Detecting and Explaining Business Exceptions for Risk Assessment
Lingzhe Liu
1
, Hennie Daniels
2
and Wout Hofman
3
1
Rotterdam School of Management, Erasmus University, Burg. Oudlaan 50, Rotterdam, The Netherlands
2
CentER, Tilburg University, Warandelaan 2, Tilburg, The Netherlands
3
Technical Sciences, TNO, Brasserplein 2, Delft, The Netherlands
Keywords: Explanatory Analysis, Interoperability, Decision Support, Risk Assessment.
Abstract: Systematic risk analysis can be based on causal analysis of business exceptions. In this paper we describe
the concepts of automatic analysis for the exceptional patterns which are hidden in a large set of business
data. These exceptions are interesting to be investigated further for their causes and explanations. The anal-
ysis process is driven by diagnostic drill-down operations following the equations of the information struc-
ture in which the data are organised. Using business intelligence, the analysis method can generate explana-
tions supported by the data.
1 INTRODUCTION
“Management by exceptions” has long been a
philosophy for business administration, in which
management can be described as a reflex arc of
monitor-control loop: the manager perceives the
environment of a company, forms an expectation,
and decides on the operations planning; additional
decisions will be made when deviations from the
expectation occur. Once an exception is detected, the
manager needs an explanation “why the exception
occurred”, so that he or she can make informed
decisions on subsequent (re-) actions – whether and
how to treat the exception.
In recent years, with the prevalance of Enterprise
Resource Planning (ERP) systems and the rising
awareness of the strategic value of business data,
companies continuously collect data about its
internal operations and external environment.
Business intelligence (BI) and analytics has been
vigorously applied in industry, translating data into a
competitive edge (Davenport, 2006). “Management
by exceptions” is then endowed with new
implication of “detecting and managing risks
proactively”, rather than the old ways of “reactive
fire-fighting” (Sodhi and Tang, 2009), with the new
terms of Risk Management or Risk Based Decision
Making. Exceptions are early risk indicators, albeit
not necessarily risky themselves. A company is
assumed to be homeostatic, that is, it can self-adapt
and operater normally unless the exception exceed a
threshold. At that point the exception turns into a
(materialized) risk. Risk management addresses the
vulnerability of the system – the condition in which
an exception will turn into a risk. In the analysis of
risk, it is important to understand the risk
propagation: how a seemingly small exception
causes a catastrophic system-wide failiure (see e.g.
Lund et al., 2011, Ch. 13). If such weak signal of
risk can be detected early in time, it leaves more
space for reaction and mitigation (Sodhi and Tang,
2009). Presumably, the pattern of risk propagation
must be implied in the historical events records of
business exception. Yet, to our best knowledge,
currently there is hardly any research on the general
methodology for analysing business exceptions
systematically.
In this paper we work towards a general
methodology on how to apply statistical methods
automatically to analyse the exceptional patterns
which are hidden in business data, based on (Caron,
2012). We also consider the method to establish a
clear view of the business events taking place in and
across companies. The paper is organized as follows.
In Section 2, we examine the concepts of BI
supported business analytics and discuss a general
model for the methodology. Section 3 discuss the
challenge of constructing data view from event logs,
especially that arises from integrating data which are
shared among companies in supply chain networks.
The practical aspects of the application are discussed
in Section 4, and the last section concludes the paper.
530
Liu L., Daniels H. and Hofman W..
Detecting and Explaining Business Exceptions for Risk Assessment.
DOI: 10.5220/0004569905300535
In Proceedings of the 15th International Conference on Enterprise Information Systems (ICEIS-2013), pages 530-535
ISBN: 978-989-8565-59-4
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
2 BUSINESS ANALYTICS
OF EXCEPTIONS
Modern management system, such as ERP, records
business data in large volume, but overloaded
information poses a problem for human decision
maker, as it confounds him/her from realizing the
true status of the system, causal relationship between
exceptions, and the effect of treatment measures
(Milliken, 1987). To avoid this, reports are
generated by aggregating the data before presented
to the manager. When the manager is examining the
report, he/she is looking for extreme or unexpected
items and try to find explanations using analytics,
i.e. reversing the process of report generation,
drilling down in a managerial model, or using
additional knowledge possibly from external
sources.
The use of analytics in business can be roughly
grouped into two parts. First, descriptive analytics
captures the pattern of systematic emergence in the
company or the environment. The description usual-
ly supports prediction. Examples are the data mining
algorithms like clustering, classification and associa-
tion, applied to identify the events which can possi-
bly lead to disasters. Although descriptive analytics
does not presume any expectations, the analyst usu-
ally looks for “interesting” patterns when interpret-
ing the results. In this process, implicit background
