ATM Cash Prediction Using Time Series Approach

Automatic Teller Machine (ATM) brought the ease of use and convenience to users and financial institutions both. Although there are many types of transaction being supported by today?s ATMs, cash dispensed is still the larger use of these machines. One of the major problem associated with ATM, is fo

Project Title

ATM Cash Prediction Using Time Series Approach

Project Area of Specialization

Artificial Intelligence

Project Summary

Automatic Teller Machine (ATM) brought the ease of use and convenience to users and financial institutions both. Although there are many types of transaction being supported by today’s ATMs, cash dispensed is still the larger use of these machines. One of the major problem associated with ATM, is forecasting the daily demand. A financial planning for the network is unavoidable as (i) if a large cash lying idle at ATM, it would cost an interest opportunity loss, the magnitude of this is directly proportional to number of ATM hosted by the institution and (ii) if the machine run out of cash, it will make customers dissatissfied. It is very difficult to acquire a new customer in this highly competitive business. The problem can be seen as a time series analysis problem. This Final Year project (FYP) wish to tackle the same problem by applying time-series approach that try to utilize the seasonal effects as well. The data is collected from one local bank. The variation of seasonal time series models will be applied and with a forecasting horizon of one-week. The forecasting accuracy will be measured with standard evaluation metrics for time series analysis.

Project Objectives

ATM Cash Prediction has always been an alternating and varying problem given a large number of feature spaces to consider at making predictions before hand. In general, cash withdrawal is a stochastic process whose characterization tends to change over time due to several factors that have strong uncertainty. This optimization should satisfy both the customer for never having to deal with an out of cash ATM likewise with the more precise amount of cash present in the respective ATM’s will leave banks not having to deal with delays, customer unsatisfactory and a respectable increase in business opportunities. Replenishment of low amount of money often would not be a solution because each replenishment has a cost for out-of-service time and overtime pay of employees. Our Final Year Project focuses on making a prediction on behalf of banks for reimbursing ATM with cash based on individual ATM forecast of a location. The forecast is used for refilling of each ATM. The approach uses Time Series implementation. Daily time series prediction is beneficial for many applications, such as predictions for financial returns, stock market prices, retail transactions and so on. Withdrawal values are divided into trend, seasonality and daily input which are then marked with a Root Mean Square Value (RMSE) of all withdrawals as predicted and original. Test values are used to predict T number of days into the future. 

Project Implementation Method

1. Preparing Data. Considering the official dataset recieved, there are many empty or NA values in between which mark for non-completion or missing values. The values present explain scenarios based on the ATM usage, 0 denotes that the ATM was out of cash or a replenishment was needed similarly a missing value states that the transactions were not recorded that day. Time Series require complete datasets with no missing values as oppose to other statistics methods. Hence we used interpolation technique to fill the missing values.

2. Methodologies. Our approach requires understanding the data clearly and completely because for training a univariate time series we will be using Auto Regressive Integrated Moving Average with exogenous variable types or ARIMAX model, followed by Neural Networks like Deep AR and recurrent neural network (RNN) for better accuracy of the former model. The implementation uses Python as the programming language and uses Keras library for Neural Networks, StatsModel for the Time Series approach, Amazon Sagemaker for DeepAR, and Matplotlib is required to generate the plots for a graphical output of the results. Therefore the data is distributed in to multiple parameters and divided further in to columns using multiple StatsModel. Trend explains about the setting of ups and downs in ATM withdrawal depending on usual trends seen usually around specific times and dates of a month, week or yearly such as summer vacations when withdrawal is more due to people seeking tourism. Likewise Seasonality describes more about seasons or time specific to a region such as holidays or festival dates when money is more dependent on than usual. Seasonality is usually a greater bump as compared to trends when the rates are climbing slowly, but consistently and maintaining for a greater time periods than that of seasonality. Where as seasonality are a quick change in values for a small period of time, but are far consistent. Error measuring is for evaluation purposes where the methods or models implemented are checked for reliability and results. From our collection of data the Bank ATMs have a rather explicit schema explaining all sorts of requirements for a ATM placement such as ATM ID, ATM Events, Replenishment and Position of an ATM. These factors are put in to consideration as well for the training and testing purposes of a model.

