![]() |
|||
![]()
|
![]() |
![]() |
![]() |
3.2.2.2. Knowledge Representation and Reasoning Mechanism CREDEX is built on the SNARK inference engine and formalism. The knowledge is represented in first- and, in some cases, second-order logic. The system has about 300 rules, of which over half are for the experiential experts. Reasoning heuristics are formalized in classic fashion through production rules. Variables are instantiable by the objects of the memory and propagated in the rule conclusion. In the working memory, the properties of the objects, the associations between them, and statements are represented by associative triplets in the form <object relation value>, where relation and value can also be structured objects. The CREDEX control strategy is mixed: it is both goal- and data-driven. 3.2.2.3. System Performance The system operates on two modes: Diagnosis (D) mode, which gives a final report after each function of the company has been diagnosed; and Assessment (A) mode, which, using specific heuristics, can skip certain functions and therefore draw quicker conclusions. The system starts by asking the user questions about:
The reasoning process then starts and the system guides the user through a logical sequence of questions on the company, provides an elementary diagnosis on each company function, prints out a quality judgment on the function, and recommends actions to be taken to improve this quality if possible in the case of this analysis. It ends by printing out a global qualitative judgment of the company in terms of corporate risk. The system is very fast. A session in A mode takes 1.3 seconds, and a session in D mode takes 6 to 8 seconds on an IBM 3090 computer under MVS. Reference: Condensed from Pinson, S. A multi-expert architecture for credit risk assessment. The Credex system, in Expert Systems in Finance, edited by D.E. O'Leary and P.R. Watkins, 1992. 3.2.3. COMPASS COMPASS was developed for the Bank of Scotland over a 7-year period, coming online in 1994. It was initially applied to all commercial loan applications in excess of $375,000. The development costs were recouped during the validation phase and, since its introduction, the bank has suffered no losses on new loans of that magnitude. COMPASS was written in C++ and runs on both Windows and OS/2 platforms. Its modular design means that the data and information specific to the Bank of Scotland can be removed, leaving a generic shell that can be adopted by any bank in any market to perform similar assessments on their loan portfolios and loan applications. By mid-1995, COMPASS was being used or tested by banks in the U.K., Ireland, Europe, and a number of other locations worldwide. 3.2.3.1. Model COMPASS is a rule-based system. Its knowledge base contains the encapsulated expertise of a loan officer that seeks the relevant knowledge from the loan officer conducting the assessment, applies it to the financial ratios calculated by COMPASS and other relevant information held, and makes recommendations concerning the lending proposal. The bank's loan officers are required to use the system, though they are not required to adhere to its recommendations. During a consultation, COMPASS accesses centrally stored information on the customer and on existing agreements with the customer, and utilizes accounting data for the customer spreading over 7 years: three historic, the present year, and three future periods. Through the inclusion of three key modules -- VIABILITY (the potential of the customer to attain and support the proposed activity), SAFETY (the assessment of the means by which the loan facility may be recouped), and CREDENTIALS (how well the customer is known to the bank, including qualitative factors such as its management style) -- quantitative and qualitative customer-specific information is filtered by COMPASS in the context of the environment in which the customer operates, accounting values are adjusted accordingly, and the resulting recommendations are presented to the loan officer in context so that he or she may explain the recommendations to the customer in a relevant and informative way. To assist this process, at the end of each COMPASS consultation, a natural language report is produced containing recommendations to the bank loan officer on the customer's request for lending. The entire consultation process takes between 5 and 30 minutes, compared with up to a week under the system in use prior to COMPASS being introduced. 3.2.3.2. Validation The first prototype was completed in 1988 and the first fully working system started testing in 1992. Testing was initially by the bank's staff. In 1993, they were augmented by a team of chartered accountants from one of the largest international accountancy firms. 3.2.3.3. Implementation By mid-1994, COMPASS was installed in 20% of the bank's corporate branches throughout the U.K., covering, as a result, 80% of the bank's corporate lending portfolio. Within 1 year, it was installed throughout the bank and it had been extended for use in respect of all loans in excess of $15,000. Steps were then taken to redefine the terms of all existing loans that it assessed as being at risk, thus minimizing the risk they held for the bank. Reference: Condensed from Sangster, A. The Bank of Scotland's COMPASS -- the future of bank lending?, Expert Systems with Applications, Vol 9, 1995, 457-468. 3.2.4. CUBUS CUBUS provides an evaluation of a commercial customer's annual accounts. Testing of over 1000 cases has shown it to be as good as an experienced expert at indicating the financial strength or the default probability of a company. It is written in ART-IM and takes about 3 minutes to complete a full analysis of a customer's annual accounts. Reference: Condensed from Wolf, M.F. New technologies for customer rating: Integration of knowledge-based systems and human judgement. Intelligent Systems in Accounting, Finance and Management, Vol. 2, 1995, 289-301.
|
![]() |
|
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. |