knowledge is applied in searching for (mental) ex-
ceptions (Keil, 2006).
Secondly, diagnostic analytics reason about the
causal relations of those patterns. The goal for this
type of the analysis is to restore or verify the mecha-
nism of a sequence of events (Keil, 2006), e.g. the
operations in the company. The conclusion usually
leads to decisions for adjustment and improvement
of the system. Exemplary analysis questions are
“why the company performance is not as expected”
– for improving performance of the managed sys-
tem, and “why certain exceptions have not been
detected by current monitors” – for adjusting the
management system. Audit analytics also falls in
this category, analysing the risk of fraud and/or
unintentional errors in accounting systems
(Vasarhelyi et al., 2004); (Bay et al., 2006). In the
framework we propose (see Section 2.2), we gener-
alize and combine these two types to the detection
and the diagnosis phases in an integrated process of
business analytics.
We argue that business analytics is a strategic
important process of organizational learning that
extends the philosophy of "management by excep-
tion". The importance of analytics lies in the neces-
sity of "meta-control" to cope with the internal and
external changes. The management system of the
company (ERP) monitors and controls the business
processes, which deliver value to customers and
form competitive competence. It automates the rou-
tine tasks of detecting and treating operational ex-
ceptions, because the business knowledge are codi-
fied into the build-in controls of the system (in form
of business rules or constraints) in a “plan-do-check-
adjust” cycle. With automation, management sys-
tems can help with handling these routine tasks in
large volume data (big data), e.g. managing thou-
sands of accounts in finance and cost accounting
systems. However, their monitor-control capability
is limited to the codified rules, so they cannot deal
with the “new” changes or the exceptions out-of-
scope of the rules. These exceptions are left to the
responsibility of human managers. Though the
“new” exceptions are on a higher system level than
the management system ergo not directly visible,
they affect the performance of the managed system
(the company): therefore, they must be detectable by
analysing the data collected / generated by current
management system. The analysis results in new
business knowledge that equips the management
system for controlling similar exceptions in the
future. Ideally, the managers hope to continuously
meta-control the management system, automating
the process using BI (Vasarhelyi et al., 2004).
2.1 BI Supported Business Analytics
Business Intelligence is the collection of procedures
to reduce the volume of information that the manag-
er need to take into account when making decision.
The information-reduction is done by organising
(extract-transform-load, ETL) transactional data into
a multi-dimensional database (data warehouse or
OLAP), in which large volume of operational details
can be abstracted, aggregated or computed into
business reports, using BI techniques (see Figure 1).
This process involves both the managerial model
and the technical model of information organization.
On one hand, the organising of information is in
essence driven by managerial purpose, i.e. the man-
agerial model. For example, the accounting process,
which in general is a BI process, aggregates transac-
tion records in various documents such as journals,
general ledgers and financial statements for operat-
ing, financing and investing purposes respectively
(Bay et al., 2006). The organization of these docu-
ments codifies the managerial model. For instance,
the general ledger, recorded using double-entry
book-keeping, is a codified management system
DetectingandExplainingBusinessExceptionsforRiskAssessment
531
which internally controls balance between two ac-
counts involved in each transaction (Bay et al.,
2006).
On the other hand, the technical model organises
information for an analytical purpose. Organising
business data in the form of tables helps to highlight
contextual similarities among the data, providing
important support for the business analyst. For in-
stance, aligning records chronically, e.g. sales in
multiple periods, can show the temporal changes and
trends in the record set. As a special case, OLAP is a
useful tool to analyse multi-dimensional, hierar-
chical data interactively, with the standard drill-
down, roll-up and slice operations (Caron, 2012).
From an analytics viewpoint, the managerial model
provides an ontological structure of the information
(Hofman, 2013), while the technical model gives a
storage structure, also known as data structure in
computer science. Combining these two models
gives a data view of the business activities taking
place in the managed and the management systems.
We will come back to discuss the data view later in
Section 3.
2.2 A General Model for Business
Analytics
Before the analytic process can be automated, its
procedure should first be formalized. The lexical
definition of exception is “an instance that does not
conform to a rule or generalization” (thefreediction-
ary.com), which implies the comparison of the actu-
al instance to a norm. Our discussion on business
analytics is largely based on previous works of caus-
al analysis and explanations in (Caron and Daniels
2009); (Caron and Daniels, 2008); (Feelders and
Daniels, 2001); (Caron, 2012). The analysis of ex-
ceptions takes the canonical format of (Feelders and
Daniels, 2001):
,,
because
, despite

