To Reduce Gross NPA and Classify Defaulters Using Shannon Entropy

Ambarish Moharil, Nikhil Sonavane, Chirag Kedia, Mansimran Singh Anand

Abstract

Non Performing Asset (NPA) has been in a serious attention by banks over the past few years. NPA cause a huge loss to the banks hence it becomes an extremely critical step in deciding which loans have the capabilities to become an NPA and thereby deciding which loans to grant and which ones to reject. In this paper which focuses on the exact crux of the matter we have proposed an algorithm which is designed to handle the financial data very meticulously to predict with a very high accuracy whether a particular loan would be classified as a NPA in future or not. Instead of the conventional less accurate classifiers used to decide which loans can turn to be NPA we build our own classifier model using Entropy as the base. We have created an entropy based classifier using Shannon Entropy. The classifier model categorizes our data points in two categories accepted or rejected. We make use of local entropy and global entropy to help us determine the output. The entropy classifier model is then compared with existing classifiers used to predict NPAs thereby giving us an idea about the performance.

Introduction

In the banking sector, a Non-Performing Asset (NPA) represents a significant financial risk, typically defined as a loan where interest or principal payments are overdue for more than 90 days. These assets drain bank liquidity and can destabilize broader economic systems. Predicting which applicants are likely to default is a complex task due to the "noisy" and imbalanced nature of financial datasets, where successful loans often vastly outnumber defaults. Traditional machine learning models frequently struggle with this skewness. This research proposes a novel classification approach inspired by information theory, specifically utilizing Shannon Entropy to measure the "disorder" introduced by new data points into established pools of successful and failed loans.

Methodology

The proposed system utilizes Shannon Entropy to quantify uncertainty within the dataset. The core logic rests on the "Difference of Entropy Metrics" (DEM). The methodology follows a structured pipeline:

  1. Pool Creation: Training data is partitioned into two distinct pools: $C^+$ (non-defaulters) and $C^-$ (defaulters).
  2. Local Entropy Calculation: A baseline entropy, referred to as the Reference Entropy Metric ($\alpha$), is calculated for each attribute within these pools.
  3. Global Entropy Evaluation: When a new loan application is processed, it is hypothetically added to each pool. The system then recalculates the entropy to find the Final Entropy Metric ($\beta$).
  4. Classification Decision: The model calculates $DEM = \alpha - \beta$. A positive DEM indicates that the new data point aligns well with the characteristics of that specific pool, causing minimal "entropy disturbance."
Workflow of the proposed entropy-based classification system

Data and Preprocessing

The study utilized a substantial loan dataset comprising 532,428 cases, with a 3:1 ratio of defaulters (406,601) to non-defaulters (125,827). To handle this imbalance, the entropy pools are treated independently, preventing the majority class from overshadowing the minority. Preprocessing involved expanding the feature set from 45 to 88 columns and employing Natural Language Processing (NLP) via SpaCy and Linear SVM to convert applicant descriptions into binary sentiment scores. Numerical data was further discretized using quantile binning to facilitate entropy calculations.

Experimental Results

The performance of the Entropy-based classifier was benchmarked against several standard industry models, including Random Forest, Support Vector Machines (SVM), and Decision Trees. The proposed model demonstrated superior performance, achieving an accuracy of 98.24%.

Accuracy comparison of various classifiers

As shown in the comparison, the entropy model outperformed the next best classifier (Random Forest) by over 11%. This suggests that measuring system disturbance is a highly effective way to identify patterns in skewed financial data.

Conclusion

By treating classification as a measure of information uncertainty rather than simple boundary detection, the Shannon Entropy-based model provides a robust solution for NPA prediction. While the model requires careful categorization of continuous data and may involve higher computational overhead for global entropy recalculations, its high accuracy makes it a compelling tool for modern risk management in banking.

Summary updated

Abstract

Non Performing Asset(NPA) has been in a serious attention by banks over the past few years. NPA cause a huge loss to the banks hence it becomes an extremely critical step in deciding which loans have the capabilities to become an NPA and thereby deciding which loans to grant and which ones to reject. In this paper which focuses on the exact crux of the matter we have proposed an algorithm which is designed to handle the financial data very meticulously to predict with a very high accuracy whether a particular loan would be classified as a NPA in future or not. Instead of the conventional less accurate classifiers used to decide which loans can turn to be NPA we build our own classifier model using Entropy as the base. We have created an entropy based classifier using Shannon Entropy. The classifier model categorizes our data points in two categories accepted or rejected. We make use of local entropy and global entropy to help us determine the output. The entropy classifier model is then compared with existing classifiers used to predict NPAs thereby giving us an idea about the performance.
BibTeX
@misc{moharil2020reduce,
  author = {Ambarish Moharil and Nikhil Sonavane and Chirag Kedia and Mansimran Singh Anand},
  title = {{To Reduce Gross NPA and Classify Defaulters Using Shannon Entropy}},
  year = {2020},
  eprint = {2004.14418},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2004.14418},
}
Figure