“Optimization of ATMs filling-in with Cash”

B.Sc. Industrial Engineering and Management

EOT (LAB) MIDS Paper

Submitted By:
Mohammad Fahad [IEM-21-22]

Submitted To:
[Engr. Abdul Hannan]

Department of Industrial Engineering and Management,

 Faculty of Quality and Industrial Systems Engineering
University of the Punjab

Executive Summary

An easy-to-understand summary of the report is as follows: This report suggests a new

method for X Bank to determine how much money to place in its ATMs at Sundar
Industrial Estate in Lahore. The method uses past data and mathematical formulas to
predict how much money customers will need each day. This helps the bank avoid both
running out of money (which is bad for customers) and having too much money sitting
in the ATMs (which is expensive for the bank). The report shows that this new method
is better than the one the bank is currently using.

Three primary cost factors are incorporated into the model: cash freezing (holding
excess cash), transportation (replenishing ATMs), and insurance (securing cash). The
effectiveness of the generated forecasts is assessed against the bank's current
prediction system, which relies solely on historical data to determine ATM refill timing
and amount.
Contents
1 Introduction 4
2 Modeling Daily Cash Demand 7
2.1 Time Series Analysis 7
2.2 Linear Regression 7
2.3 Compound Poisson Process 9
3 Optimization of ATM Filling 10
4 Decision Making 14
5 Conclusion 14
1 Introduction
Automated Teller Machines (ATMs) are 24-hour self-service machines that enable bank
customers conducting their financial transactions without visiting the bank branch. In
spite of online banking facilities expansion, need for ATMs transactions remains high
over years and makes ATMs an irreplaceable devices in everyday life. In order to meet
growing cash needs of bank clients, banks have to increase continually the number of
their ATMs in different location to make cash available. The number of ATMs in the
world in 2013 was over 2.2 million.
ATMs are still widely used despite online banking, and banks need to keep them
stocked with cash. Predicting the right amount for each ATM is tricky, as it involves
balancing customer needs with three costs: holding extra cash, transportation, and
insurance. This report focuses on a bank with various ATM types (active/inactive) and
locations (inside branches, outside isolated, or clustered in sectors). Understanding
these different types of ATMs is crucial for predicting their cash needs accurately.
Even with online banking on the rise, ATMs remain crucial for many. But stocking them
with the right amount of cash is a challenge. Banks need to balance meeting customer
demand with keeping costs down. This report tackles that issue for a bank with diverse
ATMs: some active, some inactive, some inside branches, and others outside, either
isolated or grouped in sectors. Understanding these different types of ATMs is key to
predicting their cash needs accurately and efficiently.
The Bank in question operates with different types of ATMs, based on their activity and
location. According to their activity, ATMs can be active or currently inactive. For our
problem, location of ATMs is crucial. Thus, it is important to make distinguish between
internal (inside of the branches) and external ( outside of the branches). External ATMs
can be isolated or clustered into sectors, where each sector consists of 2, 3, 4 or 5
closely located ATMs.
 Sundar Industrial Estate phase-1

To save on transportation costs, the bank refills grouped ATMs together. This raises
another crucial question: how often to refill them? The bank wants to avoid empty ATMs
but also minimize costs:

 Holding too much cash (freezing) is expensive.

 Transporting cash frequently adds cost.
 Insuring cash also contributes to expense.

Total
4000000
3500000
3000000
2500000
2000000
1500000
Total
1000000
500000
14-02-2011
0 13-01-2011
16-01-2011
17-02-2011
19-01-2011
20-02-2011
22-01-2011
23-02-2011
25-01-2011 02/01/2011
26-02-2011 02/02/2011
28-01-2011 01/04/2011
30-01-2011 02/05/2011
01/07/2011
02/08/2011
01/10/2011
02/11/2011
 Figure 2: ATMs daily withdrawals

To save on transport costs, the bank groups ATM refills and uses a loop route rather
than going to each one individually. They currently rely on an app that uses historical
data to suggest refill times and amounts, but they're looking for a better way to minimize
costs while still meeting customer needs. This includes avoiding empty ATMs,
especially during peak times like holidays and paydays. The report uses data from 2011
and 2013 to analyze ATM cash transactions and identify any patterns or trends that
could help improve refill strategy. Essentially, the bank wants to find the sweet spot
between keeping ATMs stocked and minimizing costs like cash holding, transport, and
insurance.