(1)
where
,,
is the triple for exception detection,
and the exception is to be explained by the non-
empty set of contributing causes
and the (possi-
bly empty) set of counteracting causes

. The di-
agnosis analysis is to explain why the instance
(e.g. the ABC-company) has property (e.g. having
a low profit) when the other members of reference
class (e.g. other companies in the same branch or
industry) do not.
The information structure of has the general
form of
, where 
,
,⋯,
is an n-
component vector. In words, certain property value
of which is important for decision making, denot-
ed by, is dependent on other property values in
the information structure of.
We can use the information structure to estimate
the norm value of, given the actual values of.
Exception-detection is done by studying the differ-
ence between the actual and the norm value of.

|
,
(2)
where ~
0,
. If the difference is significant,
i.e.
|
|
,
is viewed as a symptom to be ex-
plained. The user defined threshold parameter
depends on the application domain, and the estima-
tion method for
|
depends on both
and
the application. A more general form of (2) is

|
info

(3)
where info stands for all kind of information availa-
ble. For example, Alles et al. (2010) uses the infor-
mation of sales of prior period

to estimate the
profit of current period
. The symptom is ex-
plained by the influence of each
, and the influ-
ence is measured as
inf
,

,

,,
(4)
Figure 1: Business analytics supported by BI.
ICEIS2013-15thInternationalConferenceonEnterpriseInformationSystems
532
where1,2,⋯,, and

,
denotes the
value of
with all variables evaluated at their
norm values, except
.
For clarity, we distinguish the technical model
from the managerial model in the information struc-
ture. For example in OLAP (see equation system
(5)), the variables in a managerial model (shown as
the functional relation) can be organised into a
hierarchy by aggregation, such as summation or
average (shown as the functional relation). Verti-
cally, all variables in the managerial model are or-
ganised based on the same aggregation relation.
Given that, the variables on a specific level of ag-
gregation follow the same business relation, just as
those variables on other aggregation levels horizon-
tally do.
In (5), the variables and are organized in an
OLAP cube with dimensions. Each dimension has
a hierarchy of
levels, where1,2,⋯. In a
specific dimension, variables on the hierarchy
level
are aggregated from the elements in the
lower hierarchy level
1, and these elements
are denoted respectively as
and
, where
1,2,,. Here,
is an n-component vector,
whose components are denoted as
,
.


,
,⋯,

⋯


⋯

,
⋯
⋯
,⋯,
,
⋯
⋯
⋯
⋯


⋯


⋯
⋯
⋯


⋯


⋯
(5)
With the information structure available, we can
look at lower level of detail for explanation by drill-
ing down. For example, if there is a significant
symptom
in the OLAP model, detected
by



, we can drill down the
managerial model for explanations, using



. A necessary condition to obtain
sensible explanations by drilling down is consisten-
cy of the normative estimation, i.e.
y
|








