There’s a constant refrain we keep hearing in lending – that it is a business of skirting risk. Mitigating credit risk is tricky because a number of factors go into the creation of risk profiles of individual and business borrowers.
The cost of inaccuracies in risk assessment, along with sup-bar underwriting, inefficient portfolio monitoring and poor collection models, is extremely high – in 2020, Citigroup was fined $200 million for credit risk deficiencies. As a result, lenders are continuously devising new ways to mitigate credit risk and improve their profitability.
Artificial intelligence (AI) and machine learning (ML) models of credit assessment can help analyse large data volumes more efficiently because these assessment models are capable of learning from complex data sets and become more accurate incrementally.
Beyond logistic regression
The probability of default, an essential parameter in the calculation of regulatory capital under the Basel framework for international banking standards, is traditionally measured through the logistic regression model. This method measures the probability of an event occurring as a binary outcome. It tells you whether or not a certain profile poses the possibility of defaulting.
However, now that there is an exponential increase in the availability of data and advanced storage capacities, several ML models have been developed to incorporate this big data and detect patterns with more precision. In fact, logistic regression, a traditional statistical model, has also been redeveloped for machine learning.
AI/ML models are more flexible compared to statistical models, in that they are capable of analysing hard-to-detect relationships between variables. Through ML-based models like random forest, gradient boosting, and stacking methods, this data can be leveraged for enhanced accuracy.
Early warning signals
Credit risk management is incomplete without an early warning system. High-risk borrower segments can be identified long before defaults occur using a large number of indicators such as negative cash flows, high-interest rate loans and GST and MCA filings and more. Here, artificial intelligence can be used to identify patterns from large volumes and high velocity of data from an array of sources with increased accuracy.
Early warning signals can be generated using AI tools like predictive analysis, natural language processing (NLP) and clustering. NLP, for instance, helps analyse text from device data and is used to segment borrowers into high-risk, medium-risk and low-risk buckets.
Transactional fraud prevention
Lenders can pivot away from rules-based systems in fraud detection to machine learning ones. Through ML, lenders can adjust the existing rules and learn new ones as more data is processed. This occurs through both supervised and unsupervised learning.
Supervised learning is based on annotated historical data where instances of fraud are labelled to help identify existing fraud patterns. Under unsupervised learning, the system learns from data sets that haven’t been labelled as such yet. From this data, it seeks out relationships between variables that may not be apparent to humans.
A combination of both approaches allows lenders to analyse transactions holistically. It can recognize patterns and flag fraudulent activity in real time without the need for human intervention.
Challenges of mitigating credit risk through AI
The biggest problem posed by using AI for credit risk mitigation is that these models tend not to be explainable. They have been described as ‘black boxes’, since it is difficult to explain the relationship between the inputs and outputs in a machine learning model. For this reason, machine learning models have often invited the scrutiny of regulators.
This lack of transparency makes ML-based credit risk models less auditable – which exposes them to regulatory scrutiny. For instance, the absence of hard-coded rules to prevent discrimination against borrowers on the basis of race or gender may occur as the result of widespread use of ML models that incorporate atypical data if not policed properly.
However, efforts are now being made to mitigate this risk posed by ML models by introducing improved transparency, explainability, and auditability. Lenders can assess the depth of the credit risk model by measuring the prediction performance of the model and the factors that influence its performance and individual predictions.
They must also verify the ‘global’ aspect of model explainability, viz the systemic biases that are present in the data. Alongside this, the ‘local’ aspect – eg, the rationale behind its assessment of individual credit applications – must also be explained.
FinTech is at the vanguard of refining ML models
Though the adoption of ML-based risk assessment models in lending is on the rise among traditional lenders, FinTechs have been able to build on the finer points of the game through in-house data intelligence arms. This keen focus on data intelligence has helped them refine their ML models in many ways.
FinTechs act as aggregators of risk-related data as they cater to diverse portfolios. They test feedback data from across these variegated sample sizes, which makes their models more resilient in the face of macroeconomic uncertainties.
Moreover, FinTechs’ undivided attention on the data intelligence function allows them to revise their models periodically, which sets in motion a cycle where successful challenger credit scores continuously replace the prevailing scores, creating a durable credit risk model.
Here’s the good news – FinTechs are capable of extending the superior risk assessment capabilities of their ML models to lenders through a wide array of plug-and-play products. Lenders can outsource these capabilities to FinTechs in order to gain exponential value from AI/ML in risk assessment.
(This article is written by Anant Deshpande, Co-founder, FinBox, and the views expressed in this article are his own)