Assessing the Potential Impacts of Microfinance with Agent Based Modeling
Salim Rashid, University of Illinois, Youngeun Yoon, Ministry of Finance, Govt of South Korea, Shakil
Bin Kashem, University of Illinois
Abstract:
Recent empirical results have cast doubt on the value of Microfinance as a tool for reducing poverty. The
difficulty lies in evaluating the impact of Microfinance in a world where the Microfinance institutions,
their borrowers, and government policy are constantly changing. In this environment, policymakers are
particularly challenged. They must usually act before full information arrives, and before econometric
results can unambiguously assess the data. This can lead to poorly targeted policies, such as inappropriate
or ineffective Microfinance regulation, and can also create a poor environment for assessing the impact,
positive or negative, of Microfinance in practice.
There is a need for a comprehensive and transparent framework to develop the theoretical grounds for
believing in (or against) the efficacy of Microfinance, which can at the same time be used as a testing
ground for policymakers. This study is a first attempt to develop an agent-based modeling (ABM)
framework for pre-policy-implementation testing of the effects of Microfinance. Under the ABM
paradigm, a set of behaviors for individual agents in the economy is used to construct a simulation
whereby random interaction allows agents to self-organize into various groups (borrowers, consumers,
moneylenders, etc). Once the basic principles are accepted, at the policy stage, the underlying economic
interactions between agents can be tightly built upon supporting empirical data e.g. profit rates of small
enterprises, default rates on loans, etc.
Using ABM, the effectiveness of alternative Microfinance policies can be analyzed prior to
implementation. Thus, policy tools for Microfinance can be first tested within a simulation framework
'sandbox', and can then be more effectively targeted for implementation in the real world.
Assessing the Potential Impacts of Microfinance with Agent Based Modeling
Introduction
Since its inception in the villages of Bangladesh in the 1970s, the modern Microfinance revolution is
emerging in many countries of the world as a tool for poverty reduction (Robinson, 2001). The award of
the Nobel Peace prize to Dr Yunus and the acceptance of Microfinance as the primary tool to attack
poverty seem to have galvanized its opponents. Severe doubts have been cast upon the effectiveness of
Microfinance (Roodman and Morduch 2009).
“Thirty years into the movement, it might seem strange that researchers are still asking whether
microfinance reduces poverty. In fact, by the standards used to judge whether drugs are safe and effective
in the bloodstreams of people, the safety and effectiveness of microfinance injected into
the fabric of villages and barrios remains unproven.” (Roodman, 2009)
The actual causality between Microfinance and poverty reduction is yet to be well established. The use of
different methodologies, variables and assumptions by different researches are giving contradictory
results (Pitt and Khandker, 1998; Khandker 2005; Roodman and Morduch 2009). The difficulty also lies
in evaluating the impact of Microfinance in a world where the Microfinance institutions, their borrowers,
and government policy are constantly changing. In this environment, policymakers are particularly
challenged. They must usually act before full information arrives, and before econometric results can
unambiguously assess the data. This can lead to poorly targeted policies, such as inappropriate or
ineffective Microfinance regulation, and can also create a poor environment for assessing the impact,
positive or negative, of Microfinance in practice. There is a need for a transparent framework, which
focuses upon essentials, to develop the theoretical grounds for believing in (or against) the efficacy of
Microfinance, which can at the same time be used as a testing ground for policymakers. It is time to leave
aside the institutions within which Microfinance may be implemented and to focus upon the primary
ideas. Microfinance is, in essence, micro-finance ; if finance works, so must microfinance.
Agent-based modeling (ABM) framework provides decentralized, dynamic environments with
populations of evolving, heterogeneous, boundedly rational agents who interact with one another,
typically locally (Duffy, 2006). This study attempts to develop an agent-based modeling framework for
pre-policy-implementation testing of the effects of Microfinance in a simplified environment. Under the
ABM paradigm, a set of behaviors for individual agents in the economy is used to construct a simulation
whereby random interaction allows agents to self-organize into various groups (borrowers, consumers,
moneylenders, etc). Here the underlying economic interactions between agents are tightly built upon
supporting empirical data of e.g. profit-making for small enterprises, default rates on loans, etc.
Aggregate outcome of varied activities performed by heterogeneous agents can be analyzed in this
approach which gives an effective framework for policy analysis.
