pexels
One of the problems of emerging markets and low-income countries is the high cost of credit and credit ratings. It contributes to financial discrimination of borrowers such as small and medium-sized enterprises and households. Automating credit decisions based on analyzing big data can significantly improve efficiency of the process. The IMF's working report notes that advent of financial technology to the banking sector can accelerate issuance of loans and reduce their cost by increasing coverage. The study is aimed at economists and recommends them to study machine learning technologies (ML). Besides, the paper explains the basic principles of technology for non-specialists in detail.
The traditional methodology for assessing creditworthiness is based on five parameters - ability to borrow, capital structure, collateral, nature of the loan and its conditions. Machine learning can improve quality of evaluation of several of them and scoring in general. At the same time, ML models are able to highlight data patterns that cannot identify standard econometric models. But when using complex algorithms, such patterns cannot be easily verified by analysts and, although they provide a more efficient use of data, can generate erroneous results based on insufficiently relevant information. For example, in case of rapid structural economic changes that do not fully reflect in the data set, blindly using ML models can lead to outweighing outdated information and incorrect estimates of the credit risk of borrowers.
Both traditional econometrics and ML technologies are based on evaluation of models explaining links between indicators and output parameters, which should facilitate decision-making process. However, classical econometrics seeks to find formal relationships between them. In turn, ML models are set to compile the most accurate prediction within the existing sample without trying to explain the structure of cause-effect relationships.
The IMF states that the main advantages of using ML-models for scoring is economic feasibility of working with small borrowers. In addition, the largest technology companies with a large amount of user data are arriving in the financial sector, banks can use data on users' business activity accumulated as a by-product from provision of e-commerce, payments or telecommunications services for scoring and thereby further increase financial inclusion. In addition, ML models can structure unstructured information, which makes it possible to include a broader data flow in the evaluation process, and also increase efficiency of evaluation of non-linear relationships between risk drivers and output risks. This solves the problem of information asymmetry when banks overestimate requirements for borrowers, fearing inaccurate data on their part.
However, IMF analysts warn that the cheaper process of issuing loans has a downside. The disadvantage of ML is possibility of issuing inaccurate estimates when the initial data is “polluted”. This may entail “machine-based” discrimination of quite creditworthy borrowers. In addition, ML-models cannot quickly take structural changes into account (which is especially important for dynamically developing economies), and users can manipulate data - for example, “twist” activity indicators in social networks. Use of such data may threaten rights of consumers themselves - in case of discrimination of the latter on the basis of a computer solution that is inexplicable in human logic.
source: imf.org
The traditional methodology for assessing creditworthiness is based on five parameters - ability to borrow, capital structure, collateral, nature of the loan and its conditions. Machine learning can improve quality of evaluation of several of them and scoring in general. At the same time, ML models are able to highlight data patterns that cannot identify standard econometric models. But when using complex algorithms, such patterns cannot be easily verified by analysts and, although they provide a more efficient use of data, can generate erroneous results based on insufficiently relevant information. For example, in case of rapid structural economic changes that do not fully reflect in the data set, blindly using ML models can lead to outweighing outdated information and incorrect estimates of the credit risk of borrowers.
Both traditional econometrics and ML technologies are based on evaluation of models explaining links between indicators and output parameters, which should facilitate decision-making process. However, classical econometrics seeks to find formal relationships between them. In turn, ML models are set to compile the most accurate prediction within the existing sample without trying to explain the structure of cause-effect relationships.
The IMF states that the main advantages of using ML-models for scoring is economic feasibility of working with small borrowers. In addition, the largest technology companies with a large amount of user data are arriving in the financial sector, banks can use data on users' business activity accumulated as a by-product from provision of e-commerce, payments or telecommunications services for scoring and thereby further increase financial inclusion. In addition, ML models can structure unstructured information, which makes it possible to include a broader data flow in the evaluation process, and also increase efficiency of evaluation of non-linear relationships between risk drivers and output risks. This solves the problem of information asymmetry when banks overestimate requirements for borrowers, fearing inaccurate data on their part.
However, IMF analysts warn that the cheaper process of issuing loans has a downside. The disadvantage of ML is possibility of issuing inaccurate estimates when the initial data is “polluted”. This may entail “machine-based” discrimination of quite creditworthy borrowers. In addition, ML-models cannot quickly take structural changes into account (which is especially important for dynamically developing economies), and users can manipulate data - for example, “twist” activity indicators in social networks. Use of such data may threaten rights of consumers themselves - in case of discrimination of the latter on the basis of a computer solution that is inexplicable in human logic.
source: imf.org