3. Evaluation. The ARIMAX model exerts output from the tested values in terms of predicted and original values which are then used for the purposes of measuring the accuracy of a model and afterwards prediction. Evaluation purposes of a model. RMSE and SMAPE is used as evaluation matrix.

Benefits of the Project

The main beneficiary here would be the banks and their customers. Reimbursing or refilling any ATM cash machine requires time and effort specially of the bank teller, who is in charge of the ATM. Cash needs to be carried around in container wagons which may or may not be on a contract, nonetheless requires security, time management and continuous monitoring of the ATM, so that it does not run out of cash. In cases where the reimbursing is not done on time or the ATM runs out of cash both the bank and their customers are at loss here. The banks mainly because of loss of business and or dissatisfaction of their client and the client or the customer is not able to meet their demand. This is specially true on occasions as such Religious festivals, holidays, weekends and end of the month. Our algorithm takes in to account and caters all these factors on the basis of exogenous variables.
Further as stated, besides helping with refilling of ATM cash with the right amount and on time, this reduces the lag between carrying cash loads in transit vehicles and allowing banks to know at what time and date such transaction is predicted to know when a contract is to be assigned to the ATM teller and also maybe the contract service which carries out this operation.
This science can also help relate to real time spending/transaction of users based on trends and seasonality. It may be because of the present occasion or public dealings.
Further as stated, besides helping with refilling of ATM cash with the right amount and on time, this reduces the lag between carrying cash loads in transit vehicles and allowing banks to know at what time and date such transaction is predicted to know when a contract is to be assigned to the ATM teller and also maybe the contract service which carries out this operation.
This science can also help relate to real time spending/transaction of users based on trends and seasonality. It may be because of the present occasion or public dealings.

Technical Details of Final Deliverable

The final delivery will be a web based interface, powered purely on Python3. The site will be running the python scripts required for analysing, assessing, implementing and validating the models which will produce the graphical results and future predicted transactions on a sheet with which the user can interact. Django, a Python based web framework will be executing the whole scenario. This is what enables us to run fully developed models and obtaine results as Django compiles and executes Python scripts on the run in real time. 
On the welcome screen the user will be guided through the step by step process about obtaning results. At first the site asks for an input file in the form of an excel sheet and uploads it to the database from where our algorithm takes over and performs parsing/cleaning and make the sheet file ready for the model processing in the second step.
The whole sheet is loaded into memory once and further on no action is needed on the users behalf. In the second step where the user selects from a drop down list of ATMs ids available to predict for. The model is trained based on the ATM data the client preferred and validation is presented for realiability.
Finally the model is ready for prediction as per the processing explained in the implementation feature of this submission and the user selects the time frame or basically the limit to which the prediction is to be made.
On the welcome screen the user will be guided through the step by step process about obtaning results. At first the site asks for an input file in the form of an excel sheet and uploads it to the database from where our algorithm takes over and performs parsing/cleaning and make the sheet file ready for the model processing in the second step.
The whole sheet is loaded into memory once and further on no action is needed on the users behalf. In the second step where the user selects from a drop down list of ATMs ids available to predict for. The model is trained based on the ATM data the client preferred and validation is presented for realiability.
Finally the model is ready for prediction as per the processing explained in the implementation feature of this submission and the user selects the time frame or basically the limit to which the prediction is to be made.

Final Deliverable of the Project

Software System

Core Industry

IT

Other Industries

Finance

Core Technology

Artificial Intelligence(AI)

Other Technologies

Big Data

Sustainable Development Goals

Decent Work and Economic Growth, Industry, Innovation and Infrastructure, Responsible Consumption and Production

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com