Total
25000000

20000000 Sunday

15000000 Monday
Tuesday
10000000
Wednesday
5000000
Thursday
0 Friday
Saturday

Figure 3: Average value of cash withdrawals for different days of the week in 2011

Time series trend and volatility given in graph prevent us to identify seasonality. Using
available data, it is observable that cash withdrawals are not same for all days of the
week. Mondays and Fridays are the days in which the largest amounts are withdrawn
and Saturdays and Sundays the days with the lowest withdrawn amounts( 3). Important
and interesting part in daily ATM withdrawals analysis is also identifying monthly and
holidays seasonality. In particular, we could expect summer and winter seasons to be
the periods of highest withdrawn amount.
The report is organized as follows. Section 2 is referred to modeling cash demand. It
presents time series and regression models for daily cash demand predictions at
external ATMs and theoretical proposal for predictions at internal ATMs. Section 3 is
devoted to determining optimal quantity of cash to be stored in ATM based on the short
term prediction from section 2.
Last section 4 discusses the decision making process.

2 Modeling Daily Cash Demand

Probably the most natural approach to model daily cash demand is to fit time series
models which would allow forecasting of future demand. We tried several such
approaches. Another reasonable approach that we explored is to model cash demand
by a linear model based on various statistics of demand history as predictors.

2.1 Time Series Analysis

Daily cash demand represents a natural time series data. However, it can be organized
in time series in different ways. We considered the following ways to organize the data
into time series:

• Demand values for all days in the considered period are used to form a single time
series. This is the simplest way to arrange the data and therefore a natural first
choice.

• Separate time series were formed, consisting of demand values on specific week
days. It is reasonable to assume that the demand behavior on Friday, which
precedes the weekend, is distinct from Wednesday which is in the middle of the
week. By modeling these behaviors separately, one might hope to avoid complex
behavior in the data obtained by merging several simpler ones.

• Separate time series were formed for days that comprise the first, second, third
and fourth week in a month (e.g. all days of first week of January, first week of
February, first week of March, etc. are used to form one time series). This
approach aimed at capturing dynamics within different parts of the months which
may differ due to factors like paycheck arrival and similar.
 • Three time series were formed by different resolution of data aggregation — daily,
weekly, and monthly time series. The forecasts of these three would be averaged
to provide a prediction, the aim being to provide more robustness to simple daily
estimates.

The methods used to model time series was by using MS Excel, MiniTab and Python
Programming.

2.2 Linear Regression

Predicting cash demand can also be approached as a regression problem. We

considered linear model of form:

where αi are the parameters of the model to be estimated from the data, and hi are
different statistics of cash demand history. The model is evaluated for specific day d.
Values li are which govern the computation of statistics hi, like the length of the history
considered. Specifically, we considered the following statistics:

• The average recent demand history (h1) represents the average of cash demand in
l1 days prior to given day d.

• The average weekly history (h2) represents the average of cash demand on the
day corresponding to d in l2 weeks prior to the one d belongs to. For instance, if d is
a Monday, h2 could be an average of cash demand on previous 3 Mondays.

• The average monthly history (h3) represents the average of cash demand on the
day corresponding to d in l3 months prior to the one d belongs to. For instance, if d
is 15th of the current month, then h3 could be an average of cash demand on 15th
days of previous three months. In case that previous months do not contain the
corresponding day (e.g., if d is 31st of July, the problem arises since there is no 31
of June), the information on cash demand on the last day of that month is used
instead.

• The average yearly history (h4) represents the average of cash demand on the day
corresponding to d in l4 years prior to the one d belongs to. For instance, if d is a
 Christmas in the current year, h4 could be an average of cash demand on three
previous Christmases.

The estimation of parameters αi was performed by ordinary least squares method. In

case of strongly correlated predictors, ridge regression would be advised, but our data
exhibited no such behavior. We performed experiments by fitting the proposed linear
model to the data for various values of metaparameters li. Specifically, we used l1
{1,...,30}, l2 {1,...,52}, l3 {1,...,12}, and l4 = 1. The last choice was forced by the amount
of the data at our disposal. For all models roughly described by constraints l1 > 20, l2 >
20, and l3 = 12, the coefficient of determination R2 (a standard measure of goodness of fit
for linear models, bounded in the interval [0,1]) was approximately 0.4 with small
differences. No model achieved lower than 0.32. On the other hand, to our knowledge,
current Bank practices exploit only the information on average yearly history. Using only
such information, one obtains R2 = 0.226. Even though there is a lot of room for
improvement to the proposed model, this clearly shows that there is useful information
in the data which is still not exploited. Since the obtained results are promising and
indicate that there is the information in the data unused by current Bank practices, it is
reasonable pursue further improvement of the regression based cash demand
prediction. We envisioned several directions of research:

• Formulating new statistics (hi) to be used as predictors. Four statistics used in our
experiments were the first reasonable guesses which were based on the idea of
capturing recent activity and three different kinds of seasonality. Based on the
expert knowledge and experience of Bank’s employees, it is probably possible to
formulate more reasonable statistics which would be correlated with ongoing cash
demand.

