Using Inﬂuence Nets in Financial Informatics: A Case Study ...€¦ · Informatics: A Case Study of Pakistan Sajjad Haider1, Muhammad Shafqat2 and Shabih Haider3 1 Faculty of Computer

Journal of Computing and Information Technology - CIT 20, 2012, 2, 113–124doi:10.2498/cit.1001998

113

Using Influence Nets in FinancialInformatics: A Case Study of Pakistan

Sajjad Haider1, Muhammad Shafqat2 and Shabih Haider3

1 Faculty of Computer Science, Institute of Business Administration, Karachi, Pakistan2 UNESCO, Islamabad, Pakistan∗3 Department of Economics and Finance, Institute of Business Administration, Karachi, Pakistan

The paper presents an application of Influence Nets(INs) in the field of financial informatics. Influence Netshave primarily been used in war games to model effectsbased operations but, as shown in this paper, they canprove to be equally useful in other domains requiringdecision making under uncertain situations. The primaryadvantage of INs lies in their ability to acquire knowledgefrom subject matter experts in problem domains that relyheavily on experts’ opinion. A sample case study fromthe fields of economics and finance is presented in thispaper. The case study models the choices faced by a de-veloping country to recover her economy which is goingthrough a difficult phase due to global financial crisis,internal law and order situation and political instability.

Keywords: influence nets, Bayesian networks, AI infinance, causal maps, financial informatics

1. Introduction

Bayesian belief networks [1] have become thetool of choice for reasoning under uncertainty.BNs and their variants, such as Influence di-agrams, Influence Nets and Bayesian CausalMaps, have been applied in many fields includ-ing fault diagnosis, information fusion, foren-sics, marketing, medical sciences and financialinformatics, etc. [2-5]. Mathematically, a BN isa directed acyclic graph (DAG) where nodes inthe graph represent random variables while arcsbetween pairs of nodes represent certain con-ditional independence assumptions. The DAGdefines a factorization of the joint probabilitydistribution of the modeled random variable.This joint distribution is obtained as a product

of all conditional probabilities specified for eachvariable given its parent in the DAG. Formally,

P(X1, X2, . . . , XN) =N∏

i=1

P(Xi | pa(Xi)).

The main purpose of building a BN is to doprobabilistic inference which involves comput-ing the posterior probabilities of the variable(s)of interest after getting evidence about certainvariables. The process is also referred to as be-lief updating. Several exact and approximatealgorithms have been developed that exploit thestructure of a BN to do belief updating in anefficient manner [6-10]. It must be stated, how-ever, that, in general, the belief updating ina multiply-connected BN (where at least twonodes are connected through multiple paths) isNP-Hard [11].

When sufficient data is available, the structureand parameters of a BN can be learned fromit. However, if the data is not available, thenit becomes the job of a knowledge engineer,together with subject matter experts, to com-pletely specify a BN. Most of the complex un-precedented situations belong to this category.The BN specification process involves build-ing the structure of the network as well as fill-ing the corresponding conditional probabilitytables. One of the major limitations of BNs isthe acquisition of these conditional probabili-ties. They grow exponentially with the numberof parents of a node, thus making it impossi-ble for a subject matter expert to specify them.

∗Muhammad Shafqat was studying at the Faculty of Computer Science, Institute of Business Administration, Karachi when thework was done.

114 Using Influence Nets in Financial Informatics: A Case Study of Pakistan

Several schemes have been suggested to over-come this limitation. The list includes Noisy-Or [12], the CAST logic [13,14] and some otherextensions of Noisy-Or (including GeneralizedNoisy-Or [15] and Recursive Noisy-Or [16]).Such schemes ask for a non-exponential (pri-marily linear) number of parameters from a sub-ject matter expert and then transform these pa-rameters into conditional probabilities requiredfor the full specification of a BN.

The special instance of BNs which employs theCAST logic for knowledge elicitation is referredto as Influence Nets (INs). The CAST logicbased interface of INs allows a subject matterexpert to model both positive and negative im-pacts of an event on other events, again with alimited number of parameters. Unlike Noisy-Or, which asks an expert to specify the impactof an event on its child event, the CAST logicallows a subject matter expert to specify theimpact of both the presence and absence of anevent. In addition, instead of using a proba-bilistic scale of [0, 1], it allows an expert to usea non-probabilistic scale between −1 and 1 tospecify impacts of the presence/absence of anevent on its child event. This ability to captureboth positive and negative impacts makes INsbetter equipped to model problem domains thatrely heavily on experts’ knowledge.

Two other techniques that also aim to cap-ture positive and negative impacts are BayesianCausal Maps (BCM) [17,18] and QualitativeBayesian Networks (QBN) [19]. INs, however,offer a more expressive knowledge acquisitionmechanism than its counterparts. Both BCMand QBN model a positive or negative impactof the presence of an event (i.e., the event be-ing in the “True” state) on its child event, butdo not model the impact of the absence of suchevent (i.e., the event being in the “False” state)on its child. There is no such limitation in theINs. Furthermore, QBN, as the name suggests,only allows the modeling of qualitative posi-tive or negative strengths. INs, on the otherhand, provide a mechanism to transform posi-tive and negative impacts into conditional prob-ability tables. Similarly, BCM employs quali-tative positive and negative strengths during theearly stages of model building phase when ituses causal maps to build the initial skeleton ofthe BCM. However, when it comes to the spec-ification of conditional probabilities, BCM suf-

fers from the same knowledge elicitation limita-tions as BNs because it requires an exponentialnumber of conditional probability values.

