II INTERNATIONAL
CONFERENCE ON ARTIFICIAL INTELLIGENCE IN ACCOUNTING, FINANCE AND TAX
September 27-28 , 1996 - Punta Umbría (Huelva), SPAIN
Research Group on
AI in Accounting and Business, University of Huelva
On September 27 and 28, 1996, there took place in
Punta Umbria (Huelva) Spain, the Second International Conference on Artificial
Intelligence in Accounting, Finance and Tax. As in the last year, the
conference was organized by the Research Group on Artificial Intelligence in
Accounting at the Universities of Huelva and Sevilla.
Ilustre Colegio de Economistas de Huelva
Universidad de Huelva
Universidad de Sevilla
·
Guillermo J. Sierra. Chairperson. University of Sevilla. Spain.
·
Amelia Baldwin-Morgan. Eastern Michigan University.
·
Barbro Back. Turku School of Economics and Business Administration. Finland.
·
Enrique Bonson. Secretary. University of Huelva. Spain.
·
Carol E. Brown. Oregon State University. USA.
·
Marco de Marco. Catholic University of Milan. Italy.
·
Julio Moreno-Davila. Enginyeria
·
Alan Sangster. The Queen´s University of Belfast.
UK.
·
Hermann Siebdrat. University of Siegen. Germany.
·
Carlos Serrano. University of Zaragoza. Spain.
·
Miklos Vasarhelyi.
The conference began on 27th at 9.30 with the welcoming speeches by Prof. E.
Bonson, Prof. M Garcia-Ayuso (representing to Prof. Sierra, the
conference chairman) and Prof. A. Sangster.
This was followed by the
opening conference:
EXPLORING STUDENTS'
PERCEPTIONS RELATED TO DEVELOPING EXPERT SYSTEMS. Prof. Glen L. Gray.
Department of
Throughout the
conference, the following papers were presented:
Session 1 (Chaired by Alan Sangster)
MODELLING INTELLIGENT
INFORMATION SYSTEMS FOR AUDITING. Eija Koskivaara. Barbro Back. Turku School of Economics and Business Administration. Kaisa Sere.
University of Kuopio. Finland.
INTEGRATING STATISTICAL
AND NON-STATISTICAL AUDIT EVIDENCE USING BELIEF FUNCTIONS: A CASE OF ATTRIBUTE
SAMPLING. Rajendra Srivastava.University of Kansas. USA.
Session 2 (Chaired by Barbro Back)
EXPERT SYSTEM TASK
SELECTION IN MANAGEMENT ACCOUNTING - THE APPLICABILITY OF THE PERROW FRAMEWORK.
Carol E. Brown. Oregon State University. USA. Alan Sangster. The
Queen's University of Belfast. UK.
MODELING INSOLVENCY
DECISIONS WITH PROCESS AND HIERARCHICAL WEIGHTED-ADDITIVE METHODS. Philip A.
Collier. University of Tasmania. Stewart A. Leech. University of Tasmania.
Australia.
INDUCTION OF DECISION
TREES AS A METHODOLOGY FOR FINANCIAL ANALYSIS. AN APPLICATION TO CRISIS IN
BANKING. Enrique
Bonson Ponte. Universidad de Huelva. Mª Pilar Martin Zamora. Universidad de
Huelva. Guillermo J. Sierra Molina. Universidad de Sevilla. Spain.
EXPERT SYSTEMS
DEVELOPMENT AS BUSINESS PROCESS RE-ENGINEERING FOR ACCOUNTING & FINANCIAL
ORGANIZATIONS: OBJECT ORIENTATION, WORKFLOW AND THE INTERNET. Hermann Siebdrat.
University of Siegen. Germany. Amelia A. Baldwin-Morgan. Eastern Michigan
University. USA. & University of Tasmania. Australia.
Session 3 (Chaired by Glen L. Gray)
BANKRUPTCY PREDICTION
USING EX ANTE NEURAL NETWORKS AND REALLISTICALLY PROPORTIONED TESTING SETS.