• Incorporation of outside knowledge into the model (e.g., a sports event is

announced in the vicinity of the ATM) by adding new α parameters for such
predictors. One would need to provide such historical data in order to estimate the
corresponding parameters.

• Replacing the proposed model with a family of models, each corresponding to

different parts of the year. For instance, a separate model could be estimated for
the days of each month (e.g., the parameters for January model would be trained
only on data consisting of cash demand on days of several previous Januaries).
 • Formulating the model relying on nonlinear transformation of predictors or a
nonlinear model. Use of such models would be justified if previous efforts fail in
achieving significant boost of prediction quality over the model already evaluated.

2.3 Compound Poisson Process

Although Bank’s representatives suggested focusing only on withdrawals at external

ATMs, in this subsection we propose simple model for forecasting cash demand at
internal ATMs in theoretical manner. Motivation for this was found in [2]. Basic idea was
to simulate customer’s uncertain demand using some stochastic processes. Capturing
randomness of our problem requires modeling the following processes:

• number of customers arrived at ATM at the time interval (0,t) ;

• total amount withdrawn in the time interval (0,t).

First process (numbers of ATM users during each day) represent a situation when
outcome variable is a count. Such variable takes on a limited number of values and are
never negative. Additionally, their mean and variance are often related (which isn’t true
for normally distributed variables). Common approach for describing counting
processes is to use a Poisson process. Poisson process is one of the most important
models used in queuing theory. It is a simple stochastic process used for modeling the
times at which arrivals enter a system.

Formally, the Poisson process is described by the so called counter process Nt. The
counter tells the number of arrivals that have occurred in the interval
(0,t) i.e:

Nt = number of ATM’s customers in the interval (0,t).

We assume that customers arrive at ATM at the times of a Poisson process with a rate
λ per day and consider day as a time unit. Parameter λ is the average of withdrawals
made of ATM customers in a day. Upon this assumption, important properties of the
process can be derived. The most important one is that the number of arrivals Nt in a
finite interval of length t obeys the Poisson distribution of parameter λt which means that
we are able to calculate the probability of n ATM customers in the time t by formula:

.
Moreover, the number of arrivals in intervals (t1,t2) and (t3,t4) in nonoverlapping intervals
(t1 ≤ t2 ≤ t3 ≤ t4) are independent. Mean and variance of Nt are both equal to λt.

Assuming that the successive withdrawn amounts are independent and identically
distributed random variables, we can describe the other process of our interest (total
amount withdrawn in the time interval (0,t)) by using generilized Poisson process.
Let At be the total amount withdrawn in the time interval (0,t) and Nt the number of
customers to come to the ATM in the interval (0,t) as described above. If we denote the
amount of the nth withdrawal with An, then At can be represented as the following random
sum of random variables:

Nt

At = XAn.
i=1

Process {At} is called a compound Poisson process. Using this assumtion we are able
to derive following characteristics of the process:

E[At] = λtE[A1]

and
.

3 Optimization of ATM Filling

The report suggests that for short periods (around 20 days), it's possible to find the
perfect refill schedule for an ATM (or grouped ATMs) that minimizes costs like cash
holding, transport, and insurance. This involves checking every possible combination of
days for refills - which quickly becomes impractical for longer periods. However, since
any forecasting errors might require adjusting the schedule anyway, a perfect long-term
solution might not be crucial.
 Total
18000000
16000000
14000000
12000000
10000000
8000000
6000000
4000000 Total
2000000
Saturday Wednesday
Saturday Wednesday Wednesday
Thursday
0 Monday
Sunday Monday
Sunday Tuesday FridaySaturday

C H M N NH

Figure 4: Difference between Normal Days and 1st Week of the Month

Besides the optimization algorithm which works on precomputed data, a real-time

approach will be also required for dealing with unexpected trends in withdrawals. This
approach has to take into account the current amount of money in the ATM as well as
expected withdrawals in the next few days. Depending on whether there is too much or
too little money and e.g. a shortage can occur, the optimization may then be redone
starting with the current day.