INs and their extensions, Timed Influence Nets(TINs) [20] andDynamic InfluenceNets [21,22],have been extensively used in the field of effects-based operations. They are used to model andevaluate alternative courses of action and theireffectiveness to a mission’s objectives [23-27].Their scope, however, is not just limited to mil-itary and political analysis and they can proveto be equally useful in other domains requiringcomplex decision making. This paper presentsan application of INs in the field of financialinformatics. The paper shows how an IN can beused to model the choices faced by a develop-ing country like Pakistan which, in recent years,is going through a serious economic crisis dueto many factors including deteriorating law andorder situation, political instability, high infla-tion and many global factors. It is also shownhow techniques such as sensitivity analysis andsets of actions finder (SAF) algorithm can helpa decision maker in understanding the impactof unprecedented events on the country’s econ-omy.

The rest of the paper is organized as follows.Section 2 provides an overview of InfluenceNets, the CAST logic, sensitivity analysis andthe SAF algorithm. Section 3 discusses thecase study and explains how Influence Nets areused in financial informatics. Finally, Section4 concludes the paper and provides the futureresearch directions.

2. Influence Nets

Influence Nets are Directed Acyclic Graphs(DAGs) where nodes in the graph represent ran-dom variables, while the edges between pairsof variables represent causal relationships. Themodeling of the causal relationships is accom-plished by creating a series of cause and ef-fect relationships between variables represent-ing desired effect(s) and variables representingset of actionable events. Typically, the action-able events are drawn as root nodes (nodeswith-out incoming edges), while the desired effect ismodeled as a leaf node (node without outgoingedges). Influence Nets require a system mod-eler (or a subject matter expert) to specify the

Using Influence Nets in Financial Informatics: A Case Study of Pakistan 115

P(A| E, B) = 0.005 P(A| E, B) = 0.950 P(A| E, B) = 0.950 P(A| E, B) = 0.990

P(D| E, A) = 0.050 P(D| E, A) = 0.950 P(D| E, A) = 0.001 P(D| E, A) = 0.050

P(E) = 0.01

P(B) = 0.05 D A B

E

Figure 1. A sample influence net.

CAST logic parameters instead of the probabil-ities. The required probabilities are internallygenerated by the CAST logic algorithm with thehelp of user-defined parameters. The followingitems characterize an IN.

1. A set of random variables that makes up thenodes of an IN. All the variables in the INhave binary states.

2. A set of directed links that connect pairs ofnodes.

3. Each link is associated with a pair of CASTlogic parameters that show the causal strengthof the link (usually denoted as h and g val-ues).

4. Each non-root node has an associated CASTlogic parameter (denoted as the baselineprobability), while a prior probability is as-sociated with each root node.

Figure 1 shows an example of an Influence Net.The directed edge with an arrowhead betweentwo nodes shows the parent node promoting thechances of a child node being true, while theroundhead edge shows the parent node inhibit-ing the chances of a child node being true. Thetext associated with the non-root nodes repre-sents the corresponding conditional probabilityvalues obtained from the CAST logic param-eters (not shown in the figure) while the textassociated with the root nodes represents theprior probabilities.

2.1. CAST Logic

The specification of a Bayesian Network re-quires an exponential number of parameters formodel specification. As a model grows larger,this requirement presents a very big challengeto a system modeler. As an attempt to overcomethis limitation, Chang et al. [14] developed a for-malism called CAusal STrength (CAST) logicto elicit the large number of conditional proba-bilities from a small set of user-defined param-eters. The logic has its roots in the Noisy-Orapproach. The logic requires only a pair of pa-rameter values for each dependency relationshipbetween any two random variables. The valuesare converted into conditional probability tablesand the resulting tables are used during the be-lief updating phase. Thus, Influence Nets couldbe regarded as a special instance of BayesianNetworks. A brief explanation of the CASTlogic is provided below, with the help of an ex-ample shown in Figure 2. Readers interested ina detailed description of the CAST logic shouldrefer to [13,14].

h= +0.9, g= -0.7

h= -0.4, g= +0.8

h= +0.9, g= -0.5

b=0.3

B

A

C

X

Figure 2. An influence network with CAST logicparameters.


Figure 2 contains four nodes A, B, C and X.On each arc, two causal strengths are specified.These numbers represent the probability that aspecified state of a parent node will cause acertain state in the child node. Positive valueson arcs are causal influences that cause a nodeto occur with some probability, while negativevalues are influences that cause the negationof a node to occur with some probability. Forinstance, the arc between B and X has values−0.4 and 0.8. The first value, referred to ash, states that if B is true, then this will causeX to be false with probability 0.4, while thesecond value, referred to as g, states that if Bis false, then this will cause X to be true withprobability 0.8. Both h and g can take values inthe interval (−1, 1). All non-root nodes are as-signed a baseline probability, which is similar tothe “leak” probability in theNoisy-Or approach.This probability is the user-assigned assessmentthat the event would occur independently of themodeled influences in a net.

