Decision support system on bad debt recovery in the deposit protection fund board - Kenya

Abstract

The Deposit Protection Fund Board (DPFB), in particular Institutions in Liquidations division, is often plagued by a lengthy liquidation process, unpaid loans, collection agency fees and various legal charges. The unpaid loans and the recovery procedures contribute significantly to the rising cost and increasing length of the liquidation process. The DPFB does not have any decision support tools that can be use to guide on debt recovery for institutions in liquidation. They rely on manual methods such us response to demand notes, the existence of security and the general availability of documents on the loan to classify whether a loan is a potentially good or bad debt. Institutions in liquidation have massive loans data, which have been utilised in this project to aid the learning process of data mining tools in evaluating whether a particular debtor is likely to pay their debts. These results are meant to act as an enhancement to the mentioned manual methods. This research explored the effectiveness of various data mining tools in evaluating whether a debt is likely to be repaid. The research involved recognition of the manual methods DPFB uses in classifying debts, the IT measures taken by DPFB to enhance debt recovery and the level of success achieved so far. Loans data was collected from 27 institutions’ databases. Data was prepared by selecting suitable variables (predictive and target). The predictive variables were four, namely; amount at liquidation, contacts availability, debt type and customer type. The target variable was the indication on whether a debt is good or bad. Seven data mining methods were selected based on guidance from literature review. Predictive modelling software – DTREG was employed to train and validate the data mining techniques. The performance of each of the model was analysed using confusion matrix and area under receiver operating curve. Results were obtained from both balanced and imbalanced data. Balanced data performed better than imbalanced. The results which are detailed in chapter four depict that neural networks tools generally gave high accuracy. These findings guided to the development of a DSS prototype based on neural networks. This model can be used to aid DPFB decision support on debt recovery. A neural network based DSS software - Alyuda Forecaster XL whose details a highlighted in section 4.5 formed the Model Based Management System (MBMS) component of the developed DSS prototype. On testing the software using different number of debt records, it was noted its classification accuracy increased with higher number of records. The Data Base Management System (DBMS) part of the DSS was based on SQL 2000 while the Dialog Generation and Management System (DGMS) was based on Visual Studio and Crystal Report software. In conjunction with the manual methods currently used by DPFB to classify debts, the developed decision support system is expected to enhance the accuracy of classification that may eventually lead to curbing the lengthy liquidation period and reduce expenses incurred in the process.