The model
The key objective of the model was to evaluate the impact of Microfinance on the income level of the
poor and the repayment ratio of microcredit over time. It was carried out by a simulation in an agent
based modeling environment in NetLogo. It emulates the behavior of different agents in a simplified
environment and shows the outcome as a result of the interactions among agents. Agent based modeling
is an efficient way to get a lucid picture of the impacts of Microfinance, because it can model the
structures and procedures of Microfinance from the perspectives of different agents in real economy, and
thus can give the aggregate result of the interactions among different agents. Several parameters of the
model are chosen to reflect the orders of magnitude found in a recent study of Urban Microfinance in
Bangladesh (Rashid and Bashar, 2010)
This model considered 5 types of agents related with Microfinance: the poor in the population, the other
people in the population, a Microfinance institution (referred here as “MFI”), local money lenders, and
suppliers of raw materials. Each agent group is heterogeneous, because they are under the different
situations and pursue different objects from the interactions. By resorting to agent based modeling, the
model can reflect the various aspects of real world, make updates, and get the final results after the
Microfinance procedure starts. Here at the beginning the poor people get microcredit from MFIs to start a
microenterprise according to their productivity level. To do that, they should prepare for start up business
such as buying raw materials, renting a building and so on. After producing the product, they need to
trade it in the market and repay the microcredit. These procedures require rules to direct the motion of the
system and procedures to update and keep track of the agent’s situation; this involves many variables such
as wealth level, amount of loan, volume of production and trade of the products over time. An Agent
based approach replicates these activities in a virtual environment and shows the aggregate outcomes in
different time periods.
Model set up and agent characteristics
The model creates an environment where different agents live and interact with each other. This
environment is a virtual economy space where economic transactions such as borrowing, repaying,
manufacturing, and trading happen as time goes on. This space mimics economic activities in actual
world and helps us to capture the outputs of economic transactions. In the NetLogo program this space is
represented by a black square, which is composed of a grid of patches (Figure-1). The agents can move in
this spatial environment and interact with each other.
The model starts with given number of people living in the space. A portion of whole population (to be
stipulated by the user of the model) is assumed to be ‘poor’ agents in this model who have relatively low
wealth level. Since the purpose of this model is to predict the impact of microcredit on the wealth level of
the poor, these poor are our main focus throughout the entire procedures. The rest of the populations are
the non-poor, those individuals with more wealth than the poor. They are included in the model to reflect
the economic interactions with the poor. Initially individual agents in the population are randomly
distributed throughout the space. Each poor individual has a number of characteristics such as wealth, sex
, and productivity, which define their participation in the model. More specifically, each individual,
whether male or female, has a wealth level drawn from the normal distribution N (50, 15). [Also he/she
has sex index which is either male or female]. It was assumed in this model, in keeping with the usual
facts of LDC’s, that the overall productivity of a male is higher than that of a female, since men are
usually more educated than women in LDC’s. There are also a variety of social and cultural constraints
that enable males to earn somewhat higher returns than females of equal productivity. That is, the
productivity level for man is drawn from the normal distribution N (4, 1.5), while that for woman is from
N (3, 1). While the average productivity of a male is higher than that of female, the variance of
productivity for a male is bigger than that of a female. It is based on the fact that female borrowers are
usually considered more reliable than male borrower by the MFIs. The proportion of each sex is assumed
to be half of the population, but the exact number may vary because the allotment process utilized random
number generating function of the Netlogo.
This model assumed only one MFI located at the center of the economy space around which all the agents
are randomly distributed. The function of the MFI is to distribute microcredit to help the poor start their
own microenterprise. To simulate the impact of microcredit, several parameters on the loan of MFI were
provided: (1) the total amount of fund which can be used for loans was set to be 1,000; (2) the maximum
amount of money for each loan is set to be 10; (3) the interest rate charged by MFI is set to be 10% per
year. These are only illustrative values. There is much heterogeneity among the poor and among the
MFI’s, so no one set of values can be representative. Options were provided in the model to change all of
these parameters of the MFI (figure-1) so that users can analyze their impacts on overall outcome of
microfinance.
Local money lenders were also included in the model. The number of money lenders is set to be 4 to
reflect the fact that there are many more local money lenders than MFI’s in the real economic world. The
interest rates charged by the local lenders are usually much higher than those charged by MFI’s
(Robinson, 2001). In this model, this interest rate is assumed to be 3 times higher than that of MFI. But,
given the smaller scale of the typical moneylender, the amount of each moneylender’s loan is set to be
just half of that from the MFI.
