PaperlayerView exact revision on arXiv

Source-linked AI summary

Mobile Money: Understanding and Predicting its Adoption and Use in a Developing Economy

Simone Centellegher, Giovanna Miritello, Daniel Villatoro, Devyani Parameshwar, Bruno Lepri, Nuria Oliver

arXiv:1812.03289v1physics.soc-phcs.CY

TL;DR

Financial institutions remain difficult to access for poor populations in developing economies, while quantitative evidence on mobile-money adoption and sustained usage is limited. The paper analyzes millions of pseudo-anonymized mobile-phone records and M-Pesa transactions using descriptive analysis and machine-learning models; adoption and spending are predictable from phone activity, social-network, agent, geographic, and mobility features, with AUCs of 0.691 and 0.619. The findings identify factors associated with M-Pesa usage and inform service-design implications within the study’s setting.

  • Problem

    Quantitative evidence is limited on which factors contribute to mobile-money adoption and sustained usage, despite mobile money’s role in extending financial services to underserved populations.

  • Method

    The paper analyzes millions of pseudo-anonymized mobile-phone communications and M-Pesa transactions, combining phone, agent, social-network, geographic, and mobility information in predictive models.

  • Results

    AUC=0.691 for mobile-money adoption and AUC=0.619 for mobile-money spending; the most predictive features relate to phone activity, M-Pesa users in ego-networks, and mobility.

  • Takeaways & Limitations

    Past mobile-phone behavior can predict M-Pesa usage three months ahead and distinguish high from low spenders, supporting differentiated marketing actions.

  • Takeaways & Limitations

    The authors caution that findings may not generalize across countries, user samples, or time periods, and note the random customer sample and three-month dataset gap.

Abstract

from arXiv · show

Access to financial institutions is difficult in developing economies and especially for the poor. However, the widespread adoption of mobile phones has enabled the development of mobile money systems that deliver financial services through the mobile phone network. Despite the success of mobile money, there is a lack of quantitative studies that unveil which factors contribute to the adoption and sustained usage of such services. In this paper, we describe the results of a quantitative study that analyzes data from the world's leading mobile money service, M-Pesa. We analyzed millions of anonymized mobile phone communications and M-Pesa transactions in an African country. Our contributions are threefold: (1) we analyze the customers' usage of M-Pesa and report large-scale patterns of behavior; (2) we present the results of applying machine learning models to predict mobile money adoption (AUC=0.691), and mobile money spending (AUC=0.619) using multiple data sources: mobile phone data, M-Pesa agent information, the number of M-Pesa friends in the user's social network, and the characterization of the user's geographic location; (3) we discuss the most predictive features in both models and draw key implications for the design of mobile money services in a developing country. We find that the most predictive features are related to mobile phone activity, to the presence of M-Pesa users in a customer's ego-network and to mobility. We believe that our work will contribute to the understanding of the factors playing a role in the adoption and sustained usage of mobile money services in developing economies.

1 INTRODUCTION

Mobile money services extend digital financial access through widespread mobile-phone networks, but quantitative evidence on adoption and sustained usage remains limited. This study analyzes large-scale M-Pesa data to identify usage patterns, predict adoption and spending, and examine predictive features.

  • Motivation: Approximately 2 billion adults worldwide are unbanked, with access to financial institutions especially difficult for poor populations in developing economies.Mobile money can bridge cash and digital economies through transfers, deposits, withdrawals, bill payments, and related services.
  • Motivation: Despite mobile money’s success, deployment is constrained by demographic and economic conditions, marketing, agent-network development, and regulation.These challenges motivate closer analysis of customer adoption and usage.
  • Study aims: The study analyzes 140 million mobile-phone records and more than 27 million M-Pesa transactions collected over six months in an African country.The analysis focuses on key drivers of M-Pesa adoption and usage intensity.
  • Contributions: The paper reports large-scale M-Pesa usage patterns, including service types and communication and money-transfer flows.This descriptive analysis is one of the paper’s three stated contributions.
  • Contributions: Two machine-learning models predict mobile-money adoption and spending three months ahead using phone data, agent information, social-network users, and geographic setting.The paper also examines the models’ most predictive features and implications for mobile-money service design.

2 RELATED WORK: MOBILE MONEY

Prior mobile-money research emphasizes implementation, regulation, financial inclusion, qualitative adoption studies, and analyses of mobile-phone data. The paper addresses the scarcity of large-scale quantitative models of mobile-money usage by combining M-Pesa transactions with multiple predictive data sources.