There are four major steps in the CAST logicalgorithm that converts the user-defined param-eters into conditional probabilities:

a) Aggregate positive causal strengths

b) Aggregate negative causal strengths

c) Combine the positive and negative causalstrengths, and

d) Derive conditional probabilities

In Figure 2, there are eight conditional proba-bilities that need to be computed to obtain themarginal probability of X. Mathematically, themarginal probability of X is computed as:

P(X) =P(X | ¬A,¬B,¬C)P(¬A,¬B,¬C)+ P(X | ¬A,¬B, C)P(¬A,¬B, C)+ P(X | ¬A, B,¬C)P(¬A, B,¬C)+ P(X | ¬A, B, C)P(¬A, B, C)+ P(X | A,¬B,¬C)P(A,¬B,¬C)+ P(X | A,¬B, C)P(¬A,¬B, C)+ P(X | A, B,¬C)P(A, B,¬C)+ P(X | A, B, C)P(A, B, C)

(1)

The four steps, described above, are used to cal-culate each of these eight conditional probabil-ities. For instance, to calculate the probabilityP(X | A, B,¬C), the h values on the arcs con-necting A and B to X and the g value on the arc

connecting C to X are considered. Hence, theset of causal strengths is {0.9,−0.4,−0.5}.Aggregate the Positive Causal Strengths: Inthis step, the set of causal strengthswith positiveinfluence are combined. They are aggregatedusing the equation

PI = 1 −∏

i

(1 − Ci) ∀Ci > 0

where Ci is the corresponding g or h value hav-ing positive influence and PI is the combinedpositive causal strength. For our example

PI = 1 − (1 − 0.9) = 0.9

Aggregate the Negative Causal Strengths: Inthis step, the causal strengths with negative val-ues are combined. The equation used for aggre-gation is

NI = 1 −∏

i

(1 − |Ci|) ∀Ci < 0

where Ci is the corresponding g or h value hav-ing negative influence and NI is the combinednegative causal strength. Using the above equa-tion, the aggregate negative influence is foundto be:

NI = 1 − (1 − 0.4)(1 − 0.5) = 0.7

Combine Positive and Negative Causal Stre-ngths: In this step, aggregated positive andnegative influences are combined to obtain anoverall net influence. The difference of theseaggregated influences is taken. The overall in-fluence is obtained by taking the ratio of thisdifference and the corresponding promoting orinhibiting influence. Mathematically,If PI ≥ NI

AI =PI − NI1 − NI

.

If NI > PI

AI =NI − PI1 − PI

.

Thus, the overall influence for the current ex-ample is

AI = (0.9 − 0.7)/(1 − 0.7) = 0.66

Derive Conditional Probabilities: In the finalstep, the overall influence is used to compute


the conditional probability value of a child forthe given combination of parents.

P(child | j-th state of parent states) =

=

⎧⎪⎪⎨⎪⎪⎩

baseline + (1 − baseline) · AIwhen PI ≥ NI

baseline − baseline · AIwhen PI < NI

Using the above equation, P(X | A, B,¬C) isobtained as:

P(X | A, B,¬C) = 0.3 + 0.7 · 0.66 = .762

The steps explained above are repeated for theremaining seven conditional probabilities inEquation 1. It should be noted that, if theexperts had sufficient time and knowledge ofthe influences, then the conditional probabil-ity table for each node can be used instead ofg and h values. Furthermore, after estimatingthe conditional probability table, if some entriesdo not satisfy the expert, then those entries canbe modified and then used for computing themarginal probability of a node.

2.2. Tools to Gauge Impact of Changes inINs

Once an IN model is developed, it can be fur-ther refined and analyzed using sensitivity anal-ysis and sets of actions finder (SAF) algorithm.The sensitivity of action analysis looks at howsensitive an effect is with respect to the action-able events when actions are considered one ata time [28]. The analysis requires linear num-ber of searches, i.e., if there are n actionableevents, then n iterations are required to performthe analysis. The algorithm is presented in Ta-ble 1.

Another approach to gauge the impact of ac-tionable events on the desired effect is the setsof actions finder (SAF) algorithm. The SAFalgorithm [28,29] is a heuristic approach to de-termine the sets of actions (and their configu-ration/state) that cause the probability of thedesired effect to be above (below) a given prob-ability threshold. The algorithm achieves thistask in significantly less time than is requiredfor an exhaustive examination of the actions’search space, which is exponential in terms of

Given A, Ewhere A = Set of Actionable Events

E = Desired Effect

Iterate ∀ I where I ∈ ASet P(I Original) = P(I) //store the original

probability of the desired effectSet P(I) = 0Compute P(E). Set P(E min) = P(E)Set P(I) = 1Compute P(E). Set P(E1) = P(E)Set Diff(I) = P(E1) a P(E0) //store the difference

between the two probabilitiesSet P(I) = P(I Original) //restore the original

probability of I.

Table 1. Sensitivity of action analysis algorithm.