Marilyn M. Greenstein. Lehigh University. Pennsylvania. Mary J. Welsh. La Salle University. Pennsylvania. USA.
FORECASTING COMPANY
FAILURE: NEURAL APPROACH VERSUS DISCRIMINANT ANALYSIS. AN APPLICATION TO
SPANISH INSURANCE COMPANIES OF THE 80´s. Ignacio Martinez de Lejarza
Esparducer. Universidad
de Valencia. Spain.
ON THE TRAINING OF
FEED-FORWARD ARTIFICIAL NEURAL NETWORKS FOR ECONOMIC TIME SERIES FORECASTS
USING SCARCE AND NOISY DATA. Michael L. Kremmer. D. T. Nguyen and Louis
Sanzogni. Griffith University. Australia.
PREDICTING BANKRUPTCY IN
NON-FINANCIAL BUSINESS: AN APPLICATION BASED IN ARTIFICIAL NEURAL NETWORK
MODELS. Eugenio
del Rey Martinez. Universidad de Alicante. Spain.
Session 4 (Chaired by Rajendra Srivastava)
LEKTA II: TOWARD
REAL-TIME MACHINE TRANSLATION. APPLICATIONS IN THE BANKING DOMAIN. Jose Francisco Quesada. CICA
(Centro de Informática Científica de Andalucía). Sevilla.
Spain.
FINANCIAL INFORMATION
EXTRACTION AT THE UNIVERSITY OF DURHMAN. Marco Costantino. Russell J.
Collingham, Richard G. Morgan. University of Durham. UK.
Session 5 (Chaired by Miklos Vasarhelyi)
EXPLANATION GENERATION IN
ACCOUNTING EXPERT SYSTEMS: POSSIBILITIES FOR RUN-TIME GENERATED USER
EXPLANATIONS. Andrew Lymer. University of Birmingham. UK. Amelia A. Baldwin
Morgan. Eastern Michigan University. USA. & University of Tasmania.
Australia. Jan Scott. Charles Sturt University, Australia.
THE USE OF CASE-BASED
REASONING TO UNDERSTAND TRANSFER PRICING. Olivier Curet. Middlesex University.
Jamie Elliott. Southampton University. Mary Jackson. London Business School.
UK.
THE APPLICATION OF CASE
BASED REASONING TO THE INTERPRETATION OF FINANCIAL DATA FOR ACQUISITION
ANALYSIS. M.D. Mulvenna. R.T. McIvor. J.C.Hughes. University of Ulster. A.M.
Ward. Witt Thornton, Chartered Accountants, Belfast. UK. Presented by Alang
Sangster.
APPLICABILITY OF
CASE-BASED REASONING FOR BUSINESS PROBLEMS: A STUDY OF THREE SYSTEMS. Bonnie W.
Morris. West Virginia University. USA. Atish P. Sinha. University
of Dayton. USA.
The closing conference
was:
TOWARDS INTELLIGENT
AGENTS IN ACCOUNTING: BACK-GROUND AND POTENTIAL. Professor Miklos A.
Vasarhelyi. Rutgers University. New Jersey. Chairperson of the AI/ES section of
the AAA.
Clossing sesion
Most of the papers have
been published in a book, (G.J. Sierra and E. Bonson, editors) Intelligent
Systems in Accounting and Finance (Papel Copy, publisher: Huelva, Spain).
In
response to the growing real-world significance of expert systems and
artificial intelligence, many business schools are including these topics in
their curricula. Learning about expert systems is not just learning how to use
a new software package. It is learning about a different role for computers,
namely, aiding in decision-making. This makes expert systems particularly
challenging for both students and instructors. Unfortunately, in accounting,
accounting information system (AIS) textbooks provide little help to either the
instructors or the students.
