One post tagged with "predictive modelling"

Objective​

A Good Customer Risk Assessment Model is to keep track of borrower risk during the tenure of the loan, post disbursement by monitoring risk profile change of a borrower periodically.

A robust model can enable a lender keep track and understand adequately potential changes in risk profile of a borrower.

This can further lead to variety of benefits in terms of optimization of various operations including customer relationship, collection process optimization, cross-sell/up-sell efforts and portfolio risk concentration and thereby minimizing financial losses due to lost capital.

Challenges​

Microfinance Lenders face quite a few challenges in achieving a robust customer risk assessment model. But the most important ones can be divided into following sub categories:

• ** Availability of Quality Primary Data **

  • Large segment of customers are first time borrowers with no credit history
  • Lack of Demographics data including Income, Saving, Wealth data since most of the customers are bottom of the pyramid and come under priority sector/income generation loan.
  • Lack of industry wide framework or data for Group Risk Assessment either by center/area/pincode.
  • Unavailability of detailed customer payment data of loans taken from other MFI/Institution in a cost efficient manner to source periodically.

• ** Inefficient Data Management **

  • In-ability to keep data quality from internal/external quality sources sanitized/validated month on month in a repeatable way impacts model quality over time.
  • Single source of truth for all customer features from across various data sources. (App, LOS, LMS, Field Feedback Systems, Past Payment Pattern, Credit Bureau data etc)
  • Challenges of data size/growth
  • Avoid cumbersome and inefficient spreadsheet based tools or developer scripts.

• ** Model Trust, Explainability, Communiation and Presentability **

  This is a significantly under-rated challenge not understood by most companies deploying a model like customer risk assessment. Building trust on the model outcome can be time-consuming and at times draining.

  • Taking all stakeholders along the process of model thinking and establishing trust on model needs ability to produce reproducible results, ability to explain factors leading to a particular customer or group level risk assessment in simple explainable way. Inability to go through this phase leads to model not being deployed.

  • Ability to iterate through and help explain the business well and eliminate any bias/errors.

  • Ability to provide model outcome/updates periodically directly through BI and Data Visualization in the hands of relevant stackholders.

• ** Automation by impacting relevant lending processes impacted by customer risk assessment outcome **

  • Integrate with various other IT systems and/or impact various downstream processes such as collection, customer relationship, cross-sell/up-sell efforts and portfolio strategy as required.

• ** Model Deployment **

  • Keep periodic results of models informed to various stakeholders
  • Sync outcomes with Other IT systems and field force facing systems.

Technologies used at an MFI​

Cynepia Technologies flagship product, Xceed Analytics, a unified end to end data & AI platform came make it easy for micro-lenders to overcome some of the above challenges. Key capabilities of Xceed analytics that come handy for the above include seemless data integration across various source IT system, Data Management including version management/governance capabilites, Data Workflows enabling data preparation/transformation, Data Visualization capabilities including Dashboards providing Visualization and Advanced Predictive/ML Modelling capabilities supporting 35+ algorithms and model catalog supporting version management and explainability.

Above capabilties can enable Microlenders achieve the following key implementation benefits important for above model.

• ** Data Workflow Automation ** allowed for seamless automation enabled customer to build Customer 360, which brings in 200 plus loan and customer level attributes from various source systems including manual uploads, Credit Bureau Data, Loan Origination System(LOS), Loan Management System, Other Customer Facing IT Systems as well as other upstream data workflows including Customer Payment Behavior Analysis for existing loans.

• ** Data Catalog ** with readily available feature profile enabled Data modellers to quickly explore and bring features needed for this effort together as well as keeping track of data quality month on month.

• ** Cynepia AutoML Subsystem ** enabled the following:

  • Automatic feature engineering, feature scaling and hyper-parameter tuning using Bayesian Optimization Techniques,
  • Automatic selection of algorithm is made of the various available options including Logistic Regression, Random Forest, Logistic Regression, CatBoost, Adaboost and X Gradient Boost (xgboost)
  • Model Catalog helped keep track of model versions and results over months enabling comparison of model results across revisions, validation of results by monitoring new delinquencies month on month as well as signs of model deterioration leading to retraining.
  • Model explain-ability using model agnostic approaches and Shapley values.

• ** Xceed Analytics BI Subsystem ** Insights communicated with Credit/Business stackholders.

• ** API Integration ** with Credit Facing Internal Application using ** Xceed Catalog APIs **.

Business Benefits​

• Availability of Robust Risk Assessment for existing customers.
• Model/Data Catalog and Operations ensures audit-ability of model and model outcomes over time..
• Integrated data and visualization infrastructure from Xceed Analytics enabled ease of building and presenting the model outcome to relevant stackholders seemlessly without need for elaborate manual processes avoiding manual spreadsheets.
• Integration with credit team facing application enabled collection of field feedback on specific risk assessment providing a feedback loop.

Integrations, use case & future possibilities​

Lenders can further redesign and integrate model outcome with source systems via ** API integration ** and further acheive end to end automation of such a model

• Real time integration of customer risk with next cycle loan decision and cross sell/up sell
• Use of generated KPIs in other customer related outcomes such as portfolio risk concentration etc.
• Redesign of collection processes based on customer risk profile.
• Redesign of cross-sell/upsell process based on understanding of customer risk profile.
• Portfolio Optimization/Reallocation based on future risk concentration thereby acting in advance.