Given A, E, S, twhere A = Set of Actions

S = Set of Selected ActionsE = Desired Effectt = Threshold

1. Initialize S = null.2. Iterate ∀ I where x ∈ A.

Set P(I) = 0 //set the prior probability ofall actionable events to zero

3. Compute P(E). Set P(E start) = P(E).4. Iterate ∀ I where I ∈ A

Set P(I) = 1 where I ∈ AIterate ∀ J where J ∈ A \ {I}

Set P(J) = 0Compute P(E)Set Diff(I) = P (E) - P(E start)

5. Select the highest Diff(I) obtained above.If Diff(I) > 0 OR if the corresponding P(E) > t

Remove I from AInsert I into SSet P(E start) = P(E)Go to Step 5

Else Stop

Table 2. The SAF algorithm.

the number of actions. The algorithm runs inpolynomial time and uses a greedy approach toidentify the best (or close-to-best) sets of ac-tions. The algorithm is presented in Table 2.

3. Case Study

This section models and analyzes a financialdecision making situation using Influence Nets.Unlike traditionalmodeling approaches that relyon empirical data, Influence Nets allow a know-ledge engineer to capture experts’ belief regard-ing the impact of unprecedented and unfold-


ing events on the variable(s) of interests; andthus are better suited to model problem domainsthat rely heavily on subject matter experts’ be-lief. These unprecedented events could includetsunami, earthquake, terrorist activities, foodshortage, unusual hike in gold or oil price, etc.

The paper focuses on the economic/financialsituation of Pakistan during the past few years.Being a growing economy, Pakistan has a hugepotential in all areas, be it oil exploration, ser-vices sector, agriculture, industry, etc. Keepingthis in view and also due to favorable politicaland financial situation between 2002 and 2007,Pakistan attracted a lot of foreign investmentand saw a huge injection of funds from domesticinvestors. As a result, Karachi Stock Exchange,the biggest stock market of Pakistan, boomedfrom 1200 market index to 16000 market index.Starting from the first quarter of 2007, however,this upward trend started moving in the oppo-site direction. The primary factors were globalrecession, deteriorating law and order situation,increase in terrorism, and political instability.These factors contributed to the stoppage of for-eign investment and flight of capital. By 2008,the crisis reached its maximum and there was adanger that Pakistan could default as a country.

The Influence Net discussed in the sequel aimsto structure the external/internal factors that im-pact, either positively or negatively, the stateof Pakistani economy. The goal of this modelbuilding exercise is: (a) to integrate discreteeconomic indicators, (b) to illustrate the causeand effect relationships that exist among them,and (c) to present the use of this Influence Netsbased model as a decision support tool in struc-turing the problem and making the right deci-sion. Influence Net in this context coalesce theimpact of individual variables onto subsequentnodes, thus ultimately building consequentialoutcome for the overall economy. In addition,by applying sensitivity analysis and the SAF al-gorithm, a knowledge engineer/decision makercan determine variables that have the greatestimpact on the economy and in turn could aimto manipulate these variables to achieve the de-sired effect. It must be clearly stated that thefocus of this paper is not to completely and ac-curately model all the relevant issues, but tohighlight the advantages of IN based model-ing technique using a reasonably valid (thoughhypothetical at times) picture of the selected

scenario. In addition to sensitivity analysis andthe SAF algorithm, the built IN model is alsoanalyzed by having two different sets of priorprobabilities for its root nodes (depicting ac-tionable/observable events). These sets of priorprobabilities reflect the chances of occurrenceof the corresponding events in 2007/2008 and2010 periods. Fewof the actionable/observableevents that impact the overall economy throughchains of cause and effect relationships are de-scribed below.

External and Internal debt: Pakistan’s econ-omy has been mired with growing external andinternal debt that aggravated into crisis-like-situation after the global financial crisis. Rise ingeneral imports, flight of capital, widening tradeand budget deficits, foreign debt servicing alongwith huge unsustainable subsidies in oil sectorbecame the trigger for Pakistan’s economic col-lapse. The government though gradually beganpassing on this pressure to people, but still it wasnot enough. Resultantly, all the gains of reliev-ing economy from debt since 2002 reversed by2008. Pakistan is still in the process of receiv-ing foreign funding just to recover its economy,but things would get worse when the additionaldebt servicing starts of the newly “earned” debt.

IMF/WB-FOP Assistance: As discussedearlier that by early 2008, Pakistan camecloser to getting default. It was ultimatelyoffered loan (with strict conditions) by IMF,World Bank and Friends of Pakistan (FOP)consortium. The flow of assistance moneyfrom these sources helped it in boosting itsForeign Exchange reserves and stabilizingexchange rate.

Interest Rates: The central bank stabilizedthe market interest rates to contain inflation-ary pressure. The new policy of containingthe consumption spree was in contrast to theexpansionary policy of the previous sevenyears (till 2007) which was meant to demandpressure on import bills. Such mechanismhelped in stabilizing exchange rate and For-eign Exchange reserves.

Commodity and Oil prices: There had beena sharp ‘extraordinary’ spike in oil and com-modity prices in 2008 that virtually madeeconomies collapse. Even though the pricesfell after an abrupt spike, but the damage donein that relatively small period had a lasting im-pact on the economic gains of the past years.


