Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11In a 2018 paper, Deepak Kumar Gupta and Shruti Goyal compared a feed-forward artificial neural network (ANN) with linear regression to classify loan applicants as likely to default or not. They reported almost identical results for the two models. The paper is a small study—not a current benchmark or evidence that either approach is ready to guide real lending decisions—and its descriptions of the dataset conflict.
What Gupta and Goyal’s paper studied
Gupta and Goyal’s paper, published online on 8 May 2018 in the International Journal of Modern Education and Computer Science, frames credit risk as the possibility that a borrower will not repay a loan. Its task is binary classification: assign an applicant to a default or non-default class. The authors compare an ANN with linear regression, presenting the work as a creditworthiness prediction exercise rather than a validated lending system. The journal record identifies the article as volume 10, number 5, pages 9–16; the full paper gives its methods and results.
What data and inputs did the paper describe?
The paper’s dataset descriptions do not agree. Its introduction refers to a small dataset of residential mortgage applications from a bank, while its conclusion says the models were trained on loan-lending data provided by Kaggle. The article does not clearly resolve whether these are the same data or different descriptions, so the dataset cannot be identified with confidence from the paper alone.
The authors list loan-related inputs such as loan amount, funded amount, term, interest rate, installment, and grade. These are examples of features described in the paper, not a verified or complete data dictionary. A secondary page describes a 60,000-record random sample from a Kaggle LendingClub dataset said to exceed one million records, but that sample size is not confirmed in the primary journal record; it should not be treated as established study detail.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →#1 Best Overall
How the ANN and comparison were set up
The methods section describes a feed-forward network with 10 input variables, seven hidden layers, and a single-neuron output classifier. The authors say they selected numeric or integer fields, normalized values with a min–max linear transformation, and split the data into training and test sets. They also refer to a back-propagation model and compare the neural-network approach with linear regression.
Those descriptions do not amount to a complete reproducibility protocol. The paper does not clearly document an externally validated evaluation, and its account of the data is inconsistent. As a result, the architecture and preparation steps help explain the experiment but are not enough to reproduce or independently validate its performance.
Rank #2
What results did the authors report?
Gupta and Goyal reported the following error and accuracy figures. They described the results as approximately the same and noted that the mean squared error (MSE) depends on the training/test split.
| Measure | ANN | Linear regression |
|---|---|---|
| Mean squared error | 0.0220449, reported by Gupta and Goyal (2018) | 0.0227334, reported by Gupta and Goyal (2018) |
| Accuracy | 97.67575%, reported by Gupta and Goyal (2018) | 97.69609%, reported by Gupta and Goyal (2018) |
On these reported figures, linear regression had marginally higher accuracy, while the ANN had marginally lower MSE. The paper does not establish that either difference is meaningful beyond its particular experiment. Accuracy alone also does not show how well a model distinguishes defaulters from non-defaulters or what kinds of classification errors it makes. The reported percentages should not be read as a general expectation for credit decisions.
Rank #3
What the comparison does—and does not—show
It does not establish a universal winner
The reported metrics are close, and the authors’ abstract says model choice depends on the application and attributes. Their note that MSE varies with the training/test split is especially relevant: performance figures from one split do not establish how a model will behave on new applicants or in another lending context.
Interpretability is a trade-off
The authors describe the ANN as a black box and say its outcomes are harder to explain than those of linear regression. For lending, that matters because a score may need to be understood and justified, not merely calculated. The paper raises this consideration but does not assess regulatory compliance or demonstrate that either model meets the requirements of a real lending program.
Preparation and validation matter
Feature selection and normalization are part of the authors’ method, not incidental details. But the paper’s inconsistent dataset account and limited evaluation description mean readers cannot tell from the reported metrics alone how robust the results are, whether the test split represents future applicants, or how the models would perform in deployment. No independent reproduction or real-world validation is established by the sources.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to use the paper as a reader
For a personal-finance reader, the paper is best understood as an example of comparing two modeling approaches, not as a guide to how a bank currently evaluates an individual application. Its results support a narrow conclusion: in this study, the authors reported similar performance for an ANN and linear regression, with a trade-off in interpretability. They do not establish that neural networks generally predict default at roughly 98% accuracy or that an algorithmic score determines any particular borrower’s outcome.
Recommended Free Tools
Best Value
Readers interested in the broader field can look for Credit Scoring and Its Applications by Lyn C. Thomas, Jonathan N. Crook, and David B. Edelman, which the paper cites. The article’s reference does not establish current edition or availability.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