After getting a loan either from the MFI or from local money lenders, the poor get ready for starting a
microenterprise. For this purpose, he/she needs to buy raw materials, rent a store, and so on. To represent
starting expense and time, the number of suppliers was set to be only 4. Due to the scarcity of suppliers,
the individual(s) with money to invest has to spend some time to buy raw materials from them. This time
stands for the expense of money as well as time itself, because the wealth level of an agent decreases as
they move throughout the space (which is explained in next section).
Setup options of the model Virtual economic space for agents
Plot
Figure-1: Interface of the model in NetLogo
Monitor
Model updating process and implementation of microfinance activities
With execution of the model each of the agents start to perform their defined jobs. The poor, as well as
the non-poor, start to move randomly throughout the space. This movement represents the everyday
transactions which are carried out by individuals. The wealth levels of individuals, we focus only on the
poor, decreases as they move around. A decrease of wealth of 0.05 per day was assumed in this model,
which can be interpreted as everyday spending.
After knowing their current wealth level, the poor with promising business plan, i.e. those who have a
productivity level above 5, directly apply for Microcredit from the MFI as an individual. It was assumed
in the model that there is no depreciation of productivity in the process of turning their productivity to
actual plan. Hence, individuals with productivity level higher than a threshold, say 5, should be profitable
and can produce a business plan with high enough rate of return to convince the MFI for extending
microcredit. Hence, if a poor individual’s productivity is greater than 5, implying he has some project
with high yield, he/she directly goes to MFI to get the money for investing on that project without
considering any joint business with other individuals. But in the production procedure, the realized rate of
return is set to be same across individuals. In this sense, the productivity level plays a role towards
accessibility of microcredit for poor agents. After microcredit is given to one of the poor, the amount of
funds available to the MFI for microcredit will be decreased and the debts of the poor will be increased.
Such updating, essentially keeping accounts, is carried out automatically by the model over time.
If the productivity level of any agent is not greater than 5, his/her project is assumed as unpromising and
in such a case he/she needs to find someone who will join together to form a microenterprise. Thus,
he/she moves around to find another individual under same situation and if he/she finds one whose
productivity together with his/her productivity is higher than the threshold level for the joint loan, they
will form a group and jointly contact the MFI to apply for microcredit. Threshold level for joint loan was
set to be 6, 1 unit bigger than that for individual loan, to reflect the tradition of loan in real world that the
process for joint loan is more complicated than that for individual loan. If there are several potential
candidates to form a group, the individual found in the nearest neighborhood is assumed to be chosen.
The model also assumed that a loan from the MFI is not always available to the poor. If the available
funds for microcredit is insufficient to carry out the loan, then an individual cannot get the loan even
though his/her productivity is high enough. An individual whose productivity is low and who fails to find
a partner with whom he/she can apply for the loan, may end up becoming terribly poor, (say wealth level
less that 20), and would better give up getting microcredit from the MFI. Under these circumstances, the
poor should get the loan from local money lenders, even though the condition of the loan is inferior to that
from MFI, i.e. the interest rate is higher and the amount of money is smaller than the loan from MFI.
After a poor agent get a loan either from MFI or from local money lenders, he/she (or they in the case of
joint loan) should get ready for starting the business, i.e. he/she needs to buy raw materials. The
purchasing of raw material is possible only when he/she meets raw material suppliers. It was specified in
the model that any poor agent can buy 1 unit of raw materials with 2 units of wealth and on each
transaction he/she can buy 2 units of raw materials. Besides buying raw materials, the agent can hire other
poor persons who live in the nearest neighborhood to be employed as a worker rather than engage in self
employment. Poor individual(s) who fail to get a loan and have very low wealth levels (less than 20) may
give up the idea of starting his own business and decide to be an employee. For the employee, earnings
from employment (payment of labor was set to be 4 in the model) is an alternative way to overcome
poverty.
The model introduces several assumptions for the production process of the microenterprises. (1) Inputs
for the production are limited to only raw material and labor. That is, by using 2 unit of raw material or 1
unit of hired labor, 1 unit of output will be produced. Also (2) these two inputs are perfect substitutes.
Finally, (3) Production is possible only after he/she accumulates raw materials to produce two units of
products, i.e. 4 units of raw materials or 2 units of labors. This introduces a small startup fixed cost into
the model and prevents arbitrarily small enterprises from starting up. The output and inventory of the
microenterprises is calculated and updated systematically based on the progress of production and trading
over time.
