prologue
The global financial crisis of 2007 had long-lasting effects on the economies of many countries. Many people lost their jobs, savings, etc. in the epic financial and economic collapse. When too much risk is confined to a very small number of players, it is considered a flagrant failure of the risk management framework. Since the global financial crisis, terms such as ‘bankruptcy’, ‘default’ and ‘financial distress’ have grown rapidly in popularity. The process of forecasting distress/default/bankruptcy is now considered an important tool for making financial decisions and a major contributor to credit risk. The Great Depression of 1932, the Suez Crisis of 1956, the International Debt Crisis of 1982, and most recently, the 2007 recession, have pushed the world to make sense of the prospect of collapse as soon as possible.
Source: Image piki superstar upon free pick
Thanks to the excitement surrounding financial technology (FinTech) and rapid advances in AI, financial crisis and bankruptcy prediction is one of the applications that has received a lot of attention in recent years. Credit scores can be used by financial institutions to decide whether or not to give credit to borrowers, increasing the chances of denial of high-risk borrower requests. In addition to the problems caused by noisy and highly imbalanced data, we have to deal with new regulatory reforms such as: General Data Protection Regulation (GDPR)Regulators expect model interpretability to exist to ensure that algorithmic findings are understandable and consistent. As the COVID-19 pandemic get hurt To avoid situations like 2007, we need a better understanding of lending in financial markets.
Balance: Interpretability and Accuracy
Credit scoring or bankruptcy prediction models that predict whether an individual will return a lender’s money or whether a company will file for bankruptcy must address high accuracy and interpretability. Two friends of hers from the same college in the United States applied for a mortgage with roughly the same salary, the same age, and other comparable parameters. One was approved for a loan, but one of hers was rejected. Don’t you think the rejected person deserves an answer? What if a company applies for a loan for expansion, but the bank decides not to provide the loan?
interpretation
With the growth of several relevant independent variables (Sentiment scores from financial statements and ratios), revealing obvious interpretation difficulties. Previously, a minimal number of independent variables and a basic model were sufficient for easy model interpretation. As a result, studies aimed at selecting the most important variables and simulating bankruptcy based on selected features and simple statistical models have become very popular. , is another way to handle functions that are not necessarily limited in number.
Source: Image Guo et al. (2019) – research gate
Accuracy
Traditional feature selection-based techniques and machine learning-based approaches have strengths and weaknesses. Feature selection-based approach is easily interpretable This is because the context of this article uses a small number of variables that have been chosen to be important in predicting bankruptcy/default events. Feature selection-based approaches often use basic predictive models such as multivariate functions. However, the accuracy is much lower when compared to machine learning-based models. In contrast, machine learning-based approaches achieve higher accuracy, but such models are too complex to be easily understood.
Explainable AI for Humans – XAI
Source: Image by Caltech
Explainable artificial intelligence (XAI) attempts to simplify black-box models and make them easier to interpret. This allows humans to understand and trust the output of machine learning algorithms. Discuss the accuracy, transparency, and results of AI-powered decision-making models.Much-needed explainability and transparency helps organizationsapproach AI differenceThe success of machine learning has sparked a flood of AI applications. However, the effectiveness of these systems is limited because machines cannot explain their actions to humans. As AI advances, humans struggle to understand and track algorithmic results.
Source: Image by freepik above free pick
When presented with an ML-based model, use XAI techniques to audit the model and try its paces to deliver consistently accurate results across a variety of use cases related to credit scoring and distress forecasting You can check if you can. These techniques, for example, evaluate the prediction rules of the model to check their consistency with previous information about the model. business problemThis helps uncover challenges that may hinder model accuracy when applied to out-of-sample data. A problem with the dataset used to train the model. incomplete or biased Expressions of populations of interest or training conditions that lead the model to learn incorrect prediction rules can be uncovered by the XAI technique.
Post-Corona, XAI has made great strides to address credit-related business issues.
- Zest Artificial intelligence-based technology allows lenders to fine-tune their models for fairness by mitigating the impact of discriminatory credit data without compromising performance.
- FICO has publicly stated its intention to complement human domain understanding with machine intelligence to enable rapid development of highly predictive and explainable credit risk scoring models.
- Orcorus Free lenders from biased datasets and enable them to analyze financial data more efficiently. Ocrolus software analyzes bank statements, payslips, tax forms, mortgage forms, invoices and other documents for loan eligibility for mortgage, business, consumer, credit scoring and KYC determine gender.
