Neurocomputing 2 (1990) 17 - 27 17

Elsevier

Stock price prediction using neural networks:

A project report
E. Sch6neburg

Expert Informatik GmbH, Roenneberg Str. 5_4, D-IO00 Berlin (West) 41, FRG

Abstract. We analyzed the possibility of predicting stock prices on a short-term, day-to-day basis with the help of aeural net-
works by studying three important German stocks chosen at random (BASF, COMMERZBANK, MERCEDES). We examined
the use of PERCEPTRON, ADALINE, MADALINE and BACK-PROPAGATION networks.
The results were encouraging. Within short prediction time spans (10 days), we achieved a very hight degree of accuracy of
up to 90%. With a BACK-PROPAGATION network we carried out an absolute-value prediction. The network was thereby
able to recognize on its own an obvious heuristic and showed a behaviour similar to the exponential smoothing algorithm.
The results we achieved led us to expect that neural networks could considerably improve the prognosis of stock prices (and
more generally, the prognosis of semi-chaotic time series) in the future.
Nevertheless considerable improvements are needed in the theory of neural networks, as practicable methods to support the
design of neural networks for specific applications are not available yet.

1. I n t r o d u c t i o n B a n k s , f i n a n c i a l institutions, large-scale in-

I n t h e c o u r s e o f the p a s t t w o years, t h e i n t e r n a -
t i o n a l f i n a n c i a l w o r l d has s u f f e r e d t w o serious
s t o c k e x c h a n g e collapses. S u r p r i s i n g l y h o w e v e r ,
these crashes d i d n o t have the d e v a s t a t i n g effects
o n the i n t e r n a t i o n a l e c o n o m y w h i c h c o u l d have
b e e n e x p e c t e d f r o m a h i s t o r i c a l p o i n t o f view. A
c e r t a i n s e p a r a t i o n o f the s t o c k e x c h a n g e f r o m the
e c o n o m i c s i t u a t i o n seems to have e m e r g e d , the
s t o c k e x c h a n g e no l o n g e r being t h e direct
b a r o m e t e r o f e c o n o m i c d e v e l o p m e n t in s o m e
areas.
This p a r t i a l d i s a p p e a r a n c e o f the n a r r o w cause
a n d t i m e r e l a t i o n s h i p b e t w e e n the s t o c k e x c h a n g e
a n d the a c t u a l e c o n o m i c f l u c t u a t i o n s m a k e s the
p r e d i c t i o n o f stock prices even m o r e difficult t h a n
it h a d b e e n until n o w . In a d d i t i o n to this, s t o c k
t r a d i n g has b e c o m e a b l e a n d f o r c e d to react to
p o l i t i c a l a n d o t h e r r e l e v a n t events faster a n d faster
d u e to the ever increasing use o f d a t a p r o c e s s i n g
t e c h n o l o g y a n d to the i m p r o v e m e n t o f c o m -
m u n i c a t i o n systems.
vestors a n d s t o c k b r o k e r s t h e r e f o r e n o w m o r e t h a n
ever face the p r o b l e m o f having to b u y a n d resell
at a p r o f i t a m a x i m u m n u m b e r o f stocks within the
s h o r t e s t p o s s i b l e time. In this r e g a r d , a t i m e s p a n
o f o n l y a few h o u r s b e t w e e n b u y i n g a n d selling
stocks is n o t u n u s u a l .
U n d e r these c o n d i t i o n s , w h a t are the possibili-
ties for p r e d i c t i n g s t o c k prices with t h e help o f d a t a
processing and computer programs?
C o n v e n t i o n a l (i.e., n o n a d a p t i v e ) p r e d i c t i o n p r o -
g r a m s are w o r k i n g best in s i t u a t i o n s where the real
course o f stock prices c a n be s m o o t h e n e d for the
p r e d i c t i o n b y m e a n s o f statistical p r o c e d u r e s
w i t h o u t loss o f p r e d i c t i o n relevance. In the case o f
m e d i u m - t e r m i n v e s t m e n t , this is realistic a n d
useful, f o r instance, for p r i v a t e investors w h o
u s u a l l y keep their stocks for a n u m b e r o f weeks o r
m o n t h s . L a r g e - s c a l e investors a n d b r o k e r s can use
such p r o c e d u r e s o n a limited scale o n l y . T h e y
m a k e t h e i r p r o f i t s , for t h e m o s t p a r t , n o t f r o m a
m a x i m u m price increase o f fewer s t o c k s over a
l o n g e r p e r i o d o f t i m e , as is the case for p r i v a t e in-

0925-2312/90/$03.50 © 1990 - - Elsevier Science Publishers B.'v.

