Applying Artificial Neural Neteorks To Investemnt Analysis

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Applying Artificial Neural Networks to Investment Analysis

Author(s): George S. Swales Jr. and Young Yoon


Source: Financial Analysts Journal, Vol. 48, No. 5 (Sep. - Oct., 1992), pp. 78-80
Published by: {cfa}
Stable URL: https://www.jstor.org/stable/4479584
Accessed: 18-05-2019 11:50 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

is collaborating with JSTOR to digitize, preserve and extend access to Financial Analysts
Journal

This content downloaded from 193.19.172.190 on Sat, 18 May 2019 11:50:32 UTC
All use subject to https://about.jstor.org/terms
Applying Artificial Neural Networks to Investment Analysis

George S. Swales Jr., forecasting abilities. Forecasts of stock price appreciation) and
Associate Professor, Finance investment performance, based market valuation data from the
on quantitative and qualitative Fortune 500 and Business Week
and General Business
variables, have achieved limited "Top 1000" listing of firms.7 We
Department, Soutbwest
success, although methodological used two separate sets of data
Missouri State University, and statistical problems remain.2 from these sources. The first set
and The limitations of multiple dis- consisted of firms in the five in-
Young Yoon, Assistant criminant analysis (MDA) suggest dustries from the Fortune 500
Professor, Computer Infor- that an alternative, based on non- listings that offered investors the
mation Systems Depart- linear approaches, may better as- highest total returns each year;
ment, Southwest Missouri sist analysts and investors in mak- the sample consisted of 58 com-
State University ing decisions. Artificial neural panies. The second set consisted
networks offer such an alterna- of firms in the 10 industries re-
tive. ported by Business Week to have
Artificial neural network the highest market valuations; 40
technology can be used to A nonlinear artificial neural net- companies were included in this
work can address very complex set.
differentiate between stocks
problems.3 In particular, neural
that perform well and those net systems can accommodate We classified the firms in the For-
that perform poorly. This "fuzzy" or incomplete inputs, tune set into two groups accord-
note compares the neural hence may be particularly useful ing to their total return. Group 1
network results with the to decision-makers dealing with firms provided investors with the
results of a more com- qualitative, rather than quantita- highest total returns in their re-
tive, data. ANN techniques have spective industries; Group 2 firms
monly used linear multiple
outperformed regression meth- provided the lowest returns in
discriminant analysis ap- ods in predicting bond ratings.4 their industries. We divided the
proach. The non-linear, When applied to stock price be- Business Week set of 40 firms into
nine-variable, artifcial havior, however, the ANN tech- two groups. Again, Group 1 con-
neural network model per- nique has achieved only moder- sisted of those firms with the
ate success to date.5 highest valuations for their indus-
formed significantly better tries. Firms in Group 2 had the
than the tested alternative.
Multiple discriminant analysis lowest valuations for their indus-
(MDA) has been used with quali- tries.
(N

tative data to help forecast stock


This note addresses the applica- price performance. One piece of For each company, the study fo-
tion of an artificial neural net- qualitative information-the cused on the president's letter to
work (ANN) to investment analy- president's letter to stockhold- stockholders from the annual re-
sis. Specifically, can an ANN ers-provides valuable informa- port for the period immediately
differentiate between stocks that tion that is often overlooked by prior to the group-selection year.
LU
perform well and those that per- investors. This study uses the data Content analysis was used to
H form poorly? If so, how do the compiled from the president's classify and tally recurring themes
ANN's results compare with the letter to determine if an ANN can identified by similar words or
results of a commonly used mul- differentiate between firms phrases. The themes included ref-
tiple discriminant analysis ap- whose stock performed well and erences to confidence, economic
z
proach? those that performed poorly.6 factors, growth, strategic plans,
The study also compares the ANN new products, anticipated losses,
-J Multivariate discriminant analyti- technique's results with those of anticipated gains, long-term opti-
z
cal techniques such as bankruptcy more familiar multiple discrimi- mism and short-term optimism.
prediction, credit scoring of loan nant analysis. We used the frequency of allu-
z applicants and bond rating analy- sions to each theme and the per-
sis are commonly used in finan- The Study centage of the letters devoted to it
cial management.' These tech- In the original study, we com- to construct the data set. The fre-
78 niques have improved managers' piled total returns (dividends and quency data set was then used by

This content downloaded from 193.19.172.190 on Sat, 18 May 2019 11:50:32 UTC
All use subject to https://about.jstor.org/terms
Figure A An Artificial Neural Network Model to Predict Stock Price Performance Glossary

*' Content Analysis:


Systematically classifying data
Confidence
by counting and coding simi-
lar and recurring words or
Economic Factors phrases in order to analyze a
message's content. The cur-
Growth rent study analyzes the presi-
dent's letter to stockholders.
Strategic Plans Well Performing Firms

New Products "model building" in a statistical


environment.
Anticipated Losses Poorly Performing Firms
The network we used (see Figure
A) has nine input units in the
Anticipated Gains Hidden Units input layer and two output units
in the output layer. Each input
Long-Term Optimism
layer represents one of the nine
independent variables and each
Short-Term Optimism output unit represents one of two
possible outcomes-good stock
Input
price performance or poor stock
Units
price performance. The Business
Week group of companies formed
the training set for the network,
while the Fortune group formed
the testing set.