- Underlight.ai used nonlinear algorithmic modeling to estimate lending risk in regions with little or no credit bureau use.
- Temenos Launched a transparent XAI model delivered as SaaS to help banks and credit unions speed up digital onboarding and loan processing to address post-COVID-19 challenges for the lending industry.
Model: Modern vs. Traditional
The latest methods of applying machine learning models to credit scoring to predict defaults in the financial sector have undoubtedly yielded: Improved predictive performance It outperforms traditional models such as logistic regression.New digital technologies have accelerated the adoption of machine learning models, making them available for financial institutions to acquire and use larger dataset It consists of many features and observations for model training. Machine learning approaches have the potential to experimentally discover nonlinearities in the relationship between outcome variables and their predictor variables, and the interaction effects between the latter, in contrast to traditional models such as the logit model. I have. As a result, if the training dataset is large enough, ML models are likely to outperform traditional statistical models routinely used in credit assessment. As a result, improved default forecast accuracy can benefit lending institutions by reducing credit losses and creating savings related to regulatory capital.
Role of regulators
financial regulators recently expressed strong support for having financial institutions use both traditional and machine learning models simultaneously and assigning humans to handle significant discrepancies. Most financial regulators do not restrict where or how black boxes can be used by financial institutions.Nevertheless German Federal Financial Supervisory Authority We encourage institutions to consider the benefits of using a more complex model and document their reasons for choosing an interpretable alternative. In the words of many policy makers, there is no universal method for assessing costs and benefits. They suggest that banks consider the model’s justification, deployment context, and goals before implementing it. The patented work was done by Equifax, which uses neural networks to analyze consumer credit risk and provide reason codes to help businesses meet their regulatory obligations.
Source: Image by freepik above free pick
Actionable Insights for Customers
The XAI methodology for credit-related models readily accommodates binary consumer decisions such as “lend” or “not lend”. The explanations provided by the XAI methodology may not take into account important aspects of lending, such as interest rates, repayment schedules, credit limit changes, or customer loan preferences. However, financial institutions using XAI must inform their customers why their loan application was declined and how to improve their credentials. A customer might be offered a raise in salary or education, or asked to pay their credit card bills on time for several months to get loan approval. This is nothing but actionable information that customers can use to reapply for a loan.
Isn’t success linear?
Many financial institutions are currently investigating inherently interpretable algorithms, linear regression, logistic regression, explainable gradient boosting, neural networks, and more. Consistently, there is a desire to investigate new techniques for developing models that are transparent by design and do not require post-explanation. Explain pre-built models after the fact Another part of the story. For example, a 200-layer deep neural network or black box model passes input to an explanation algorithm. This breaks a complex model into smaller, simpler pieces that are easier to understand. These simpler outputs also consist of a list of features and parameters that are important to the business problem. In both scenarios above, Trade-off between high accuracy and interpretability the need for time.
Source: Image by Ajitesh Kumar vital flux
SHAP (SHApley Additive exPlanation) and LIME (Local Interpretable Model-agnostic Explains) are widely used explainability approaches. SHAP uses the Shapley value to score the functional impact of your model. The Shapley value considers all feasible predictions using all inputs. SHAP’s exhaustive approach ensures consistency and local accuracy. LIME, on the other hand, builds a sparse linear model around each prediction to describe how the black box model behaves locally.
More importantly, which is more relevant: an accurate model or one that is easily understood by the business as well as the customer. If the data are linearly separable, inherently interpretable and an accurate model. However, if your data is complex and the decision boundaries are not straight, you may need to look at complex models to make sure they are correct before thinking about ex post facto explanations.
Conclusion
In an ideal world, explainability would help people understand how the model works and when to use it. Lenders, regulators, and governments should work together to develop AI guidelines that protect the interests of their customers. If different groups do not articulate their explainability goals, AI will benefit the few, as it did during the 2007 global financial crisis. A well-designed XAI solution considers stakeholder needs and customizes the presentation. AI and machine learning algorithms are already lending institutionThe main points about XAI are:
- Model Accuracy/Performance and Interpretability – XAI aims to make AI models more intuitive and understandable without sacrificing accuracy/performance.
- Implementing and Scaling XAI – Mysteries have long prevented lending institutions from fully leveraging AI. However, XAI has helped institutions with more than just smooth onboarding and personalization.
- Responsible AI Transparency – How the credit model arrived at its results and what parameters it uses.
- Informativeness (giving humans new information to help them make decisions), and uncertainty estimation (capturing how uncertain an outcome is) (quantifying the confidence of a prediction).