Remittances: Remittances by expats to Pak-istan rose substantially since the start of globalfinancial crisis. It is primarily triggered bymassive layoffs by foreign companies thatprompted capital to flow into home country.Such foreign flow of money helped in stop-ping the complete devaluation of Pakistan’scurrency which was triggered by the flight ofcapital.

Debt Servicing: Pakistan’s substantial rev-enue goes to national debt servicing that hasbeen on the rise for many years. Debt resche-duling and additional loans taken over thepast decade are getting matured along withadditional short terms debt. Such measureshave a very negative impact on the foreignreserves and the exchange rate.

TightMonetaryPolicy and IncreasedTaxes:In an attempt to reduce country’s hunger forever increasing imports and to encourage sav-ings, the government pursued constrictingmonetary policy. Taxes had also been in-creased to close growing void between gov-ernment’s collection and spending.

Inflation and Cost of Production: Both fac-tors were on high as a result of increasing de-mand and supply gap which was sparked byheightened oil prices that also increased elec-tricity and transportation costs. These two,being at the core of any production, furtherincreased every day commodity prices, thusmaking them expensive for people to afford.

Political instability and security: The reper-cussion of being a front line state in theglobal war on terror and having a fragiledemocratic political setup not only scaredaway prospective foreign investment, but alsodid not help in holding previous investments.This severely affected country’sGDP growth.

The variables listed above are drawn as rootnodes in the Influence Net. They impact (ei-ther positively or negatively) other intermediateevents and ultimately all chains of cause andeffect converge to a single desired effect “Im-provement in Pakistani Economy”. The com-plete model is shown in Figure 3. There are 17root nodes which depict actionable (or poten-tially observable) variables. These root nodesimpact 12 intermediate variables (not directly

Figure 3. Influence net model of Pakistani economy.


observables) and finally all of them directly orindirectly have an impact on the overall state ofthe Pakistani economy. The strength of positiveor negative impact of the presence and absenceof one variable on the other is quantified throughthe CAST logic. For instance, the impact of ‘HiCntrl Bank Interest Rate’ is much slower to sur-face on country’s ‘exchange rate’ than the effectof ‘WB or IMF assistance’ that rescues deplet-ing rates more briskly.

Once the influence net is completely built, itsapplication lies in gauging the impact of ac-tionable events by varying probabilities. In thisstudy, two different sets of prior probabilitiesare considered. These sets of prior probabilitiesmodel the chances of occurrence of certain ob-servable variables (root nodes) in 2007/2008and 2010 periods. For instance, Pakistan hashistorically been relying heavily on assistancefrom World Bank and IMF that deemed nec-essary for bailing its ailing economy from dol-drums. In 2007/2008, however, the economicconditions deteriorated when its plea for finan-cial assistance was turned down and chanceswere not very bright of getting further aid. Thesituation is depictedwith a very lowprior proba-bility (0.1) for the event “IMF/WB FOP Assis-tance”. The situation, however, changed oncethe democratically elected government cameinto power, and since then, Pakistan has beenreceiving constant stream of financial assistancethat greatly improved its deteriorating economichealth. Thus, in 2010, the prior probabilityof “IMF/WB FOP Assistance” is quite high(0.9). The probability is not set to 1 as there isalways some element of uncertainty regardingthe next installment if Pakistan fails to follow

up on her promise to implement certain taxesand to reduce subsidies. Similarly, in 2007,international commodity prices reached a newheight and broke all the previous records. Samewas the case with oil prices which peaked toall time high. This causes the prior probabil-ities of these events to be very high (0.9) in2007/2008. The situation is not so extreme in2010 and prices have lowered down consider-ably from their peak values. This results in alower prior probability (0.3/0.4) for both vari-ables.

Furthermore, due to emergency declaration inPakistan which resulted in ban on judiciary andelectronic media and, later on, due to the as-sassination of one of the most popular politicalfigures, Pakistan faced an extremely high polit-ical instability in 2007/2008. In addition, thelaw and order situation was not very good dueto the fallout of global war against terror, whichresulted in many terrorist activities within thecountry. Thus, the corresponding events in theInfluence Net, Political Instability and Deterio-rating Law and Order Situation have a very highprior probability (0.9/0.85). The situation in2010 is slightly better as, despite many hiccups,the democratically elected government has beenable to survive for more than 2 years. In addi-tion, the army led operation against terrorists’hideout has marginally improved the law andorder situation within the country. This resultsin slightly lower prior probabilities for both ofthese events in 2010 (0.3/0.6). Table 3 listsall those factors which have significantly dif-ferent prior probabilities during 2007/2008 and

Variables 2007/2008 2010

IMF/WB FOP Assistance 0.1 0.9Hi External Debt 0.9 0.4Incr Capital Flight 0.9 0.5Incr Intl Commodity Price 0.9 0.4Incr Oil Price 0.9 0.3Political Instability 0.9 0.3Deter. Law & Order situation 0.85 0.6Hi Inflation 0.8 0.4

Improvement in Pakistani Economy 0.08 0.17

Table 3. Different sets of prior probabilities for 2007/2008 and 2010 periods.