18 E. Schi~neburg / Stock price prediction
vestors, but rather as a rule from a relatively smal 1
price increase of a great number of stocks in an ex-
tremely short time.
Under these conditions, conventional prediction
methods can only be used with a great amount of
effort. The price fluctuations are for the most part
too minimal to be successfully recognized and
predicted with the statistical procedures which
smoothened the course of stock prices for prediction
purposes. Statistical smoothing usually disregards
just that factor which is most important for predic-
tions in short periods of time: small upward or
downward price fluctuations.
A further and very important disadvantage of
conventional nonadaptive program techniques is
that for the analysis of the data, all relevant in-
fluencing factors must already be known in ad-
vance. Conventional programs must " k n o w " all
parameters and influencing factors, i.e., they must
contain the parameters along with their
assessments and weightedness a priori in their pro-
gram code. The designer of this type of program
must therefore know and determine exactly which
influencing factors he thinks are important and to
what extent the wants to consider them.
In the case of the economic development and in-
tertwinement of today's world, such as we are
witnessing with the upheavals in Eastern Europe, it
is practically impossible to recognize all the rele-
vant influencing factors and explicitly formulate
them. The fundamental analysis is basically
restricted by the incredible abundance and com-
plexity of information to be considered. No one
can determine today what relevant factors will af-
fect the world economy in the next few years. Too
many factors are unsure.
As a result, it is highly improbable that a con-
ventional program should be able to correctly
recognize and predict, even with only approximate
accuracy, the trends implicitly contained in the
economic data of the future. This would only be
possible if the program were constantly adapted to
the current situation, which is as a rule very costly
and difficult, and which would probably take more
time than the life-span of the trends themselves.
(We do, not criticize adaptive algorithms for
statistical prediction since most of them are using
ADALINE-like neural network algorithms in-
troduced by Widrow and Hoff; see [4], [1, p. 199
ff.] and [2].)
1.1. The alternative: Neural networks
In the current situation, neural networks offer a
genuine alternative to conventional nonadaptive
prediction techniques. They have two great advan-
tages over conventional methods: (1) they can
recognize " o n their own" implicit dependencies
and relationships in data, and (2) they can "learn"
to adapt their behavior (their prediction) quickly
and without complication to changed conditions.
These important capabilities of neural networks
solve, at least in principle, some of the problems
mentioned above.
The following results show that it is possible to
make surprisingly good predictions, even with
relatively simple and almost "old-fashioned"
types of neural networks. These results are part of
a comprehensive study, which we carried out in
preparation for a contract to develop neural net-
works for stock price prediction for a major Ger-
man bank (see [31).
2. The results
For the purpose of the following study, we have
assumed that stock prices do not represent com-
pletely chaotic time series. This is assumed from
time to time. We have been working on the basis
of the axiom that stock prices show, at least in
part, certain tendencies and trends which may be
very difficult to recognize but nevertheless exist.
We have also presupposed that price
predictions--even in a very restricted form--are
possible and useful using only methods of
technical stock analysis. We have until now not
considered fundamental analytical methods.
We have analyzed several classical network types
in terms of their suitability for stock price predic-
 E. SchOneburg / Stock price prediction
tion. In particular, we have examined A D A L I N E ,
MADALINE, PERCEPTRON and BACK-
P R O P A G A T I O N networks. Our aim was mainly a
short-term rise-fall prediction for the next day ( in
the case of the A D A L I N E , M A D A L I N E and
P E R C E P T R O N networks) and an absolute-value
prediction ( B A C K - P R O P A G A T I O N network) for
the next day.
As data for training the networks, we used the
prices of three randomly chosen major German
shares (BASF, C O M M E R Z B A N K and MERCE-
DES) over a period of 42 days (from 9 February
1989 until 14 April 1989). After the 42 training
days, predictions were made for up to a maximum
of 56 days.
The input vector had a breadth of 40 elements
for most o f the networks (the BACK-
P R O P A G A T I O N network had 10 input elements),
i.e., for each learning step 40 (10) data items were
fed in parallel into the networks' input layer. These
are the input information items we used:
• K = the current day stock price,
• VV = the absolute variation o f the price in
relation to the previous day,
• R V = the direction o f this variation (rise,
fall),
• RG = the direction of the variation from two
days previously,
Different Spllttings of the I n p u t v e c t o r s
a) [ 21
12 II
c) 16
d) 16
Fig. 1. Different splitting of the input vectors.
19
• G = major ( > 107oof stock price) variations
in relation to the previous day,
• the prices of the last 10 days in the case of
BACK-PROPAGATION.
In our studies, the element K has for the most
part not been directly necessary for prediction, as
it is only the relative variation in prices, and not
the absolute price value, which is of significance
for price predicition.
Figure 1 shows some of the linearizations of in-
put data we examined and their distribution on the
input vector. The predictions were strongly depen-
dent on the linearization chosen, as shown by the
result charts below.
2.1. T h e A D A L I N E network
Training was over a period of 42 days, predic-
ting over 19. A comparison of the dependencies
between various linearizations and the learning
cycles (between 2000 and 3000 presentations of
data to be learned were " s h o w n " to the network)
is presented in Table 1.
Our best result, with the BASF stock (lineariza-
tion (c), 2500 learning steps), was an accuracy rate
of 79°7o for a rise-fall prediction for the following
day. The C O M M E R Z B A N K stock, with lineariza-
tion (d) and also 2500 learning steps, was predicted
W RV RG G
E
-
[ 16
I 10
, II 9
20 E. SchOneburg / Stock price prediction