Results
both MDA and ANN techniques to an output layer. An input unit has Table I summarizes the perfor-
predict stock price performance. connections to a hidden unit in mances of the MDA and the two-,
the hidden layer, and the hidden three- and four-layered networks
We used the Statistical Analysis unit has connections to an output on the training and testing data
System (SAS) step-wise discrimi- unit in the output layer. An input sets. The Business Week nine-
nant procedure to develop a unit in this network structure.has variable MDA (training) model
nine-variable linear discriminant indirect connections to an output correctly classified 21 of 29 com-
analysis function. Discriminant unit through hidden units. Hid- panies (72%) into Group 1 and 22
analysis techniques were used to den units augment the input data of 29 companies (76%) into
classify a set of independent vari- in order to support any non- Group 2. However, for the For- a.
ables into two or more mutually linear function from input to out- tune (testing) data set, the MDA C.'

exclusive categories. The ap- put units. model yielded an overall 65% 0
proach involves finding a linear success rate; it correctly classified U
combination of independent vari- In an ANN model, a set of connec- 14 of 20 companies (70%) and 12 0
&
ables that reflects large differ- tion weights represents a function of 20 companies (60%) into
ences in group means. from input to output. The pro- Groups 1 and 2, respectively. LU

cess of computing appropriate


LUI
We developed a multilayered weights is called "learning" or The nonlinear nine-variable ANN
ANN model to predict stock price "training"; it is equivalent to model did significantly better. z
performance. We looked at three D

different network architectures-


two-, three- and four-layered net- Table I Nine-Variable MDA and ANN Models -J

works. A two-layered network has


only an input layer and an output Training Data Testing Data

layer; an input unit on the input


Model Group 1 Group 2 Mean Group 1 Group 2 Mean
layer has direct connections to an
MDA 72% 76% 74% 70% 60% 65%
output unit on the output layer. A
2-Layer ANN 63 65 64 65 40 52
three-layered network consists of 3-Layer ANN 86 79 83 90 50 70
an input layer, an intermediate 4-Layer ANN 86 96 91 90 65 77
layer (called a hidden layer) and

This content downloaded from 193.19.172.190 on Sat, 18 May 2019 11:50:32 UTC
All use subject to https://about.jstor.org/terms
The performance of the ANN should seriously consider the ap-
model also improved as the num- plication of artificial neural net-
ber of hidden layers was in- work techniques to investment
creased. As Table I shows, a two- analysis.
layered network with no hidden
units yielded only a 52% success Footnotes
rate, while a four-layered network 1. See, for example E Altman, "Finan-
improved performance to 77%. cial Ratios, Discriminant Analysis
and the Prediction of Corporate
This study did not look at net-
Bankruptcy," Joumal of Finance, Sep-
works with more layers, since
tember 1968; R Collins, "An Empiri-
four layers have been shown to
cal Comparison of Bankruptcy Pre-
be sufficient to solve a problem if diction Models," Financial
a solution exists. Management, Summer 1980; 0. Joy,
andj Tollenfson, "On the Financial
During the training phase, three Application of Discriminant Analy-
of the four models provided bet- sis," Journal of Financial and Quanti-
ter predictive capability for firms tative Analysis, December 1975.
in the lower performance cate- 2. R. Eisenbeis, "Pitfalls in the Applica-
gory than for firms in the higher tion of Discriminant Analysis in
performance category. During Business Finance and Economics,"
the testing phase, however, all Journal of Finance, June 1977; G.
Pinches, "Factors Influencing Classi-
models demonstrated better pre-
fication Results from Multiple Dis-
dictive capability for firms in the
criminant Analysis," Journal of Busi-
higher performance category. ness Research December 1980; E.
Scott, "On the Financial Application
The performance of the two-
of Discriminant Analysis: Comment,"
layered ANN was not as good as
Journal of Financial and Quantitative
the performance of the MDA Analysis, March 1978.
model. A nonlinear output func- 3. T Kohonen, "An Introduction to
tion was not useful in this ANN Neural Computing," Neural Net-
application. The three-layered works 1 (1988), pp. 3-16; P. K Simp-
ANN model, however, outper- son, Artificial Neural Systems: Foun-
formed the MDA model by 9% for dations, Paradigms, Applications and
the training data set and by 5% for Implementations (New York: Perga-
mon Press, 1990).
the testing data set. More impor-
4. S. Dutta and S. Shekhar, "Bond Rat-
tantly, the mean success rate dur-
ing: a non-conservative application
ing the testing phase for the four-
of neural networks," Proceedings of
layered network was 77%, the IEEE International Conference
compared with 65% for the MDA on Neural Networks, pp. II443-
technique. II450; A. J Surkan, andJ C. Single-
rsJ ton, "Neural Networks for bond rat-
Conclusion ing improved by multiple hidden
LU
The ANN approach used in this layers," Proceedings of the IEEE In-
study was able to discriminate ternational Conference on Neural
0
between stocks that did well and Networks, pp. II157-Ill62.
5. D. Hawley, J Johnson and D. Raina,
UJ
those that did poorly in the mar-
"Artificial Neural Systems: A New
ketplace. It also outperformed the
LU
Tool for Financial Decision-Mak-
MDA approach overall.
LU
ing," Financial Analysts JournaL No-
LU vember/December 1990.
Investors who use ANN technol-
6 The original data set formed the ba-
ogy should be able to enhance
z sisfor G. Swales, Jr., "Another Look
their analysis of investment alter-
at the President's Letter to Stockhold-
natives. However, further re- ers," Financial Analysts Journat
-J search into ANN techniques March/April 1988.
needs to be done. Testing quanti- 7. "The Fortune 500," Fortune, April
z tative and qualitative data, using 30, 1984, April 29, 1985 and April
-J various time periods and looking 28, 1986; "The Top 1000: America's
at different industries may further Most Valuable Companies," Business
U
z
validate the ANN technique and Week April 18, 1986.
4

its use in investment analysis. At


this early stage of development,
80 investors and researchers alike

This content downloaded from 193.19.172.190 on Sat, 18 May 2019 11:50:32 UTC
All use subject to https://about.jstor.org/terms

You might also like