Loan Collection using Neural Network

by Jul 29, 2016Financial Services, Success Stories

Objective

To output an ordered score which would discriminate the potential good accounts (accounts which have high likelihood of regular repayment) from the bad accounts (low propensity to repay) for the current and early delinquent (bucket 1 i.e.<30 DPD and bucket 2 i.e. <60 DPD) accounts. These scores were to be used for prioritization of collection actions/ efforts.

success-finance6

Methodology

  • All transactional tables (billing, payment, delinquency/ bounced cheque details etc) were transformed to prepare a unique customer level dataset so that every customer has a single record with all monthly information populated against it
  • Joined profile data, transactional data and other relevant data sources to prepare a single snapshot of the data
  • Data preparation steps (data cleaning – missing value treatment, incomplete/ invalid value treatment etc.) were performed to prepare the dataset for model development
  • Features creation performed- based on statistical results (correlation study, multivariate analyses, sensitivity analysis etc.) and business knowledge, various features were created for model building
  • Based on various feature selection procedures (stepwise linear regression, clustering results, correlation, cross-tab method etc.), a subset of 25 features were selected to build scorecard prototypes.
  • A series of Neural Network scorecards were developed on a combination of these 25 features. Models were developed on training datasets and validated on validation datasets. Out of time validation (out of sample) was also performed.
  • All scorecards’ performances were compared against each other on various technical and business objectives- 1. How well it is separating ‘Goods’ from ‘Bads’, 2. Statistical validation results on models’ robustness and consistency, 3. Are the important features and their weights in agreement with business logic? Basis these results, the best performing model was selected.
  • Neural Network Model output was implemented into client’s CRM system. The system scores all relevant accounts on first day of every billing cycle. Portfolio is then divided into segments of customers for targeted collections treatment for the rest of billing cycle.

 

Impact

  • The collections team now targets only top 20% accounts with highest scores (90% of potential defaults are in top 20%). This optimizes collection budget usage and minimizes delinquencies.
  • Collection expenses have been reduced by 10% as well as monthly default rates have gone down by 6% within six months of implementation.

Related Posts

[et_pb_dfbm_blog dfbm_fullwidth=”off” article_distance=”20″ item_animation=”faded” custom_posttypes=”post” include_post_categories=”248″ posts_number=”3″ post_content_background_color=”#f9f9f9″ content_absolute=”on” content_absolute_event=”off” thumb_same_height=”on” define_thumb_height=”300″ header_tag=”h3″ show_content=”excerpt” show_limit_words=”on” show_comments=”off” show_limit_categories=”on” show_author=”off” show_date=”off” read_more=”off” show_pagination=”off” use_inner_shadow=”off” disabled_on=”off|off|off” _builder_version=”3.19.1″ post_body_line_height=”1.5em” post_header_font=”Montserrat||||||||” post_header_font_size=”17px” post_header_line_height=”1em” post_meta_font=”Montserrat||||||||” post_meta_text_color=”#000000″ post_meta_font_size=”11px” post_meta_line_height=”1.2em” post_meta_link_font=”Montserrat|300|||||||” post_meta_link_text_color=”#000000″ post_meta_link_font_size=”11px” post_meta_link_line_height=”1.2em” post_content_line_height=”1em” text_orientation=”left” custom_css_filt_blogs_post_meta=”font-weight: 200;” custom_css_filt_blogs_post_meta_a=”font-weight: 200;” saved_tabs=”all” /]