  • Implementation and policy: Prior work studies successful deployment, agent networks, marketing, consumer trust, financial risks, fraud, and regulations supporting mobile-money expansion.This literature also examines how services reach critical mass and promote financial inclusion.
  • User and design studies: Ethnographic, interview, qualitative, and HCI studies examine mobile-money adoption among low-income users and how service design affects customer behavior.These approaches provide user and interaction perspectives on mobile-money use.
  • Mobile-phone data: Call Detail Record studies use mobile-phone data to predict spending, infer credit scores, and estimate socioeconomic conditions such as wealth.One cited study estimated the wealth of 1.5 million customers from mobile-phone activity in Rwanda.
  • Research gap: Very few studies analyze large-scale mobile-money data while building quantitative statistical models of usage behavior.Existing work includes analyses of mobile-phone usage and social-network mobile-money users, plus adoption models across three developing economies.
  • This study: This paper extends prior work with millions of M-Pesa transactions and machine-learning predictions using phone data, agent information, social-network users, and rural-versus-urban activity.It presents the study as the first quantitative analysis using data from M-Pesa, described as the leading global mobile-money service.

3 M-PESA, A MOBILE MONEY TRANSFER PLATFORM

M-Pesa is a mobile-network payment and money-transfer service launched by Safaricom in Kenya in 2007. It supports deposits, transfers, withdrawals, airtime purchases, and bill payments through an extensive agent network, and operates across eight countries.

  • Platform origins: Safaricom launched M-Pesa in Kenya in 2007 as a payment and money-transfer service delivered through its mobile-phone network.M-Pesa stands for Mobile “Pesa,” using the Swahili word for money.
  • Services: After registration with official identification, customers can deposit money, transfer funds by SMS, withdraw cash, purchase airtime, and pay bills.The account is associated with the customer’s mobile phone.
  • Agent network: M-Pesa uses an extensive agent network distributed across the territory to enable cash deposits and withdrawals.Agents are part of the infrastructure required to operate the service.
  • Expansion: Following M-Pesa’s rapid growth in Kenya, other mobile operators launched competing mobile-money services.Deployment also requires marketing, service integrity, investment, and supportive regulation.
  • Geographic context: M-Pesa is offered in eight countries, including Kenya, India, Ghana, Tanzania, and Mozambique.The study’s country of interest has M-Pesa’s largest market share at 42%, a 47% estimated poverty rate, and about 12 million people living in extreme poverty.

4 DATA

The study links pseudo-anonymized mobile-phone metadata with M-Pesa transactions and supporting agent, location, and user-registration information. The datasets cover distinct periods and provide large-scale behavioral, geographic, and transaction records for analysis.

  • Data sources: The study uses pseudo-anonymized Call Detail Records and M-Pesa financial transactions, matching customers across datasets through consistent hashed identifiers.Encryption preserves personal privacy while allowing the same customer to be followed over time.
  • Datasets: Dataset D1 contains more than 140 million CDR events from 100,000 customers during November 2016–January 2017.Dataset D2 contains approximately 27 million M-Pesa transactions from more than 1.2 million randomly selected customers during April–June 2017.
  • CDRs: CDR metadata includes encrypted caller and callee identifiers, call direction, timestamp, duration, and routing cell-tower ID.SMS records include encrypted sender and recipient identifiers, outgoing-message type, timestamp, and routing cell-tower ID; incoming SMS was unavailable.
  • Geography: Cell-tower mappings provide latitude, longitude, and urban, suburban, or rural classification for each tower location.These data support geographic characterization of mobile activity.
  • M-Pesa data: M-Pesa transaction records include anonymized sender and recipient identifiers, transaction type, amount, timestamp, and routing cell-tower ID.The dataset also includes M-Pesa agent locations identified through their nearest cell tower and registration status for people making or receiving calls and SMS.

5 DESCRIPTIVE DATA ANALYSIS

The analysis describes M-Pesa transaction behavior, service diversity, and geographic flows using quantitative mobile-money data. Customers predominantly use few services, transfer relatively small amounts, and send P2P money across longer distances than their phone calls.

  • Transaction types: P2P transfers and utility-bill payments are the most frequent transaction types, with P2P transactions exceeding 30% of all transactions.P2P transfers are direct customer-to-customer money transfers and remain consistent with M-Pesa’s original money-transfer purpose.
  • M-Pesa services: More than 65% of customers use M-Pesa for one purpose, while 10% use two services.The analysis examines the number and types of services used in dataset D2.
  • M-Pesa services: 89% of transactions among single-purpose customers are Customer Transfers or Deposits at Agent Till, indicating saving or sending-money behavior.These customers are characterized as “sources” of money.
  • M-Pesa services: Less than 1% of single-purpose customers’ transactions pay utility bills, so utility payment is rarely their sole M-Pesa use.The finding concerns the subset of customers using M-Pesa for only one purpose.
  • M-Pesa services: Customer Withdrawal at Agent Till appears in 70% of transaction pairs among two-purpose customers, identifying this group as recipients of money.Withdrawal is the most common transaction in this subset and represents cash withdrawal from M-Pesa accounts.
  • Calls and money flows: 72.5% of M-Pesa P2P transactions occur between districts, compared with 63% of phone calls within districts, indicating longer-distance money flows.This comparison uses 3,182 customers with 46,478 phone interactions and 320 inter-customer P2P transactions.

