Neural Network Analysis of the Credit Screening Decision Using MatLab
Gary Cao
December 19, 1996
Executive Summary
In this project, we have two raw data-sets in the credit screening decision process.
The credit screening data-set has 590 observations; the credit application data-set has 100
observations. Both of them have 15 attributes. In addition, the screening set has one class
attribute indicating the decision (yes or no). This project is only a simple example of the
actual problem which may has more than several thousand observations and more
attributes.
We plan to use Matlab and the Neural Network concept to solve the problem.
The objective is to utilize the computer systems’ capacities and let them assist us in the
credit screening decision-making process.
The procedure includes the following steps:
(1) to train the computer system to learn the pattern of credit screening decision made by
human intelligence;
(2) to minimize the difference between the computer decisions and human decisions on the
same data set; and
(3) to let the trained system make the decision for us on the new data set.
We used four Matlab programs to analyze the data and have the following results:
(1) linear network, 10 positive class attributes and 90 negative class attributes;
(2) log-sig network with trainbp, all negative class attributes for the 100 new
observations;
(3) log-sig network with trainbpx, all negative class attributes for the 100 new
observations;
(4) log-sig network with trainlm, without definite results.
Statement of the Problem
We have a data-set of 590 observations with 15 attributes and 1 class attribute.
We have another data-set of 100 observations with the same 15 attributes but no class
attribute. We plan to ask the computer system to learn the pattern on decision-making
process (criteria, or some logical and/or causal relationship between the 15 attributes and
the class attribute) on the 590 observations, and make the decisions for us on the 100 new
applications.
Description of Raw Data
The raw data-set for credit screening can be described in the matrix equation:
Y0 = function ( X0 )
Y0 = Class Attributes (one vector, two possible values as 1 or 0);
X0 = Attributes (15 vectors, continuous, nominal with small numbers of values, and
nominal with large numbers of values).
If the function is linear, then
Y0(1x15) = A(1x15) + B(1x15) * X0(15x15)
where Y0 and X0 are known as the original screening data.
From the 590 observations, we try to solve for A and B, and then plug A and B
into the new equation and solve for Y1, where
Y1(1x15) = A(1x15) + B(1x15) * X1(15x15)
and
X1 is known as the 100 new observations of the 15 attributes.
Transformations and the Reason
We have to transform the data in order to solve the problem. The computer can
only recognize a specific signal when it process the data.
For example, the first attribute A1 can have two values (“a” and “b”), but the data-
set may have missing attribute values, therefore A1 can actually have three possible
values: “a”, “b”, and “value not available at this time”. We transform A1 into a one by
three (1x3) matrix as
100=a
010=b
001=not available at this time
Based on the above reason and criterion, we transform the 15 attributes into 51
primary attributes that can be recognized and processed by computer.
Linear Network: Explanation and Discussion of Results
The Matlab program is attached in Attachment D.
The result is: out of the 100 new applicants, 10 have positive class attributes. This
indicates an acceptance rate of 10%, well below the class distribution assigned by the
Professor (44.5% positive).
Two possible explanation exist.
(1) The linear model does not perfectly fit the original screening data-set. We adjusted
different parameters to minimize the prediction errors, but still we have
SSE = 58.2512
Max(SSE) = 1.2
Sum of Absolute values of Prediction Difference (errors) = 74
(2) The new application data-set may be fundamentally different from the original
screening data-set.
Log-sig Network: Explanation and Discussion of Results
The Matlab program is attached in Attachment D. For the log-sig network with
three different patterns (trainbp, trainbpx, trainlm), we find that the third one (trainlm)
does not yield any result due to computer capacity limits, and the first two have the same
result: none of the 100 new applicants get a positive class attribute.
Two possible explanation exist.
(1) The log-sig network model does not perfectly fit the original screening data-set. We
adjusted different parameters to minimize the prediction errors, but still we have
SSE = 147.493
Sum of Absolute values of Prediction Difference (errors) = 293
(2) The new application data-set may be fundamentally different from the original
screening data-set.
Competitive Learning: Explanation and Discussion of Results
We were not able to complete this part of the project because we do not
understand the procedure.
Summary and Explanation of decision
Based on our preliminary experience on neuron network analysis, we believe that
this approach to utilize computer systems in decision-making process has a great potential.
It will dramatically decrease the cost of operation and increase efficiency in data
processing and decision making.
However, the models available do not yield satisfactory results. One possible
reason is that the models are too simple to handle the extremely complicated data-sets.
If we use the linear network in the credit screening process, the difference between
the predicted values and the known target values is still large (sum of errors is 74 out of
590 observations, a 12.5% error rate).
Attachments:
1. Raw data for credit screening information
and Prediction using Neural Network tools
2. Raw data for new credit applicants
3. Predictions for the 100 new credit applicants Matlab programs
3a. general linear
3b. general trainbp
3c. general trainbpx
3d. general trainlm.
This page hosted by
Get your own Free Home Page