with a maximum accuracy of 74o70. In contrast, the
results for the MERCEDES stock were rather
modest. The maximum accuracy achieved here
with linearizations (a) and (d) and 2000 learning
steps, and with linearization (a) and 3000 learning
steps, was only 58°70.
We found the dependency of the networks'
prediction capacity on the planned prediction
period and on the distance in time from the actual
date to be of particular interest. The results are
shown in Table 2, Fig. 2, Table 3, Figs. 3 and 4.
Figures 2 - 4 are to be interpreted in the follow-

Table 1
Prediction accuracy (070) dependent on linearization and number of learning cycles with the ADALINE network

BASF COMMERZBANK MERCEDES

2000 2500 3000 2000 2500 3000 2000 2500 3000

Linearization (a) 58 58 58 47 68 68 58 53 58
Linearization (b) 47 68 63 63 58 58 42 53 53
Linearization (c) 63 79 74 68 63 63 53 53 53
Linearization (d) 68 68 68 63 74 68 58 53 53
Linearization (c) 47 63 63 58 63 63 53 53 53

Table 2
Prediction accuracy (070) dependent on linearization and prediction time for BASF stock with the ADALINE network

Prediction period in days

10 20 30 40 50 58

(a) 70,0 55.0 46,7 40,0 42,0 41,4 49,2

(b) 70,0 65,0 56,7 50,0 50,0 48,3 56,7
(c) 80,0 70,0 60,0 55,0 60,0 56,9 63,7
(d) 80,0 75,0 70,0 57,5 58,0 55,2 66,0
(e) 70,0 60,0 50,0 40,0 38,0 37,9 49,3
74,0 65,0 56,7 48,5 49,6 47,9 57,0

Table 3
Prediction accuracy (°7o) dependent on linearization and prediction time for COMMERZBANK stock with the ADALINE network

Prediction period in days

I0 20 30 40 50 58

(a) 60,0 65,0 63,3 52,5 54,0 51,7 57,8

(b) 80,0 65,0 63,3 60,0 62,0 58,6 64,8
(c) 70,0 65,0 56,7 47,5 54,0 53,4 57,8
(d) 90,0 75,0 60,0 52,5 58,0 56,9 65,4
(e) 70,0 65,0 60,0 52,5 56,0 53,4 59,5
74,0 67,0 60,7 53,0 56,8 54,8 61, I
 E. Sch6neburg / Stock price prediction 21

ing way: the thin line represents the training placed in front of the bold line. It is included to
period, the bold line the prediction period and the allow a comparison between the training data and
dotted line the average prediction succes in °70 of the recall data. The straight line represents the
the several 10 day periods. The thin line should be overall trend in prediction accuracy.)