(6)
This condition in relation with usually holds
for the OLAP model, but should be checked for
(statistical) managerial models in general. This issue
is studied in depth for ANOVA models in OLAP
databases (Caron, 2012).
3 ANALYTICS IN SUPPLY
CHAIN NETWORKS
The method for business analytics can be applied in
a company, a supply chain, or even supply chain
networks, since a supply chain system can also be
seen as “a big company”. This generalization is
relevant, as activities taking place in a company
influence, and can be influenced by those in other
companies in a supply chain. With this dependence,
the analytics of supply chain exceptions should in-
volve event logs shared by multiple companies.
In the supply chain context, risk analysis is per-
formed over the data which are shared during busi-
ness transactions between trading partners. Integrat-
ing these data to form a data view gives rise to the
challenges of interoperability (Hofman, 2013). Here
we limit the discussion to logistic services. Interop-
erability comprises three aspects that are closely
interrelated, namely 1) the logistic services resulting
in business transactions, 2) the semantics of shared
data, and 3) the choreography of business. The se-
mantics of data is a precondition for processing data
automatically. The choreography needs to be known
to derive the status (
) of physical processes and
business transactions which refer to logistic activi-
ties that are performed, e.g. transport of cargo con-
tainers. As such, these three aspects are part of the
managerial model relevant for monitoring supply
chain networks.
Under the assumption that companies share data
electronically, a data capture algorithm can crawl
these event logs regularly. And the data can be fused
to compose a supply chain view, organized in a
“business event store”. A condition is that all the
involved companies adhere to the same semantic
model. Transformations can be implemented in case
a company adheres to another semantic model than
agreed.
The business event store may contain duplicated
data for different business events, i.e. (almost) iden-
tical data can be stored for two or more business
events that are related to different companies. For
instance, two reports for a container may be stored,
referring to the delivery and the acceptance events of
the container. The data fusion component needs to
identify that these two reports are related, referring
DetectingandExplainingBusinessExceptionsforRiskAssessment
533
to the same business transaction involving the logis-
tics service provider and the cargo receiver.
The data fusion functionality has to mine the as-
sociation amongst the event logs by matching the
following properties of logistic service:
Business transaction identifier: e.g. a Unique Con-
signment Number assigned to each complete chain
of transportation
Sender/recipient: which construct the custom-
er/service provider relation for each transaction
Place and time: each business event associates to a
place and time, e.g. place and time of acceptance
and of delivery
Transaction hierarchy: this allows for decomposi-
tion of logistic activities, e.g. a journey of contain-
er transport may consist of several stretches of
transportation
4 PROCEDURE FOR ANALYTICS
Based on the discussion above, we can summarize a
general procedure for business analytics, with con-
sidering the practical methodology of data analysis
(Feelders and Daniels, 2000):
1. Define problem: define analysis goal and choose
the variable which is important for decision.
2. Establish context: abstract and explicitly specify
the information structure (or load from a
knowledge base, if available). The context is
usually connoted by the source of information
from which the business report was generated.
Sometimes external sources need to be included
to enlarge the context, depending on the analysis
goal.
3. Identify exceptions: choose appropriate reference
class, estimate the norm, and apply it to actual
data. Despite the wishes for fully automated
analysis, the derivation of the norm remains an
interactive process in which several practical as-
pects demand lots of background knowledge
from the analyst (see Section 4.1).
4. Generate explanations: relate the exceptions in
different parts of the business system and reason
about the causal relations, using equation (4).
Method for developing the relations has been
well studied in previous works (Caron and
Daniels, 2008); (Caron and Daniels, 2009), in-
cluding greedy and top-down explanation.
5. Interpret results: review the explanations. In case
the results does not sufficiently supports deci-
sion, repeat step 2 to 5.
4.1 Practical Aspects
The following two key tasks are the most intricate in
the process of business analysis:
1. How to find an appropriate normative model to
detect exceptions, and
2. How to find the real causes to explain the rela-
tionship between the exceptions.
4.1.1 Exploration: Finding an Appropriate
Norm
Business analysis is in any case an exploratory pro-
cess. The normative model plays a central role in
qualifying a feature as normal or exceptional. The
firstly used normative models to detect symptoms
are usually the codified business constraints in the
management system, such as plans or budgets. Pecu-
liarly, in the subsequent diagnostic analysis to ex-
plore a sensible explanation, the choice of the nor-
mative model for the lower level of analysis relies to
a large extent on the choice of the analysis context,
because the analysis goal is usually an open ques-
tion. For instance, a decrease in profit may due to
the drop in internal efficiency or the deteriorated
global economy.
In the exploration for the subsequent normative
models, statistics are usually applied to the analysis