Since a real-time approach is secondary to the optimization the latter is more important
for the solution. The optimization algorithm requires quantified costs associated with the
filling-in schedule:

• cost of insurance - dependent on the amount of money filled-in at the given ATM,

• cost of cash freezing - dependent on the amount of money at any given time in the
ATM,

• cost of filling-in - any filling-in has a fixed cost associated with the distance to the
closest branch from which the cash for the ATM can be supplied.

The result based on those would include the amounts of money as well as days at
which filling-in should take place. To take into account a margin of money that should
be at all times in the ATM an additional set of assumptions can be included if desired to
fine tune the results. These assumptions can include:
 • setting a minimum and maximum limit of the number of days which can occur
between sequential filling-ins,

• extra percentage of money that an ATM should be filled with (e.g. 20 % so that
assuming the forecast hold 20% of the filled-in money will be left when the next fill
occurs).
An example algorithm works by first assuming that a filling-in occurs every day and then
checking other possibilities. If no other than basic assumptions are given then Tables 1-
3 show a sample of three steps conducted by the algorithm when optimizing total costs
of filling-in schedule.
Date Forecast Fill-in ( morning
)
2011-01- 6000 6000
13
2011-01- 9000 9000
14
2011-01- 3000 3000
16
2011-01- 1000 1000
17
2011-01- 10000 10000
19
Table 1: Step one of the algorithm

Date Forecast Fill-in ( morning

)
2011-01- 6000 6000
13
2011-01- 9000 9000
14
2011-01- 3000 3000
16
2011-01- 1000 11000
17
2011-01- 10000
19
 Table 2: Step two of the algorithm

Date Forecast Fill-in ( morning

)
2011-01- 6000 6000
13
2011-01- 9000 9000
14
2011-01- 3000 4000
16
2011-01- 1000
17
2011-01- 10000 10000
19
Table 3: Step three of the algorithm

The steps of the algorithm are as follows:

1. assume that the filling-ins occur in a given order;

2. calculate the amount of money to be filled in as a sum of forecasted withdrawals
in the consecutive days between every filling-in and next filling-in (here for
example a 20% extra amount can be added
in);

3. calculate the total cost for assumed schedule;

4. if total cost is less than previous minimal cost then set the discovered schedule as
optimal;

5. modify schedule;

6. repeat from step 1.

The final results will describe at what days should the filling-in occur and what amount
of money should be filled in. Example showed in Table 4. gives three such days with
accordingly 15000, 4000 and 10000 money units to be filled in at those days. Since the
algorithm checks every solution the result is the global minimum cost for the schedule
optimized for these days.
Date Forecast Fill-in ( morning
)
2011-01- 6000 15000
13
2011-01- 9000
14
2011-01- 3000 4000
16
2011-01- 1000
17
2011-01- 10000 10000
19
Table 4: Example final results of the algorithm

A real-time approach will than check the amount of money left at the ATM at the end of
the day (e.g. 7000 on first day in Table 4) and should a shortage be expected (in Table
4 because of withdrawals of 9000 on the next day) the optimization algorithm can be
ran again.

4 Decision Making
Given precise predictions on cash demands, in order to determine optimal filling of
ATMs with cash, it suffices to use the described optimization procedure. However,
predicting cash demand is a rather tough task, especially several days in advance.
Namely, one of the statistics used — average of cash demand in last several days is
not available for the future since demand history is not available for the given or future
days. For future days, that statistic can only be approximated by the average
corresponding to the day on which the prediction is made. Therefore, for better decision
making, it would be better to perform optimization each day as soon as new demand
data becomes available, and that way correct potentially bad decisions caused by bad
prediction. For example, if one predicts on day one that an ATM will need refilling in 5
days, surprisingly high cash demand on that same day might change the prediction on
the second day so that the refilling period is shortened.

5 Conclusion
Aim of the study group team was to minimize total costs of the X Bank when supplying
ATMs with cash. Main task was to find an optimal balance between two opposite
approaches such that ATMs are supplied with sufficient amount of money for satisfying
customers’ needs. Philosophy of the first one is that ATMs should be rarely refilled with
large amounts of cash which brings low transport cost, high freezing and high insurance
costs. On the contrary, other approach is based on frequently supplying ATMs with
small quantities of cash which reduces freezing and insurance cost and increases
transport cost. Numerical experiment was focused only on transactions at external
ATMs. Time series analysis and linear models were employed for obtaining predictions
of daily cash demand at each ATM. Obtained results were later incorporated in
optimization model of filling-in schedule. Final outcome of optimization algorithm gives
information on days when the filling-in should occur and about amount of money should
be filled in. The algorithm checks every solution and it results with the global minimum
cost for the schedule optimized for these days.