To
better understand how students develop expert systems, Gray[1995] had 80
students maintain diaries as they worked on developing expert systems using
VP-Expert. The content analysis of the diaries resulted in a list of variables
that the students found to be important. This qualitative research became the
foundation for a standardized survey about expert systems development. This
paper presents the survey results in order to (1) describe students'
experiences developing expert systems and (2) provide guidance on methods and
tools for expert systems instruction.
Results
indicated that students experienced difficulty in developing the logical flows
and decision rules. Additionally, students encountered a variety of problems
related to the programming task, at least in part because many found the
VP-Expert book difficult to use. Nonetheless, most students avoided a high
level of frustration by using support resources, including assistance from the
professor and their peers. The on-line debugger in VP-Expert also proved
helpful to the majority of students. Overall, students found the project to be
of "average" difficulty despite the need to master both a new program
and the logic of expert systems. Most found the project relevant and useful for
their career and professional goals.
Accounting
reports are an important source of information for managers, investors,
creditors, and financial analysts. Moreover, developments in the capital
markets and in information technology have emphasised the importance of
reliable accounting information. Computerised bookkeeping provides more timely
and exact information than manual bookkeeping. The auditor´s task is to inspect
the realibity, completeness, and accuracy of accounting records. However,
auditors have to achieve an appropiate balance between effectiveness and
efficiency in their control task. In practice, this means that their statements
are based on aggregated data and ratios. Therefore, auuuditors and other
interest groups would benefit from a model that enables them to monitor the
financial performance of a firm based on more detailed available data than
aggregated data and ratios.
Researchers
have tried out different statical methods and experts systems to support in
different auditing tasks for example in the analytical review process. howevew,
these methods have not found their ways to practice to any larger extent,
mainly because they have been too complicated to use, too unreliable and the
cost of applying them has been too high compared to the benefits.
Artificial
neural networks (ANNs) are a computing paradigm that can be used as a basis for
building intelligent information systems. ANNs are suitable for a se of
applications like classification, prediction, control, and inference. Unlike
tradicional experts systems, where knowledge is made explicit in the form of rules,
neural networks generate their own rules by training examples. ANNs have been
management science techniques. For example, ANNs are feasible for those
business applications which require the solution of very complex system of
equations, recognising patterns from imperfect inputs, or adapting decisions to
changing environments.
Integration
of neurocomputong and auditing expertise in a computer system might be an
answer to auditors´ paradox of time and cost when auditing firms. Thre are
already preliminary research results on applying neural networks on detecting
errors in monthly balances of firms. Our aim in this study is to get further
evidence on neural networks´ capabilities in this area of auditing.
The
main purpose of this article is to show how one can integrate statical evidence
from attribute sampling with non-statinal evidence within the Dempster-Shafer
belief function framework. In particular, the article shows: (1) how
determinate the sample size in attribute sampling to obtain a desired level of
belief that the true attribute occurrence rate of the population lies in a
given interval; (2) what level of belief is obtained for a specified interval
given the sample result; and (3) how to integrate non-statical evidence with
the statical evidence arising from the attribute sampling. These issues are
important to the auditor ans therefore we use auditing examples to illustrate
the process. As intuitively expected, we find that the sample size increases as
the desired level of belief in the interval increases. In evaluating the sample
results, we again find results that are intuitively appealing. For example,
provided the sample occurrence rate falls in the interval B for a given number
of occurence of the attribute, we find that the belief in B, Bel(B), increases
as the sample size increases. However, if the sample occurence rate falls
outside of the interval then Bel(B) is zero. Note that, in general, both
Bel(notB) are zero when the sample occurrence rate falls at the points of the
end points of the interval. These results extend similar results already
available for variables samplings. However, the auditor faces an additional
problem for attribute sampling: how to convert belief in an interval for
control exceptions inte belief in an interval for material misstatements in the
financial statements, so that can be combined with evidence from other sources
in implementations of the Audit Risk Model. We discuss this problem, and
investigate conversion methods that are consistent with current auditing
practice.