context, i.e. the members of the reference class.
With a data driven (bottom-up) approach, the meth-
od for choosing a proper reference class can be “sof-
tening” the set of business constraints used in the
management system for a particular monitor, using
an un-slice operation.
Softening business constraint is a useful tech-
nique for analysis. The un-slice operation takes the
union of the data sets which correspond to different
parts of the system. It thus expands the analysis
scope, so that the patterns on a larger system scale
can be revealed. For example, in the time dimension,
the trend or fluctuation of a variable over time can
only be seen on a time period, but not at a time
point. Besides, expanding the scope by un-slicing is
in itself an attempt of exploration, for instance in
searching for those exceptions whose impact only
takes effect after a time lag (Alles et al., 2010). This
in general helps the analyst to involve extra data by
extending the current information structure: in any
case, one can always organize the information of the
analysis context into an OLAP-like structure, and
then start to expand.
The reference class is always defined by a set of
constraints. Reminding of the codified business
constraints in the first place, the exploration for an
ICEIS2013-15thInternationalConferenceonEnterpriseInformationSystems
534
appropriate reference class can be regards as a “me-
ta-control” process that diagnoses and reflects upon
the detective power of the current set of constraints,
performed by the analyst (see Section 2). The explo-
ration thus iteratively applies the detective and diag-
nostic processes on the design of the business analy-
sis method.
4.1.2 Validation: Finding the Real Cause
Correctness and relevance are two important criteria
for evaluating the explanation. The correctness of
the models in the information structure is a premise
for finding the real cause. If the model doesn’t cap-
ture the business correctly, the reference model
would be based on a false assumption, and it would
then be incapable even in explaining a normal effect.
As a result, the model will possibly raise many false
alarms.
The relevance concerns the usefulness of the ex-
planation for decision support. A counter-example is
the explanation presented at the wrong level of detail
(also pointed out in Keil, 2006). The method for the
evaluation of the correctness and relevance generally
rely on the background knowledge of the application
domain.
5 CONCLUSIONS
Current business databases contain massive amounts
of data that carry important explicit and implicit
information about the underlying business process.
In this paper we have shown how general statistical
methods can be applied to automatically detect im-
plicit patterns that are interesting to be investigated
further for risk assessment. In many cases the data
itself include enough information to discover unusu-
al patterns or trends to be explored further, like in an
OLAP database. The process of examination is driv-
en by accounting equations or drill-down equations
and can generate explanations supported by the data.
In the future we want to investigate the incorpora-
tion of heterogeneous external data sources to obtain
a richer structure for causal analysis as described in
this paper. A case study in risk management in
global supply chains is currently explored.
ACKNOWLEDGEMENTS
This work was supported by the EC FP7 project
CASSANDRA (Grant agreement no: 261795).
REFERENCES
Alles, M. G.., Kogan, A., Vasarhelyi, M. A. & Wu, J.
2010. Analytical Procedures for Continuous Data
Level Auditing: Continuity Equations. Available at:
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.
1.1.174.240 [Accessed February 3, 2013].
Bay, S., Kumaraswamy, K., Anderle, M. G., Kumar, R. &
Steier., D. M. 2006. Large scale detection of
irregularities in accounting data in Data Mining, 2006.
ICDM’06. Sixth International Conference on. IEEE,
pp. 75–86.
Caron, E. A. M. & Daniels, H. A. M. 2009. Business
Analysis in the OLAP Context in J. Cordeiro et al.
eds., ICEIS 2009. Milan, Italy, pp. 325–330.
Caron, E. A. M. 2012. Explanation of Exceptional Values
in Multi-dimensional Business Databases. Erasmus
University Rotterdam. Available at: http://www.
emielcaron.nl/papers/Thesis.pdf.
Caron, E. A. M. & Daniels, H. A. M. 2008. Explanation of
exceptional values in multi-dimensional business
databases. European Journal of Operational Research,
188(3), pp.884–897.
Davenport, T. 2006. Competing on analytics. harvard
business review, 84(1), pp.98–107.
Feelders, A. & Daniels, H. A. M. 2001. A general model
for automated business diagnosis. European Journal
of Operational Research, 130(3), pp.623–637.
Feelders, A. & Daniels, H. A. M. 2000. Methodological
and practical aspects of data mining. Information &
Management, 37(5), pp.271–281.
Hofman, W. 2013. Compliance management by business
event mining in supply chain networks in VMBO.
Delft, the Netherlands.
Keil, F. C. 2006. Explanation and understanding. Annual
review of psychology, 57, pp.227–54.
Lund, M.S., Solhaug, B. & Stølen, K. 2011. Model-driven
risk analysis: the CORAS approach. Springer-Verlag
Berlin Heidelberg.
Milliken, F. J. 1987. Three Types of Perceived
Uncertainty about the Environment: State, Effect, and
Response Uncertainty. The Academy of Management
Review, 12(1), p.133.
Sodhi, M. S. & Tang, C. S. 2009. Managing Supply Chain
Disruptions via Time-Based Risk Management in T.
Wu et al. eds., Managing Supply Chain Risk and
Vulnerability. London: Springer, pp. 29–40.
Vasarhelyi, M. A., Alles, M. G. & Kogan, A. 2004.
Principles of analytic monitoring for continuous
assurance. Journal of Emerging Technologies in
Accounting, 1(1), pp.1–21.
DetectingandExplainingBusinessExceptionsforRiskAssessment
535