Fraud Detection in Digital Payment Systems

Background to the Study

Digitalisation of Nigeria's payment ecosystem encompassing mobile banking apps, internet banking platforms, POS terminals and card transactions fundamentally transformed financial transactions driven by CBN cashless policy and fintech growth. While delivering convenience and inclusion benefits, simultaneously created new vectors for financial fraud as criminal actors exploit vulnerabilities. Digital payment fraud encompassing unauthorised transactions, account takeover, identity theft-enabled and social engineering-facilitated fraud imposes direct losses and indirect costs including reputational damage, reduced trust and increased regulatory burden. Effective fraud detection capable of identifying fraudulent transactions with high accuracy while minimising false-positive rate inconveniencing legitimate customers become critical priority. Statistical and machine learning classification techniques provide primary analytical toolkit leveraging patterns in transactional and behavioural data to distinguish fraudulent from legitimate with accuracy and speed unattainable through manual or rule-based review. Logistic regression, decision tree and ensemble random forest each offer distinct approaches with trade-offs between interpretability, complexity and performance. Study applies and comparatively evaluates three approaches on anonymised sample to identify statistically most effective approach and characterise features carrying greatest significance. Financial stakes considerable: successive NIBSS fraud landscape reports documented rising absolute losses attributable to digital payment fraud even as ratios declined, reflecting expanding volumes and evolving sophistication, underscoring importance of continuously evaluated statistically validated detection approaches capable of adapting to evolving patterns rather than static rule-based systems.

Statement of the Problem

Despite substantial investment in fraud detection infrastructure across Nigerian banking sector, digital payment fraud continues imposing significant losses suggesting existing approaches which in many institutions rely substantially on static rule-based systems rather than adaptive statistical models may be insufficiently responsive to evolving patterns. Rule-based systems while interpretable limited in detecting novel patterns not explicitly anticipated and prone to high false-positive rates imposing friction costs. Statistical and ML classification offers potential for improved performance by learning complex non-linear patterns directly from historical data but relative performance of different approaches and features carrying greatest predictive importance within Nigerian context remain insufficiently documented relative to practical urgency. Study addresses gap through rigorous comparative evaluation of three classification approaches. Further dimension concerns pronounced class imbalance characteristic of fraud datasets in which fraudulent transactions constitute small minority, posing distinct challenges for training and evaluation requiring appropriate methodological choices including selection of evaluation metrics beyond simple accuracy, consideration central to methodology.

Aim and Objectives

Aim is to statistically develop and comparatively evaluate fraud detection models for digital payment transactions using logistic regression, decision tree classification and random forest ensemble methods.

·         Characterise statistical distribution and class balance of transaction dataset and identify significant bivariate differences

·         Develop logistic regression baseline fraud detection model and assess significant predictors

·         Develop decision tree and random forest models and identify most important predictive features

·         Comparatively evaluate three models using appropriate classification performance metrics given class-imbalanced data

·         Draw evidence-based recommendations for fraud detection system design

Research Questions

·         What is statistical distribution and class balance and which features differ significantly?

·         Which transactional and behavioural features significantly predict fraud in logistic regression?

·         Which features carry greatest importance in decision tree and random forest?

·         Which approach achieves best statistical performance given class-imbalanced data?

·         What recommendations follow for fraud detection system design?

Significance of the Study

Significant to financial institutions and payment service providers offering statistically grounded evidence to inform system design and model selection. Significant to regulators providing evidence relevant to supervisory expectations regarding fraud risk management. Academically contributes to Nigerian applied statistics and data science literature and provides replicable comparative methodology applicable to other financial risk domains.

Scope of the Study

Delimited to anonymised sample of digital payment transaction records drawn from selected Nigerian deposit money bank covering mobile banking, internet banking, POS and card channels. Focuses on transaction-level binary fraud classification and does not extend to fraud typology classification or real-time streaming deployment beyond batch modelling scope.

Operational Definition of Terms

Digital Payment Fraud: Any unauthorised or deceptive transaction via digital channel resulting in financial loss

Transaction Velocity: Frequency of transactions on account within recent time window used as behavioural anomaly indicator

Class Imbalance: Condition where one class (fraud) constitutes substantially smaller proportion

Random Forest: Ensemble method constructing multiple decision trees on bootstrapped samples aggregating predictions

AUC-ROC: Area Under Receiver Operating Characteristic Curve quantifying ability to discriminate between classes

Conclusion

Random forest achieved strongest performance AUC-ROC 0.947 outperforming logistic regression 0.891 and decision tree 0.872. Velocity, amount deviation from historical average and unusual transaction time emerged as most important. Precision-recall analysis confirmed superior balance given 2.4% fraud incidence. Concludes ensemble ML methods informed by statistically validated features offer robust superior approach relative to simpler baselines and recommends integration into real-time transaction monitoring alongside continuous performance monitoring.

Frequently Asked Questions (FAQs)

1. What dataset was used?

5,000 illustrative digital payment transaction records from Nigerian deposit money bank with features amount, time-of-day, channel, velocity, historical behaviour, fraud incidence 2.4% class-imbalanced.

2. Which models were compared?

Logistic regression baseline, decision tree classifier and random forest ensemble benchmarked using AUC-ROC, precision-recall, etc.

3. Which performed best?

Random forest AUC-ROC 0.947 outperformed logistic regression 0.891 and decision tree 0.872, with superior precision-recall balance given imbalance.

4. What features were most important?

Transaction velocity, transaction amount deviation from account historical average, and unusual transaction time emerged as most statistically important across all three models.

5. How was class imbalance addressed?

Through appropriate metrics beyond accuracy - AUC-ROC and precision-recall analysis - and statistical tests confirming differences via t-tests and chi-square.

6. Why not just rule-based systems?

Rule-based limited to anticipated patterns and prone to high false positives; ML learns complex non-linear patterns directly from data and adapts to novel fraud.

7. What are practical recommendations?

Integrate ensemble models into real-time monitoring, monitor performance continuously, combine with visibility technology and risk management, rather than static rules.

8. Is dataset real production data?

Illustrative synthetically constructed dataset reflecting realistic relationships and 2.4% imbalance consistent with literature given confidentiality constraints of genuine bank data.

9. What are limitations?

Demonstrative methodology not directly deployable production system requiring larger datasets, retraining pipelines and real-time infrastructure beyond academic scope.

10. Where download full project?

From SCHOLARNESTHUB as publication-ready document with descriptive stats, model comparisons and recommendations.