfaces
Groupware & Collaboration
Content Management
Productivity Applications
Hardware
Fun & Games

EarthWeb Direct EarthWeb Direct Fatbrain Auctions Support Source Answers

EarthWeb sites
Crossnodes
Datamation
Developer.com
DICE
EarthWeb.com
EarthWeb Direct
ERP Hub
Gamelan
GoCertify.com
HTMLGoodies
Intranet Journal
IT Knowledge
IT Library
JavaGoodies
JARS
JavaScripts.com
open source IT
RoadCoders
Y2K Info

Previous Table of Contents Next


3.4.1.1. Model

The rule-induction method is based on the learning-from-example approach. The rule induction system is presented with examples of a decision (its inputs and outcomes) and attempts to induce a decision model. A software system called ACLS (Analog Concept Learning System) was used to analyze past examples and formulate decision rules. Rules were generated to predict both an expert market analyst's prediction of the market and the actual market's movement. A single expert with 12 years of experience as stock market analyzer was used as the source of expertise. The expert also had experience in giving verbal advice to a large audience -- he was writing a newsletter in which he gave biweekly recommendations on the stock market. These recommendations provided the basis for the model. The Dow Jones industrial average (DJI) was used as reference point for the system.

3.4.1.2. Knowledge Representation

The knowledge was represented in the form of induced rules. The expert identified the 20 cues that he found most relevant from a list of 20. The cues and outcomes set up the structure of the example database. Three types of outcomes were used to categorize the weekly recommendations: bullish (forecasting an upward trend), bearish (forecasting a downward trend), and neutral (indicating that either call was too risky). The example database consisted of data at closing time on a certain time on a weekday for each of a number of weeks. For the prototype, it was data at closing time on Friday for each of 108 weeks during the period of 20 March 1981 to 9 April 1983. Most of the data was collected from the Wall Street Journal.

3.4.1.3. Validation

The validation was performed in two steps. In the first, an initial database of 80 examples of the expert's predictions from a 1.5-year period was validated. The first test split the examples into two groups of 40 on a random basis. One group was used to induce a rule using ACLS; the other was used to test the induced rule. The degree of consensus between rule-generated predictions and the expert was 57.5%. The hit rate was improved to 65% when the number of examples was increased to 60 for rule inducing, leaving 20 examples for testing.

In step 2, each of the example databases was expanded to 108 examples. Two situations were replicated: one predicting the expert's predictions and one predicting actual market movements. The average hit rate predicting the actual market's movement was 64.4. The expert was correct only 60.2% of the time.

Reference: Condensed from Braun, H. and Chandler, J.: Predicting stock market behavior through rule induction: an application of the learning/from/example approach. Decision Sciences, Vol. 18, 1987, pp. 415-429.

3.5. NEURAL NETWORK APPLICATIONS

There have been many attempts to apply neural network technology to the finance domain. One such system is in use in a major financial institution in the U.K. where it is used to predict the type of customers who may go into arrears on their house financing loans. The neural network was built using 150,000 past loan records and is used in order to assess whether or not to proceed with a loan application. In testing, it was found to correctly predict the risk in 91.8% of 160,000 loans.

Reference: Torsun, I.S. A neural network for a loan application scoring system, The New Review of Applied Expert Systems. Vol. 2, 1996, 47-62.

3.6. HYBRID SYSTEMS

This class of product has yet to make a major impact in this domain. However, such systems are being developed and put to use. IEEE Expert, for example, carried a description of an intelligent software system, called ProfitMax, in its "new product" section in February 1996. ProfitMax provides transaction-based, real-time authorization action decisions for managing the profitability of credit card portfolios. It uses neural networks, expert rule bases, and cardholder behavior profiling technology to analyze each cardholder account and predict its profit. Historical data and three neural network-based models -- credit risk, revenue, and attrition risk -- are used to predict profitability. The profit evaluation is customized to the using issuer's definition of financial profit. ProfitMax updates the profitability prediction whenever a new transaction or event takes place. Transaction-based scoring enables online, real-time decision-making, even during authorization.

The system also offers profitability-based decision-making. Business strategies and account management tactics can be entered into the system's rule base to automatically apply actions (for example, to extend a credit line) based on each account's expected behavior and profit. ProfitMax implements the actions and monitors to measure their impact. Portfolio managers can use simulation to perform decision-making in a test mode before implementing the actions in the live cardholder base.

ProfitMax interfaces with the issuer's authorization system, whether installed on a mainframe or a UNIX platform. Client PCs handle management and reporting functions.

Source: IEEE Expert, Vol. 11 (1), 1996, 87.

4. RESEARCH ISSUES

The research issues can generally be classified into in three main groups. The first includes statistical, model development, and validations issues; the second deals with cognitive issues, including understanding cognitive tasks; and the third is directed at explanation and user acceptance issues.

4.1. STATISTICAL, MODEL DEVELOPMENT, AND VALIDATION ISSUES

For the most part, statistical, model development, and validation issues are not just issues related to expert systems, but to any statistical procedure. To be useful in a real application, a trading system, for example, should be validated using an objective function that is very specific to the trading strategy and objective (e.g., profit maximization). Fitting systems into trading strategies and market mechanisms may require use of a hybrid system.

Another validation issue relates to the comparison of expert system results with neural networks, traditional statistical models, and human experts. Different systems cannot always be evaluated in the same way or through use of the same measurement metrics. However, in the end, the comparison of one system to another, or to a human expert, must be a comparison of how well the clearly defined objectives are met.


Previous Table of Contents Next

footer nav
Use of this site is subject certain Terms & Conditions.
Copyright (c) 1996-1999 EarthWeb, Inc.. All rights reserved. Reproduction in whole or in part in any form or medium without express written permission of EarthWeb is prohibited. Please read our privacy policy for details.