When a poor agent successfully completes the production process, he/she then sells the product to other
people in the population. It was specified in the model that an agent can sell the product only after he/she
get an inventory of more than 2 units of product. The price of the product was set to be 10 units of
wealth, implying the seller get 6 units of wealth by selling 1 unit of product whose variable cost is 4 units
(the price for 2 units of raw materials or 1 unit of labor). Although one poor agent can sell the product to
another poor, such a transaction is just a redistribution of wealth within the poor. Thus it has no impact on
the aggregate wealth level of the poor, which is the main focus of this model. From this point of view, it
was assumed that the transaction occurs only between one of the poor and one of the other in the
population.
Using the profits from the sales, agents can repay the loan to the MFI or local money lenders. The rules
differentiate the type of repayment. More specifically, the borrowers repay the microcredit only after they
sell 5 units of products, by which they get 30 units of profit. But they repay the loans from local money
lenders sooner, after they trade 4 units of products with 24 units of profit, due to the presumed inferior
financial terms of loans from money lenders compared to microcredit. The amount of repayment is
principal plus the interest for the time of the loan. Such times are also calculated in the model and the
amount of required repayment is specified accordingly for each of the borrowing agents. Figure-2
provides a flow chart of the model which graphically shows the whole procedure followed in the
YES
NO
model.
Figure-2: Flow-chart of the model
Other features of the model
The model was designed to be versatile for a variety of different Microfinance worlds. To achieve
this end, several user defined options were provided in the model (figure-1). The user can manipulate the
interest rate, population, number of poor agents in the space, the time period of analysis and presence of
any idiosyncratic disaster that may happen during the analysis period. By manipulating these parameters
users can easily evaluate the sensitivity of each of the parameters on the overall outcome of microfinance.
Each time period in this model stands for a day and the profiles of the agents are updated each time up to
specified time period.
In reality, there can be many idiosyncratic disasters such as natural disasters or an abrupt decline
of demand in the market, which cannot be covered by the insurance or precautionary actions by the
agents. An option for this kind of idiosyncratic disaster was also included in the model to represent such
real world phenomena. Here the average probability of the occurrence of the idiosyncratic event is set to
be 5% of the running time. If any idiosyncratic disaster occurs, it would affect the income of individuals
with microcredit so that there will be no gain from trade, whereas sellers get 6 units of profit from each
trade when there is no disaster. To encode this disaster, a switch termed idiosyncratic disaster is used
(figure-1). If this switch is turned on the simulation get the result from disaster, which make it possible to
compare the results with and without idiosyncratic disaster.
The Netlogo program updates the result of interactions among agents in the monitors and shows
these in dynamically changing plots. Several monitors and plots were listed in the model (figure-1) to
show the progress over time. In the NetLogo program, monitors show the resulting number and plots
show the graphs. The monitors and graphs update the results while the procedure is going on.
The monitors that were included in the model are (figure-1): (1) the repayment ratio by the poor
to MFI, (2) total amount of loan from the MFI, (3) total amount of loan from the local money lenders, (4)
the number of repayment for the MFI, and (5) the amount of repayment to the local money lenders.
The model also contained following plots (figure-1 and figure-3) that show the progress over
time: (1) Average wealth level either for male or female in the Average Wealth plot. (2) The total amount
of loan either from the MFI or local money lenders in the Total Loan plot. (3) The number of repayment
either for the MFI or for the local money lenders in the Payment plot. (4) The total amount of trade in the
Trade plot. (5) Finally, the number of producers in the Producers plot.
Model simulation and analysis of results
Several simulations of the model were conducted changing different parameters of microfinance. It was
evaluated how income level of poor and repayment ratio of microcredit appears in different scenario (i.e.
change of interest rate, loan availability, credit limit, time period of analysis etc.). Before each simulation,
the setting of the economic space was specified and several simulations were conducted for the same
setting to get an average picture of the variability created by the randomness of agent characteristics.
Table-1 shows one such setting of the parameters of the model. Several simulations were conducted with
this setting of the model. Figure-3 (a-e) shows the output plots of one such simulation.