6 PREDICTING M-PESA ADOPTION AND CUSTOMER SPENDING

The paper builds machine-learning models to predict future M-Pesa usage and expenditure from mobile-phone, social-network, mobility, agent, and geographic features. Mobile activity, M-Pesa contacts, and mobility are among the strongest predictors, while adoption and spending show distinct feature patterns.

  • 6.1 Feature extraction: 77 hand-crafted features describe mobile-phone behavior, M-Pesa agents, social networks, and geographic location for two prediction tasks three months ahead.The tasks predict future M-Pesa usage and customer expenditure.
  • 6.2 Classification tasks: 70,269 of 92,574 registered customers were inactive, while 22,305 were active based on whether they completed an M-Pesa transaction.The models used an 80% training split and a 20% test split with the same active-to-inactive ratio.
  • 6.2 Classification tasks: 0.691 AUC was achieved for predicting future M-Pesa usage, while 0.619 AUC was achieved for predicting customer spending.The usage model performed at parity with the state of the art; the spending result was averaged across 5-fold cross-validation test sets.
  • 6.3 Feature Analysis: For future usage, sent SMS count and active days were the strongest predictors, followed by M-Pesa contacts and mobility.The radius of gyration was especially relevant, while M-Pesa agent availability did not appear among the important features.
  • 6.3 Feature Analysis: For spending, the percentage of M-Pesa friends was the most important feature, while mobile activity and mobility also ranked highly.Greater mobility and in-degree were associated with higher likelihood of high spending, whereas more M-Pesa friends and longer incoming weekend calls were associated with lower likelihood.
  • 6.3 Feature Analysis: The spending model found a counter-intuitive negative association between the percentage of M-Pesa friends and high spending.The paper presents lower socioeconomic status as a possible explanation and identifies the hypothesis for future investigation.

7 DISCUSSION AND IMPLICATIONS

The study finds that mobile money usage is strongly associated with mobile phone activity, social-network presence, and mobility, while usage patterns support targeted service design. Predictive performance is moderate, and the findings have important country-specific limitations.

  • Usage patterns: More than 65% of customers used M-Pesa for only one purpose, while money transfers were the most frequently performed transactions.The authors identify sources and recipients of money and report that utility-bill payment alone is unlikely to drive M-Pesa use.
  • Usage patterns: M-Pesa transactions flowed over longer distances than phone calls, supporting the importance of transfers across long distances and between urban and rural regions.The paper links these patterns to money moving from urban, richer regions to rural, poorer regions.
  • Predictive performance: AUC=0.691 for adoption prediction, 38% better than the baseline, while expenditure prediction achieved AUC=0.619, 32% better than the baseline.The adoption model performed at par with the state-of-the-art; expenditure results could not be compared with prior work.
  • Predictive features: Mobile phone activity, M-Pesa users in customers’ ego-networks, and mobility were the most predictive feature groups.Examples include SMS sent, active days, and radius of gyration.
  • Agent network: Agent distance and density did not appear among the top 15 predictive features, possibly because coverage was already sufficiently dense or location features were approximate.The paper notes that agent access remains necessary for withdrawals and deposits, which comprise more than 30% of transactions in the datasets.
  • Implications: The models may help providers identify potential customers, anticipate consumption, and understand drivers of mobile money adoption and usage.The authors describe these findings as having business value for mobile money service providers.
  • Scope and limitations: The findings should not be assumed to generalize across countries, user samples, or time periods because prior CDR-based models did not necessarily transfer across countries.The authors leave investigation of broader generalization to future work.

8 CONCLUSION AND FUTURE WORK

The paper analyzes anonymized mobile communications and M-Pesa transactions to describe usage patterns and predict adoption and usage intensity. It concludes that mobile phone behavior, social-network characteristics, and mobility help predict future M-Pesa use, while acknowledging sampling and temporal limitations.

  • Study design: The study analyzes millions of anonymized mobile phone communications and M-Pesa transactions from an African country.It combines M-Pesa data with data from a leading mobile phone operator.
  • Study design: The authors report aggregate M-Pesa usage patterns and build machine-learning models using mobile phone, agent-network, social-network, and geographic features.The models predict M-Pesa adoption and intensity of usage.
  • Limitations and future work: The study has limitations including a random customer sample and a three-month gap between the two datasets.Future work includes replication in another country, real-world model testing, and further study of agent and social-network effects.
Loading 1812.03289v1…