Explore the evidence.
Understand the reach.

Financial institutions predict the probability of default when making lending decisions using conventional credit risk scoring models. However, in emerging economies, the underbanked, widows, small and medium enterprises owners and youth are not able to access traditional forms of collateral or identification required by financial institutions due to lack of data for them to get access to loans. Financial institutions cannot obtain much of the information required about an applicant and they use alternative data sources such as public data or social media to deal with the problems of information asymmetry, adverse selection and moral hazard. To eliminate such problems, credit risk assessment models must be built on the foundation of artificial intelligence and machine learning approaches in an effort to perform an accurate credit risk analysis, to properly assess the behaviour of the customers and subsequently perform a thorough verification of the potential of the applicant to repay the loan thereby allowing less privileged people to access credit. However, big data sourced from the internet and public data sources is characterized by huge volume, variety, veracity, the curse of dimensionality, class imbalance, concept drift and non-linearity among others. These challenges affect the generalization performance of credit risk assessment models used by most financial institutions. This paper proposes an efficient and effective credit risk assessment model capable of learning incrementally based on machine learning algorithm called Adaptive Heterogeneous Dynamic Ensemble Selection (AHDES) that uses big data from alternative sources for financial for the underbanked. The algorithm is agile and adaptive to unexpected world events, changes in customer behaviour and cognitively counter bias likely to be introduced by information asymmetry. Experimental results on four large credit datasets and five evaluation metrics show that our proposed algorithm performs better than other selected benchmark models allowing the underbanked and less privileged to access loans to stimulate economic growth.

This article maps the knowledge structure and evolution of bankruptcy/default prediction research in financial institutions using a bibliometric co-citation approach, which identifies works that are often cited together due to the proximity of their theories, methods, or themes. Bibliographic data were collected from the Web of Science (WoS) for the period 1970-2025, then extracted in a complete bibliographic format and prepared for network analysis. The co-citation procedure is carried out by extracting the reference list, normalizing and unifying citation variations, and preparing a co-citation matrix, which is then visualized and classified using VOSviewer with minimum threshold parameters and normalized association strength. The results of the document ranking based on the number of citations and Total Link Strength (TLS) show the development trajectory of the field from the ratio and Z-Score approach to the logit probabilistic model, the market foundation based on the structural theory of credit risk, the governance perspective supported by agency theory, the dynamic panel methodology/GMM for longitudinal data, to computational waves (neural networks and machine learning) to capture nonlinear patterns of credit scoring. The co-citation map uncovers four main clusters: econometrics & banking risk, structural theory, ratios, and computational, with Merton and Altman acting as the cross-cluster binding nodes. These findings support the design of a more comprehensive early warning system (EWS) by integrating ratio signals, market indicators, governance/agency factors, and ML models to improve accuracy and adaptation.    

Financial institutions predict the probability of default when making lending decisions using conventional credit risk scoring models. However, in emerging economies, the underbanked, widows, small and medium enterprises owners and youth are not able to access traditional forms of collateral or identification required by financial institutions due to lack of data for them to get access to loans. Financial institutions cannot obtain much of the information required about an applicant and they use alternative data sources such as public data or social media to deal with the problems of information asymmetry, adverse selection and moral hazard. To eliminate such problems, credit risk assessment models must be built on the foundation of artificial intelligence and machine learning approaches in an effort to perform an accurate credit risk analysis, to properly assess the behaviour of the customers and subsequently perform a thorough verification of the potential of the applicant to repay the loan thereby allowing less privileged people to access credit. However, big data sourced from the internet and public data sources is characterized by huge volume, variety, veracity, the curse of dimensionality, class imbalance, concept drift and non-linearity among others. These challenges affect the generalization performance of credit risk assessment models used by most financial institutions. This paper proposes an efficient and effective credit risk assessment model capable of learning incrementally based on machine learning algorithm called Adaptive Heterogeneous Dynamic Ensemble Selection (AHDES) that uses big data from alternative sources for financial for the underbanked. The algorithm is agile and adaptive to unexpected world events, changes in customer behaviour and cognitively counter bias likely to be introduced by information asymmetry. Experimental results on four large credit datasets and five evaluation metrics show that our proposed algorithm performs better than other selected benchmark models allowing the underbanked and less privileged to access loans to stimulate economic growth.

This article maps the knowledge structure and evolution of bankruptcy/default prediction research in financial institutions using a bibliometric co-citation approach, which identifies works that are often cited together due to the proximity of their theories, methods, or themes. Bibliographic data were collected from the Web of Science (WoS) for the period 1970-2025, then extracted in a complete bibliographic format and prepared for network analysis. The co-citation procedure is carried out by extracting the reference list, normalizing and unifying citation variations, and preparing a co-citation matrix, which is then visualized and classified using VOSviewer with minimum threshold parameters and normalized association strength. The results of the document ranking based on the number of citations and Total Link Strength (TLS) show the development trajectory of the field from the ratio and Z-Score approach to the logit probabilistic model, the market foundation based on the structural theory of credit risk, the governance perspective supported by agency theory, the dynamic panel methodology/GMM for longitudinal data, to computational waves (neural networks and machine learning) to capture nonlinear patterns of credit scoring. The co-citation map uncovers four main clusters: econometrics & banking risk, structural theory, ratios, and computational, with Merton and Altman acting as the cross-cluster binding nodes. These findings support the design of a more comprehensive early warning system (EWS) by integrating ratio signals, market indicators, governance/agency factors, and ML models to improve accuracy and adaptation.