Theoretically,
management accounting is well suited to expert systems development yet,
compared to other accountancy domains, there are relatively few reported
management accounting expert systems,. Recent reported analysis of the Perrow
framework appears to provide an explanation for this situation. This paper
considers whether the Perrow framework can be applied in this way, concludes
that it can not, but proposes that it does present a tool that could usefully
be applied to help explain expert system success and failure, and to assess
whether or not to proceed with expert systems development in the first place.
The
aim of this research is to create a validated descriptive model of insolvency
decisions (MIND) using an expert systems research approach.
A
decision made by insolvency practitioners, in attempting to determine the
future of a company in financial distress, typically involves numerous factors.
These result from a financial analysis of the situation, judgments about the
attitudes and abilities of stakeholders and environmental factors about the
type of business that is in distress.
While
insolvency practitioners often follow some proprietary procedures and
guidelines, there is no widely accepted or explicit framework to assist in the
decision-making process. This is reflected in a paucity of extant literature.
In this paper, related literature on corporate financial distress and on
modeling aspects of human attitudes and abilities is discussed.
In
the context of the sparse literature on insolvency decision making, a
preliminary goal of this work is to identify the factors and processes that
reflect the current practice of insolvency practitioners. For this we have
selected appropriate methods of knowledge acquisition.
We
argue that many judgments in insolvency are best modeled by combining several
items of evidence robustly using weighted-additive methods, in preference to
more complex models of uncertain reasoning. These judgments are then combined
using process models.
In
this paper, we aim to show the usefulness that algorithms of induction of
decision trees may have for financial analysis. For this purpose, we use this
method to explain the crisis of Cajas Rurales in
Elliott
(1992) described the present wave of information technology (IT) changes as the
third wave. The movement of this wave highlights the need for changes in
accounting and financial organizations that reflect the current changes in
business and management that are drivenby the evolution of information
technology. The purpose of this paper is to show how the concepts of BPR can be
integrated naturally with the development of expert systems to re-engineer
accounting processes, and how new concepts of object orientation, workflow and
the internet are evolving to impact how business processes are designed.
Bankruptcy
prediction is a topic of vital interest to accountants, financial advisors,
investors, and creditors. Each of these constituents has a vested interest in
accurately assessing the reasonableness of a going-concern assumption. From an
accounting standpoint, rules and standards, such as SAS 59, detail an auditor's
responsibility to evaluate an entity when there is "substantial doubt
about the entity's ability to continue as a going concern for a reasonable
period of time . . ."(par. 6). Further, this auditing guideline lists a
number of categories of evidence that provide a basis of evaluation, including
negative trends in key financial variables, credit difficulties, and external
business conditions. Unfortunately, the statement does not provide any
guidelines on weighting the various factors; however, it does place a burden on
the accountants to diligently assess the financial statements along with other
external data in order to attest to the viability of the entity as a going
concern enterprise.
Given
the recent litigious climate, investments in research projects aimed at
developing and improving bankruptcy prediction models by Big Six accounting
firms should be no surprise. Neural networks are one alternative to traditional
multi-discriminant analysis and LOGIT regression that are currently being
investigated as a bankruptcy prediction modeling technique. For example, KPMG
Peat Marwick recently funded a study examining the usefulness of neural
networks in predicting commercial bank failures. Prior neural network research
has, for the most part, been conducted using ex post facto research designs
with sample sets that do not accurately reflect the true relative proportions
of bankrupt to non-bankrupt firms in the population. Thus, most of the studies
are not generalizable to other data sets.
The
objective of this study is to assess the viability of a neural network in a
more realistic setting than has generally been examined in the past. The data
sets which are tested in this study reflect the true proportion of firms that
actually fail, which is less than one percent of all firms. Further, all models
are developed ex ante to test subsequent years' data sets. In order for neural
network, bankruptcy prediction models to be useful and relevant, they must be
designed to work in realistic settings, rather than carefully designed
matched-pair settings. This study also examines whether the neural network
models are relatively static or dynamic over time. Further, the performance of
neural networks is also evaluated relative to traditional logit models.