Table 1: Example of a Setting and Conditions of Parameters in the Model
Classifi
cation
Agents /
Procedure
Variable
Settings
General
Periods
5 years i.e. 365 * 5 days
Number
1
Fund
200
interest rate for
microcredit
10% / year
maximum amount of
microcredit
10
raw material
supplier
Number
4
local money
lender
Number
4
Number
100
Wealth distribution
N (50, 15)
Productivity for woman
N (3, 1)
Productivity for man
N (4, 1.5)
Number
200
MFI
Setup
poors
abovemiddles
Move
Forward
Wealth decrease
4 per movement
(poors, abovemiddles)
0.05 per movement
productivity requirement productivity >= 5
Borrow
buy raw
materials
Group formation
requirement
sum of productivities >= 6
Condition to loan from
local lenders
wealth < 20
Condition after
repayment
wealth >= 60
bargaining condition
4 wealth : 2 material
precondition
raw material >= 1
Employee requirement
wealth < 20, no loan either from
MFI or from local lenders
bargaining condition
1 labor : 4 wealth
precondition
raw material >= 4 or labor >= 2
production
transformation
(2 raw material or 1 labor) → 1
output
precondition
inventory >= 2
Bargaining condition
1 output : 10 wealth i.e. profit 6
wealth with marginal cost 4 ( 2 raw
materials or 1 labor)
probability
5%
loss when idiosyncratic
happens
0 gain from trade (actually loss)
Go
Employ
Produce
Trade
idiosyncratic
disaster
(a)
(b)
(c)
(d)
(e)
Figure 3: Plots from one simulation of the model
The results from the simulations of the model gave an optimistic result for the effects of Microfinance. It
was found that in all of the simulations the average wealth level of the poor decreased for a while (for
example, figure 3(a)), because it takes some time to make products, trade and gain the fruit of
microenterprise. But after a certain period of time (usually one year), the wealth of the poor starts to
increase and then maintains a higher rate of increase. The result for the repayment rate was also found to
be encouraging, being no lower than 90% and at times almost 99%. These two results together show the
bright prospects of microfinance in increasing the wealth of the poor.
From the simulation results of this agent based model under different scenario, two policy directions were
suggested to enhance the success of Microfinance. First, to magnify the positive impacts indicated in the
simulation we need to increase the amount of loanable MFI funds, make them easily available for start up
business, and make the investment rate of return higher by providing easy market access of the products.
Secondly, minimizing the negative impacts is also important. This can be done by decreasing the interest
rate charged by MFI and local money lenders, by keeping a closer look on the competitive behavior of
local money lendersi.
Conclusion
Any modeling exercise creates an abstract picture of reality and tries to find out the outcome of different
activities through that abstraction. The Agent based modeling exercise presented in this paper also
abstracted from reality to show the impact of microfinance through the interaction of different agents in a
simplified virtual environment. Through this abstraction it proves the positive impact of microfinance and
shows the issues to be considered for making it more efficient in the real world. Considering the
uncertainty and variability of the real world this model was made flexible, so as to be applicable in
various setups. But it still has lots of scope for improvement. By designing the model to cover more
delicate issues, such as diffusion of information across agents, various methods of group lending, and
competition between lenders, it would be better able to capture the interactions of Microfinance with
society. So, further extensions of this modeling exercise are suggested for making it much more
applicable in the real world and thereby employing it as a policy instrument in different social and
political settings.
Salim Rashid, University of Illinois
Youngeun Yoon, Ministry of Finance, Govt of South Korea
Shakil Bin Kashem, University of Illinois
References
Duffy, J., 2006, Agent-based models and human-subject experiments, In Handbook in AgentBased Computational Economics, L. Tesfatsion andK. Judd (Eds) (New York: Elsevier).
Khandker, Shahidur R. 2005. Microfinance and Poverty: Evidence Using Panel Data from
Bangladesh. World Bank Economic Review 19(2): 263–86.
Pitt, Mark M., and Shahidur R. Khandker. 1998. The Impact of Group-Based Credit on Poor
Households in Bangladesh: Does the Gender of Participants Matter? Journal of Political
Economy 106(5): 958–96.
Rashid, Salim and Toriqul Bashar, “Urban Microfinance in Bangladesh “. Report commissioned
by the Institute of Microfinance, Dhaka. 2010
Roodman, David. 2009. What Do We Really Know About Microfinance’s Impact? Available online at: http://www.microfinancegateway.org/p/site/m/template.rc/1.26.11408 (accessed on June,
2010)
Roodman, David and Jonathan Morduch. 2009. "The Impact of Microcredit on the Poor in
Bangladesh: Revisiting the Evidence." Center for Global Development Working Paper #174.
Robinson, Marguerite. 2001. The Microfinance Revolution: Sustainable Finance for the Poor.
Washington, DC, IBRD/The World Bank.
i
In countries like South Korea, where there are usury laws, to check whether or not they follow the usury law.