Enhancing Banking Stability: Leveraging Artificial Neural Networks to Prevent Non-Performing Assets

Banks play very useful and dynamic role in the economic life of every modern state. They are an important constituent of the financial market. They have control over a considerable part of the stock of money and their lending and investing activities cause changes in the quantity of money in circulation which in turn influences the nature and character of production in any country. In fact, a country’s economic progress is judged and determined by the progress of its banking system. To quote Bhaba (1956), “banking is the kingpin of the chariot of economic progress. As such its role in expanding economy of a country like India can neither be underestimated nor overlooked.”

The winds of freedom and change sweeping across the financial sector due to reforms have thrown open several challenges to the banks and financial institutions. The major challenges relating to introduction of prudential norms, technological up-gradation and transparency, improving operational efficiency in the deregulated interest rates, reduction of non-performing assets (NPAs) for survival in the competitive market and product innovation. (Reddy et. al., 2002; Ghosh, 1995). One of the major causes of low profitability is the high presence of NPAs in Indian banking. Going by the number of committees, expressions of concern, papers, programmes, etc. perhaps the issue of NPA has become problem number one continuously for about a decade in the banking industry. (Chhimpa, 2002).

One of the major hurdles Indian banking is facing today, is its ever-growing size of NPAs over which the top management of almost each bank is baffled. On account of the intricacies involved in handling the NPAs, the ticklish task of assets management of the bank has become a tight rope-walk affair for the controlling heads, because a little wavering ‘this or that side ‘may land the concerned bank in trouble (Banmali, 2001). Gujral (2003) described NPAs problem as one of the greatest and the most formidable problems that has shaken the entire banking industry in India like an earthquake. Narsimham Committee report of 1998 has also pointed out that NPAs of the banks constitute a real economic cost to the nation as the scarce capital is locked up in unproductive use and consequently adversely affects the recycling of funds with the banks.. The spiraling and the devastating effect of NPAs on the economy has made this problem an issue of public debate and of national priority as well. Therefore, any measure or reform on this front would be inadequate and incomprehensive, if it fails to make a dent in NPAs reduction and stall their growth in future, as well (Gujral,2003).

Managing NPAs is perhaps the greatest task ahead of Indian bankers as well as planners. Over the years many strategies have been chalked out to control its menace, but resulted in limited success. Although total elimination on NPAs is not possible in banking business owing to externalities but their incidence can be minimized. It is always wise to follow the proper policy for appraisal, supervision and follow-up of advances to avoid non-performing assets (Reddy et.al.,1999 and Naik,2001; Muniappan ,2002;. Sood,2001; Kulkarni,1999; Singh,2002; Rao, 2002; Bose, 2005) .To control NPA menace, preventive measures would be necessary. One of the preventive measures is credit appraisal and credit audit. Credit appraisal usually suffers from failures to : a) assess promoters ability to adapt to change, understand industry and market and raise adequate margins, b) forecast sale, and c) monitor end use of funds, cash flow etc. Documentation of credit policy, credit audit immediately after sanction and human resources development through training interventions are some of the measures necessary to upgrade quality of credit appraisal in banks (Taori,2000).Therefore, the key to handle the huge volume of NPA lies with the efficient decision making at the branch as well at the corporate level. In this context the implementation of adequate Information System is highly essential and desirable as banking industry in India is on the accelerated path of computerization (Godse, 2002).The focus in this book is on the development of Information Technology driven decision system to manage the new loan assets, which is to be sanctioned by the banks. Configured ANN Models(Artfiical Neural Networks) tryto overcome these problems and provide support to decision makers in Banking field.

The methodology is based on the results of the extensive survey of financial analysts, from the academic and banking field, as well as on the international and national literature concerning Non–Performing Assets. The aim is initially to identify and select the criteria (financial ratios and qualitative criteria) that are the most appropriate to use in the evaluation of NPA, in a second stage, to achieve the modeling of the selected criteria and its estimation, and in a third stage representation of knowledge based on modeling. Finally there is development of neural network followed by validation and refinement of the network .

It is found that majority of the factors identified through the primary survey match the factors collected from the secondary sources. However, some new factors like borrowers dealing with the bank, experience in the same line of business, repayment period, no of installments, no of years of experience in business management, track record of repayment of loan , net sale to total loan ratio, general administrative expenditure to total expenditure ratio etc. are also surfaced during the primary survey.

A technique of numerical representation of knowledge is presented by modeling the selected criteria variables. The test data (i.e. Applicant Data) is given a number representation using the modeling criteria and estimation process. Two neural networks have been trained to classify this data. This results into a technique for solving classification problems. A classification procedure applied to twenty-six criteria variables is also developed. It is found that the Perceptron network only isolates NPA and PA cases. As a refinement the Probabilistic Neural network is used to isolate the doubtful cases as well. Whereas in the Perceptron network the error percentage is found to be high this gets reduced considerably in the PNN.

The artificially generated random data set is used in the absence of realistic training data set. It is felt that the performance of the neural network will be significantly improved if realistic data sets are used for training the network. In this case network will approximate an expert. A practical limitation in training the network is that one could not use huge training data sets (this could be an inherent limitation of MATLAB or system resources).It may be possible that if we train the network in a batch mode the error level may significantly dropped.

It is possible that in future the credit granting process may become more complex and other parameters will be found to be more important in establishing the credit worthiness of the loan applicant. In that case the number of criteria variables may significantly increase .The classification process i.e. the finding the NPA and PA applicants will remain the same. However, other fields may be incorporated by modifying the system design and software by following the outline given in the work. For creating the database Ms-Access is used where database security is minimum as it is a DBMS. Whereas the RDBMS such as Oracle, SQL Server has more data security features. In realistic situations we recommend that these databases should be used so that Applicant confidentiality is maintained. However, with certain minor changes this program which is written in VB6.0 will work with any RDBMS.

It is suggested that both network should be run as it is not possible to say a priori whether a case is doubtful or not. It is possible that by increasing the layers of the perceptron accuracy of the network can be increased. However, we have not examined this possibility because increasing the number of layers will require higher processing speed cpu and more memory which may not available in all banks. Further the time to train/setup the network will be significantly increased. Using realistic training data sets can also reduce the error level. Further; this technique can be generalized to solve other classification problems such as investment decisions, risk management, financial analysis, etc.