Although prior studies have found that traditional statistical techniques
perform as well as neural networks, the limitations noted for previous neural
network models apply to comparative studies as well. Therefore, the research
gives insights into the performance of a logit model in a more realistic
setting.
Results
show that neural networks can be applied in more realistic settings than have
previously been studied. However, a traditional logit model often outperforms a
neural network. Further analysis of the results indicate that the two methods
should be viewed as complementary, rather than as alternatives, in predicting
corporate failure.
A
brief review of prior neural network studies is presented in the next section.
In the sections that follow, an overview of neural networks is discussed, the
methodology and sample selection procedures are presented, and the model
training techniques and testing results are detailed. The paper concludes with
a discussion of the results.
In
this work a Backpropagation Neuron Net Model is proposed to forecast the
company failure. The model is applied to the spanish insurance companies of the
80's. An alternative Discriminant Analysis-based model is carried out, and both
performances are compared concluding a better forecasting results by the Neural
model.
This
paper is devoted to the issue of training and generalisation of feed-forward
artificial neural networks for forecasting economic time series. The problem of
overtraining, when using scarce and noisy data, is addressed using the method
of cross validation. The drawbacks, contradictions and limitations of this
method, as described in the literature, are reviewed in detail. A variant of
the cross-validation method is then discussed. The training method described
here is based upon the authors's own experiments, and the work of Masters
(1993:180-185), and Hecht-Nielson (1990:110-117). It is specific to the
training of a fully connected feed forward neural network using scarce and
noisy macroeconomic variables, and has been successfuly implemented in an application
which generates forecasts of Australian interest rates.
The
aim of this study is to present additional empirical evidence on bankruptcy
prediction in non-financial copmpanies, adjusting a model based on neural
networks in which financial ratios are used as input data. In relation to the
study carried out in our coutry with this type of companies, the present
workpresents the novelty of considering a different method of classification
with a longer period and larger sample, different financial ratios as well as
some peculiarities in the methodology in the treatment of the data.
Consequently, two models with an extremely good capacity for prediction are
obtained. The first, with data from the year before the failure, allows the
crisis to be successfully anticipated with in 94.3% of the cases, using five
financial ratios. The second, with six ratios, anticipates failure with success
in 82.1% of the cases two years beforehand.
The
study is organized in the following way. In section 2 the definition of
bankrupt is specified as well as the sample used. In section 3 the ratios to be
used as initial predicting variables are selected. In the fourth section the
application of the connectionist paradigm to classification tasks is briefly
reviewed. Then, the results obtained are presented and finally, some relevant
conclusions are drawn.
Natural
Language Processing, or Computational Linguistics, is an area in Artificial
Intelligence that attempts to get computers to understand and produce human
language. In fact, processing natural languages has always been one of the
central research issues in artificial intelligence, both because of the key
role language plays in human intelligence and because of the wealth of
potential applications. We can say that natural language is the perfect human
computer interface.
Section
1 outlines the basics in this research field. Section 2 discusses their
application in Accounting, Finance and Tax, including natural language
interfaces, message handling, on line news stories or text generation. Section
3 is devoted to Lekta II, a machine translation tool developed by the author
which was used in a research project between Telefonica I+D and the
This
article describes the financial information extraction system under development
at the University of Durham. Differently from many others developed in the
past, the system has been designed for use in real situations and to alleviate
the "data overload" from which traders, brokers, fund managers etc.
suffer nowadays. The system is based on the financial activities approach, for
the identification of the relevant templates to be extracted from the source
articles. The goal of the system is to summarise financial news (either from
newspapers or on-line services) producing specific templates associated to the
various financial activities. The templates produced can be successfully used
for a "meta-analysis" of the news on price behaviour. The system uses
natural language processing techniques developed at Durham University which are
based on deep natural language processing techniques, as opposed to
pattern-matching or statistics.
