Following the 2008 financial crisis, regulators have tightened the leash on financial institutions in terms of credit risk to protect the economy from another crash. In addition to stricter regulations, poor credit risk management is one of the greatest factors affecting financial institutions’ ROI.
Credit decisioning is a complex process that relies on several elements, including the accuracy of data. Manual decision-making to some extent is still based on “gut feel” and is prone to errors that can be costly to the organization. Moreover, the need for human intervention to review every single credit request greatly increases the decision time and customers are forced to wait for an approval.
AI based automation of the credit decisioning process improves operational efficiencies and lets your organization optimize the credit approval and credit amount decision, based on a comprehensive process involving data analysis, modelling and decision trees.
Adastra’s AI Driven Credit Decisioning and Risk Management Solution
Adastra’s AI-driven credit decisioning solution takes a stepwise approach that allows financial institutions to make faster, better informed decisions that maximize profitability and mitigate risk.
The process can be broken down into four discrete steps:
Data Preparation – This step involves optimizing the weight of evidence (WoE) and conducting an informational value (IV) analysis. The process starts by creating a data source lineage and data flow mapping. The data is then cleaned and processed, and numerical variables are divided into classes and the WoE and IV values are calculated.
Based on the information value of a predictor, we can assess the efficacy of a variable in determining an individual’s creditworthiness. This process does not account for the interaction between the variables and is, therefore, purely descriptive. However, this step is crucial for identifying the features to exclude for the multivariate analysis. The WoE values determine the directionality of relationship between the variables and an individual’s creditworthiness.
Profiling – This step involves a correlation analysis to determine the relationship between the variables and creditworthiness of an individual. A pairwise regression helps us understand how closely variables relate linearly to another, so that correlated variables that can potentially result in spurious conclusions from the regression model can be identified.
Credit Approval – The credit approval process first requires selection of a variable subset. When faced with correlated variables, we can either filter them out based on a suitable metric, use business logic to remove them from consideration or perform a stepwise regression to incrementally remove variables that have negative coefficients in the WoE analysis. Next, a scorecard is created using modelling techniques like logistic regression, Lasso & Ridge regression with stepwise selection or gradient-boosted decision trees. Grid search or parameter sweeps are frequently used to determine the best parameter options for different modelling techniques, and cross-validation helps ensure maximum model generality.
Model assessment is done via AUC and Gini throughout development and on profitability post-development. The model is assessed to determine revenue potential and profit rate for all previously approved and rejected applications. The probability of default is calculated based on insights from the model and the global default rates across hold out sample. The process also helps determine the optimal threshold for application approval which maximizes profitability.
Credit Amount Decision – The financed amount needs to be calculated based on the customer’s financial information, such as their debt-to-income ratio, etc. Decision trees are used for a consistent, automated rule-based model for each client. These recommend optimal credit limits, guided by past performance and the institution’s risk appetite. Separate models are typically used for different customer demographics.
This process significantly shortens the decision time for credit approval and amount. By minimizing human intervention and guesswork, the margin of error is reduced, and the model is optimized to maximize the financial institution’s profitability, keeping well-within the acceptable range of risk.
Interested in implementing a credit risk solution? Contact our experts today.
Book a Free Consultation
We will contact you as soon as possible.