2010 periods. Many events such as “Hi InternalDebt”, “Tight Monetary Control”, “Increase inTaxes”, etc are not listed as their prior prob-abilities have not changed significantly from2007/2008 to 2010. The table also lists the cor-responding marginal probability of the desiredeffect “Improvement in Pakistani Economy”. Itcan be seen that during the 2007/2008 period,the chances of improvement in Pakistani econ-omy were only 8%. This number, however, hasreached 17% in 2010, primarily due to financialassistance from IMF, World Bank and Friendsof Pakistan consortium, running the countryby a democratically elected government, anda substantial decrease in international commod-ity and oil prices. The value is still not veryhigh as the country is going through a seriouslaw and order situation and an extreme energycrisis which is resulting in lack of foreign/localinvestment and flight of capital.

Another way to analyze the impact of certainvariables is through sensitivity analysis. Asbriefly discussed in Section 2.2, sensitivity analy-

sis selects each node from the set of actionablenodes in an iterative manner. Its probability isfirst set to zero and then to one, while keepingthe probabilities of other actionable events totheir initial prior values. The sensitivity of theeffect with respect to the selected action is de-termined by computing the probabilities of thedesired effect when the action’s probability isset to zero and one, respectively. At the endof this iterative process, the events which causethe highest difference in the likelihood of theoccurrence of the desired effect are consideredfor further analysis. When sensitivity analysisis run on the Influence Net under considera-tion, the following five variables are identifiedas having the most significant individual impacton the state of Pakistani economy.

Hi Production Cost, Tight Monetary Policy, De-teriorating Law & Order Situation, Hi ExternalDebt, Hi Raw Material Prices

The sensitivity analysis is good to analyze theimpact of individual events,but its outcome can-not be generalized to estimate collective impact

Actionable Events State of Actionable VariablesSet 1 Set 2 Set 3 Set 4

IMF/WB FOP Assistance T T T THi Current Acct Deficit F F F FHi Cntrl Bank Interest Rate T T T THi Internal Debt F F F FHi External Debt F F F FIncr. in Intl Commodity Price F F F FIncr. in Oil Price F F F FIncr. in Remittance T T T TIncr. in Capital Flight F F F FDebt Servicing F T F TTight Monetary Policy F F F FIncr. in Taxes F F F FHi Inflation F F T THi Raw Material Price F F F FHi Production Cost F F F FPolitical Instability F F F FDeteriorating Law & Order Situation F F F F

Improvement in Pakistani Economy 0.799 0.798 0.795 0.793

Table 4. Results of the SAF algorithm.


due to non-synergistic and inhibitive impact ofdifferent variables, when taken together. Thusan actionable event, when considered alone,might have a positive impact on the desired ef-fect, but the same variable might impact nega-tively when taken in synchronization with othervariables. The collective impact of a set ofactionable events is analyzed through the SAFalgorithm, as briefly explained in Section 2.2.Primarily designed to support Influence Netsbased analysis in effects based operations, thealgorithm provides a mechanism to identify thebest configuration(s) of actionable events thatmaximize the chances of achieving the desiredeffect. It starts with a single actionable eventwhich, when considered individually, causes thehighest increase (for a maximization problem)in the probability of the desired effect beingtrue. This is followed by the selection of a sec-ond action from the remaining set of actions thattogether with the first action cause the highestincrease in the probability of the desired effect.Other actions are added iteratively in a similarmanner. The process stops at a point where(i) the inclusion of an action causes the proba-bility of the objective node to decrease and tofall below a given probability threshold or (ii)there are no more actions to add. Once alter-native sets of actions are obtained, they can begrouped together to form more general sets ofactions.

When the SAF algorithm is run on InfluenceNet of Pakistani economy, it suggests 4 con-figurations of the set of actionable events thatcauses up to 80% chance of improvement inPakistani economy. The actionable events andtheir configurations are listed in Table 4. Acursory look at these sets suggests that the con-figurations also match with the intuition onecould develop by analyzing the model. All theactionable events have identical configurations(either true or false) except two events: “DebtServicing” and “Hi Inflation” which have bothtrue and false values. “IMF/WB and FOP As-sistance” is true in all the solutions. Same istrue for “Hi Remittance” and “Hi Cntrl BankInterest Rate”. Similarly variables like “Politi-cal Instability”, “Deteriorating Law and OrderSituation”, “Hi External Debt”, “Hi Intrl Com-modity Price”, etc are false in all the solutions.As discussed above and explained in .[28], these4 solutions can be further generalized to forman aggregated solution.

4. Conclusions

The paper presents an application of InfluenceNets in financial informatics. Influence Netshave been extensively used in modeling effectsbased operations, but, as shown in this paper,they can be equally useful in other domains thatrely extensively on subjective knowledge. Theprimary advantage of Influence Nets lies in theircapability to model both positive and negativeimpact of the presence/absence of an event onother events and that too with very limited num-ber of parameters. They are also good for mod-eling unprecedented events which may not beadequately handled by empirical data based tra-ditional approaches. Together with sensitivityanalysis and the sets of actions finder algorithm,they can aid a decision maker/knowledge engi-neer in identifying the best sets of actions (andtheir corresponding state) which could maxi-mize the chances of achieving the desired effect.The selected case study: (a) modeled the situa-tion faced by a developing country like Pakistanas its economy suffered from global recession,political instability, deteriorating law and ordersituation during the past few years, and (b) sug-gested ways to overcome the burgeoning eco-nomic crisis. As mentioned earlier, the purposeof this study was not to accurately model all theeconomic factors, but to highlight the impor-tance of Influence Nets based modeling tech-nique. Nevertheless, the analysis and resultspresented in this paper are quite intuitive andthe approach can easily be applied/generalizedto other countries and/or situations with an en-hanced set of variables.