Explanation
facicitieshave historically been considered fundamental to the succesful
implementation of Expert Systems. However, first generation ES explanation
provided limited useful user feedback often typified by llittle more than rules
tracing or simple one line, pre-defined, fixed text responses. Whilst
explanation facilities of this form were found in most early commercial
systems, evidence suggests this type of explanation is little used in practice.
This
paper argues that explanations as a vital element of ES has yet to archieve its
real potential. It argues that explanation facicities can offer vital access to
system knowledge if constructed properly and development appropriately thereby
significantly improving the value of a system in a commercial domain. It argues
that the ability to generate explanation at runtime would significantly improve
the way ES could be applied in commercial domains and that the ability to do
this is now possible. There are however, a number of considerations to be
hightlighted in attempting on the nature of the knowledge of the domain and the
ways in which it can be represented. This paper discusses these issues and uses
examples from the accounting and finance domain to illustrate the commercial
possibilities of runtime explanation generation.
To
the authors' knowledge this is the first study to use Case-Based Reasoning
(CBR) to explore the domain of Transfer Pricing. Data has been collected from a
sample of UK multinational enterprises (MNEs) concerning their transfer pricing
policies. This paper illustrates the applicability of CBR technology to
evaluate the relationship between the occurrence of transfer pricing audits and
companies' characteristics (eg organisational features and contingent
variables). Initial findings suggest that there are patterns of transfer
pricing audits which need further explanation. CBR may have a valuable part to
play in learning about both transfer pricing audits and the problem domain in
general (including transfer pricing policies).
This paper illustrates
how knowledge based systems technology, and in particular case based reasoning
technology, can assist in the interpretation of financial data for acquisition
analysis. The authors discuss a prototype model designed to help a company
formulate an effective acquisition decision.The model includes both rule based
and case based reasoning technology. There is a particular focus on the issues
involved in the application of case based reasoning techniques to the
automation of the acquisition decision-making process. The development of this
hybrid system has shown that it is possible to use knowledge based systems
methods to build a hybrid decision support system in the area of financial
strategic decision making, especially if the domain is well defined, has a
large number of factors to be considered and the relevant knowledge is
available. This system has illustrated the immense impact that knowledge based systems
technology and case based reasoning can have upon business processes.
Three
case-based reasoning models are examined. Task characteristics that point to
the applicability of case-based reasoning are identified and discussed. This
paper presents some preliminary work toward the development of a framework for
choosing among case-based reasoning, expert systems, and neural network models.
Intelligent
agent technology is one of the faster growning areas of research and
Internet-related commercial endeavours. It is, however, an ill defined field,
with many overstated claims and few specific areas of applications. Both
accounting and finance have great potencial as fields of application. However,
at this stage there are very few, if any, known applications. Most of the
current proposed applications are in the primary research stage. For further
development of the field, it is necessary to create an operacional definition
of the field, understand its extant composition, and to postulate a program of
research and application development. Such a theoretical work should be of great
value as a foundation for an emerging field.
Many
applications today,- particularly those for the World Wide Web, claim somo
functional "intelligence" where agents within software will perform
tasks for a principal. This paper, explores the spectrum of software agency;
from the automated "softbots" that are presently being implemented,
to the concepts and projects of the future that are more accurately described
as intelligent agents.
The
operational definition section should provide some assessment of the current
state of intelligent agent technology and who some of the key players are,
to-date. The analysis focuses on academic and comercial research. The paper
describes the basic mechanics for agency and how agent developers are tackling
the challenge of intelligent agents within networked computing environments.
The
state-of-the-art section explores the commercial agent landscape. Commercial
efforts are fust the beginning of the capabilities and potencial of intelligent
agents. Agents are classified inte categories, and examples, from practice, are
provided.
Important
questions are raised in this discussion, somo technical, some statutory and
some behavioral. For example, does accounting/finance require special gent
technology, wath are the laws necessary to allow agents in electronic commerce,
and whether consumers will achieve the level of trust for intelligent agents´
capabilities to effectively represent them when making important decisions.