Mobile money service enables people to use their phones to carry out financial transactions, which includes person-to-person funds transfer, bill payments, merchant payments, sending and receiving remittances, along with many other use cases. Customers who have bank accounts can top-up their mobile money account directly from their bank, and can similarly withdraw mobile money balance to their bank.
Unbanked or underbanked customers, however, rely on a network of agents to carry out top-up or withdraw their mobile money account. These agents may be small scale traders, shop owners, large scale merchants, or employees of the mobile money service provider, authorized by the service provider to provide last mile services to unbanked or underbanked customers. Agents cannot just help customers in cash-in or cash-out, but can also register new customers and carry out payments (P2P or P2M) on behalf of the customers, and receive commissions for each of these transactions.
In order to ensure that they can perform the above-mentioned transactions for the customers, mobile money agents are provided with a wallet account by the mobile money service provider, and agents can request e-money stocks (e-float) from distributors or wholesalers in exchange of cash based on their requirements.
Agent’s e-float also gets replenished whenever a customer requests cash-out (withdrawing cash in exchange of mobile money), for which the agents need to maintain sufficient level of cash with them. Mobile money agents do not prefer keeping high stocks of e-money or cash because of high cost of capital as well as security reasons.
As a result of this, they often witness shortage of e-money or cash at their store to carry out transactions for the customers until they rebalance their e-money/cash inventory. This not just results in loss of commission for the agent, but also a potential loss of revenue for the service provider because many of those customers use mobile money accounts by different providers. Also, the cash and e-money inventories are inter-related in the sense that one gets replenished every time the other is extracted.
Therefore, each day’s requirement of cash and e-money can also be fulfilled by much lower starting levels of cash and e-money for that day. For example, starting with 120 worth of e-money and 140 worth of cash, an agent can also perform transactions worth 500 that day, based on the occurrence of those transactions.
This study is done to suggest an approach for determining the optimum cash and e-money requirement for each agent outlet every day. This study in done to figure out what type of data will one require and what model will work the best with the data.
The main task here is to determine/forecast the right amount of inventory i.e. the cash as well as the ideal amount of e-float balance which must be present so that there are least stockouts of cash and e- float, minimum loss and higher returns. My approach to the problem is using time series ARIMA modelling technique to predict the amount of cash and e-money balances which should be present for a day with an agent using the historical data.
The ARIMA (p, q, d) model stands for Auto Regressive Integrated Moving Average. There are multiple variants of the ARIMA models, for a seasonally varying time series data SARIMA (Seasonal Auto Regressive Integrated Moving Average) can be used which takes few more input parameters. Similarly, there is ARFIMA (Auto Regressive Fractionally Integrated Moving Average) which can used when long-memory processes are required it is same as ARIMA only that the d here can take up fractional values.
The steps given below describe the way of approach
1. Exploratory data analysis on our time series data of the balances of both cash and e-float for a frequency of a day/week/month (sample period of 6 months is taken in the reference paper) of a particular agent to find out seasonality of the data.
2. The differencing required to make the data stationary can be found out that is the order d. Unit root tests like ADF test (Dickey-Fueller test) can be done to check whether the data is stationary.
3. We look at the ACF and PACF plot’s properties (decay, geometry, nature... Etc) of the stationary time series data to estimate p, q the AR and MA order respectively.
4. Run diagnostic tests on the estimated model.
5. Find out the model which is the best fit after looking at various indicators like the MAPE (Mean Absolute Percentage Error), the behavior of the residuals of the models, the value AIC (Akaike's Information Criterion) and many others.
The dataset should have cash_balance, efloat_balance along with the timestamp (date of the month) for each day before the agent begins and ends his/her transactions for the day. Because we here are concerned about predicting the inventory for a single day itself. The dataset should essentially contain such details for 6 months or a year. For ARIMA to be efficient 100-200 data points are enough meaning the data for 100-200 days is a must.
The objective of this study is to predict the demand for e-money and cash each day at the agent outlet by analysing historical patterns. The main application of ARIMA is in the area of short-term forecasting requiring at least 40 historical data points. It works best when your data exhibits a stable or consistent pattern over time with a minimum number of outliers.
I feel that using other neural network algorithms is a great idea at the same time algorithms like LSTM (Long Short-Term Memory) which is a class of RNN (recurrent neural networks) is preferable and more efficient when you have a huge dataset and lot of input parameters. Moreover, from my experiences I can say that neural network models are nothing but black boxes it all depends on how you train it to behave.
The ground rule in forecasting is comparing the efficiency of the models and choosing whichever is the best. Therefore, ARIMA and LSTM are the two most suitable approaches for predicting the inventory.
Inventory Management for Mobile Money Agents in the Developing World by Karthik Balasubramanian, David F. Drake, and Douglas Fearing, Harvard Business School Publications
Author: Aniket Biswal