Acknowledgment

The authors would like to thank the Editor andanonymous reviewers for their valuable com-ments and feedback, which were helpful in im-proving the paper.

References

[1] J. PEARL, Probabilistic Reasoning in IntelligentSystems: Networks of Plausible Inference, MorganKaufmann, 1988.


[2] A. MITTAL, Bayesian Network Technologies: Ap-plications and Graphical Models, IGI Publishing,2007.

[3] O. POURRET, P. NADM AND B. MARCOT, BayesianNetworks: A Practical Guide to Applications, Wi-ley, 2008.

[4] F. TARONI, C. AITKEN, P. GARBOLINO AND A. BIE-DERMANN, Bayesian Networks and ProbabilisticInference in Forensic Science, Wiley, 2006.

[5] R. E. NEAPOLITAN AND X. JIANG, ProbabilisticMethods for Financial and Marketing Informat-ics, Morgan Kaufmann, 2007.

[6] J. E. CANO, L. D. HERNANDEZ AND S. MORAL, Im-portance sampling algorithms for the propagationof probabilities in belief networks, Int. J. Approx.Reasoning, Vol. 15, 1996, pp. 77–92.

[7] J. CHENG AND M. J. DRUZDZEL, AIS-BN: An Adap-tive Importance Sampling Algorithm for EvidentialReasoning in Large Bayesian Networks, Journal ofArtificial Intelligence Research, Vol. 13, 2000, pp.13–155.

[8] R. DECHTER, Bucket elimination: a unifying frame-work for reasoning, Artif. Intell., Vol. 113, 1999,pp. 41–85.

[9] S. L. LAURITZEN AND D. J. SPIEGELHALTER, Lo-cal computations with probabilities on graphicalstructures and their application to expert systems,Readings in uncertain reasoning, Morgan Kauf-mann Publishers Inc., 1990, pp. 415–448.

[10] D. POOLE AND N. L. ZHANG, A simple approachto Bayesian network computations, Proceedingsof 10th Canadian Conference on Artificial Intelli-gence, pp. 171–178.

[11] G. COOPER, The computational complexity of prob-abilistic inference using bayesian belief networks,Artificial Intelligence, Vol. 42, Mar. 1990, pp. 405,393.

[12] J. M. AGOSTA, Conditional inter-causally indepen-dent node distributions, a property of noisy-ormodels, Proceedings of the 7th Conference (1991)on Uncertainty in Artificial Intelligence, Los Ange-les, California, United States: Morgan KaufmannPublishers Inc., 1991, pp. 9–16.

[13] J. A. ROSEN AND W. L. SMITH, Influence Net Mod-eling with Causal Strengths: An Evolutionary Ap-proach, Command Control Research TechnologySumposium, Monterary, CA: 1996, pp. 699–708.

[14] K. C. CHANG, P. E. LEHNER, A. H. LEVIS, A. K.ZAIDI AND X. ZHAO, On Causal Influence Logic,George Mason University, 1994.

[15] F. DIEZ, Parameter adjustment in Bayes networks:The generalized noisy or-gate, Proceedings of the9th Conference on Uncertainty in Artificial Intelli-gence, 1993, pp. 99–105.

[16] J. LEMMER AND D. GOSSINK, Recursive noisy OR-a rule for estimating complex probabilistic inter-actions, IEEE Transactions on Systems, Man, andCybernetics, Part B: Cybernetics, Vol. 34, 2004, pp.2261, 2252.

[17] S. NADKARNI AND P. P. SHENOY, A Bayesian net-work approach to making inferences in causalmaps, European Journal of Operational Research,Vol. 128, Feb. 2001, pp. 479–498.

[18] S. O. SAHIN, F. ULENGIN AND B. ULENGIN, ABayesian causal map for inflation analysis: Thecase of Turkey, European Journal of OperationalResearch, Vol. 175, Dec. 2006, pp. 1268–1284.

[19] M. J. DRUZDZEL AND M. HENRION, Efficient Rea-soning in Qualitative Probabilistic Networks, Pro-ceedings of the 11th National Conference on Artifi-cial Intelligence, 1993, pp. 548–553.

[20] L. W. WAGENHALS, I. SHIN AND A. H. LEVIS, Creat-ing executablemodels of influence nets with coloredPetri nets, International Journal on Software Toolsfor Technology Transfer (STTT), Vol. 2, Dec. 1998,pp. 168–181.

[21] S. HAIDER AND A. H. LEVIS, Modeling time-varyinguncertain situations using Dynamic Influence Nets,Int. J. Approx. Reasoning, Vol. 49, 2008, pp. 488–502.

[22] S. HAIDER, From dynamic influence nets to dynamicBayesian networks: A transformation algorithm,Int. J. Intell. Syst., Vol. 24, 2009, pp. 919–933.

[23] LEE W. WAGENHALS, J. T. RIED, R. J. SMILLE ANDALEXANDER H. LEVIS, Course of Action Analysisfor Coalition Operations, Proceedings of the 6thInternational Command and Control Research andTechnology Symposium, Annapolis, MD: 2001.

[24] LEE W. WAGENHALS AND ALEXANDER H. LEVIS,Course of Action Analysis in a Cultural LandscapeUsing InfluenceNets, Proceedings of the IEEESym-posium on Computational Intelligence for Securityand Defense Applications, Honolulu: 2007.

[25] L. WENTZ AND L. WAGENHALS, Effects Based Op-erations for Transnational Terrorist Organizations:Assessing Alternative Courses of Action to Mit-igate Terrorist Threats, Proceedings of the 9thInternational Command and Control Research andTechnology Symposium, San Diego, CA: 2004.

[26] L. WAGENHALS, R. JANSSEN AND E. DEGREGORIO,Model Interoperation for Effects Based Planning,Proceedings of the 9th International Command andControl Research and Technology Symposium, SanDiego, CA: 2004.

[27] L. WAGENHALS AND L. WENTZ, New Effects-basedOperations Models in War Games, Proceedings ofthe 8th International Command and Control Re-search and Technology Symposium, Washington,DC: 2003.

[28] S. HAIDER, A. K. ZAIDI AND A. H. LEVIS, Identifi-cation of best sets of actions in Influence Nets, Int.J. Hybrid Intell. Syst., Vol. 5, 2008, pp. 19–29.


[29] S. HAIDER, A. ZAIDI AND A. LEVIS, A heuristicapproach for best sets of actions determination ininfluence nets, Information Reuse and Integration,2004. IRI 2004. Proceedings of the 2004 IEEEInternational Conference, 2004, pp. 600–605.

Received: August, 2011Revised: June, 2012

Accepted: June, 2012

Contact addresses:

Sajjad HaiderFaculty of Computer Science

Institute of Business AdministrationKarachi, Pakistan

e-mail: [email protected]

Muhammad ShafqatUNESCO

Islamabad, Pakistane-mail: [email protected]

Shabih HaiderDepartment of Economics and Finance

Institute of Business AdministrationKarachi, Pakistan

e-mail: [email protected]

SAJJAD HAIDER received the B. Sc. (Hons.) degree in Statistics andM. Sc. degree in Computer Science from the University of Karachi,Pakistan, in 1997 and 1999, respectively, and the MS degree in In-formation Systems and the Ph. D. degree in Information Technologyfrom George Mason University, Fairfax, VA, in 2002 and 2005, respec-tively. He is currently an Associate Professor of Computer Scienceand the Head of the Artificial Intelligence Lab, Faculty of ComputerScience, Institute of Business Administration, Karachi. He also holds adual position of Honorary Associate at the University of Technology,Sydney where he is a member of the Innovation and Enterprise Re-search Lab. From 2005 to 2007, he worked as a Predictive Analyst atFannie Mae, USA. His areas of interest include decision sciences, prob-abilistic reasoning, cognitive robotics, machine learning, computationalintelligence and systems engineering.

Dr. Haider was a recipient of the 2004 Gary F. Wheatley Best StudentAward at the Command and Control Research and Technology Sym-posium and the Best Paper Award from the International Council ofSystems Engineering. He also won the best student awards during hisB. Sc. (Hons.), M. Sc. and MS studies. He was recently awarded theBest Teacher Award by the Higher Education Commission of Pakistan.

MUHAMMAD SHAFQAT has done Bachelors in Business Administration(BBA) from the Institute of Business Administration, Karachi. He hasworked as policy consultant at the Planning Commission of Pakistandevising the ’Innovation Roadmap’ for Pakistan’s 10th five-year devel-opment plan (Framework for Economic Growth 2012–2016). Since2011 he has been working with UNESCO Islamabad, first as ForesightConsultant for its Science Sector, and now as Country Program De-veloper for the organization. His research interests are in the areas ofcrowdsourcing, social innovation and policy forecasting.

SHABIH HAIDER received the BE degree in Electronics from NED Uni-versity and Masters of Business Administration degree from Instituteof Business Administration, Karachi in 1979 and 1991, respectively.He is currently working as an Assistant Professor of Economics andFinance in the Institute of Business Administration, Karachi (IBA). Healso serves as Management Advisor in Biogenics Pakistan Private Ltd.Prior to joining IBA, he taught at Hamdard University and worked asmarketing manager in Alcatel, senior engineer in NESPAK (the largestengineering consultants in Pakistan) and senior application and serviceengineer in Zelin Private Limited. Mr. Haider’s interests lie in costaccounting, financial economics and business statistics.

Documents

Using Inﬂuence Nets in Financial Informatics: A Case Study ...€¦ · Informatics: A Case Study of Pakistan Sajjad Haider1, Muhammad Shafqat2 and Shabih Haider3 1 Faculty of Computer