BASF- Stock-Prediction
DM %
320 100
315 - - 90
310 80
3on : ' : - ~ ..... 70

300 60
295 50
290 ........ 40

280 -- 20

275 10
270 l : l l l l l l , ' , : ' , l l ~ ' , : ] [ l l l i l , : : l l l l l l l ' , : l l l l l l ] l ' , ' , : : l l l ' , ' , ' , l l l l l : 0
19.04.89 05.08.89 22.06.89 07.06.89 21.06.89 06.07.89

Fig. 2. BASF stock prediction with the ADALINE network: Training (thin line), prediction (bold line), average prediction (dotted line),
and overall trend (straight line).

Commerzbank Stock-Prediction
DM %
280 100
275 ~ / 90
2 7 0 - . . . . . . . . . . 80

260 60
255 ..~ . . . 50
250 40
245 30
240 20
235 10
230 I I ' , = : : l l ' , ' , ' , : l : : = , ' , ' , l ' , = l i l l = = ' , : l t t l l l t l l l t l i = = , I p l l : i = = l : r r ' l ' r 0
19.04.89 06.06.89 22.06.89 07.06.89 21.08.89 06.07.89

Fig. 3. COMMERZBANK stock prediction with the ADALINE network: Training (thin line), prediction (bold line), average prediction
(dotted line), and overall trend (straight line).
22 E. Sch6neburg / Stock price prediction

Mercedes- Stock-Prediction
DM %
590 100

580 90

570 //~ 80

,o o,,g,,,,,,
.......... ,,,
,o
550------_____ _ _ ~ ~ A \ A . eo

530 \ \A . J .. " ~ / . ~ "~ .....

v,,\ t-- ,o

510 ~ ~ 20
5OO 10
4g01 t t I I I I I I I I I I I I I I I I I III '11, : I : : : : ', ', : ', 'l I I I I : : : ', ', ', I '1 '1 ~ ~1, I I I ', ', ', ',0
19.04.89 05.06.89 22.06.89 07.06.89 21.06.89 05.07.89

Fig. 4. MERCEDES stock prediction with the A D A L I N E network: Training (thin line), prediction (bold line), average prediction (dot-
ted line, and overall trend (straight line).

It must be noted that, although the success rate M A D A L I N E networks with 3 to 21 A D A L I N E

drops along an almost linear pattern as the time elements (due to combinatorial considerations, we
distance to the actual date increases (which was to only used odd numbers o f A D A L I N E PE's). At
be expected), it was nevertheless possible to make first, we worked with a fixed linearization (a) and
very good predictions for some ten-day periods put the network through 4000 learning steps.
even though the time distance to the actual date The results of these tests (based on a rise-fall
was very large. This is probably because the net- prediction for 19 days) can be seen in Table 4. For
work can make good predictions for periods which all three stocks, 17 A D A L I N E processors in the in-
have occurred in a similar form in the past and for termediate layer were a local optimum (as we
which the future then was similar to the future to guess). The relationship between the number of
be predicted. In these cases the network was able A D A L I N E processors needed, the basic lineariza-
to recognize similar historical situations relatively tion and the prediction time are presented in a
well. chart (Table 5), using the C O M M E R Z B A N K stock
as an example. We have, however, not yet analyz-
2.2. The M A D A L I N E networks ed these interrelationships mathematically, as they
are quite complex.
A M A D A L I N E network can be seen as a com-
bination of A D A L I N E networks, in which there 2.3. The P E R C E P T R O N network
are A D A L I N E neurons in the intermediate layer of
the network which generate a result by using the The P E R C E P T R O N network showed the worst
majority function in the output layer. For a prediction capacity o f all, a fact which we had ex-
M A D A L I N E network, therefore, it is necessary to pected in view o f the well-known limitations o f this
determine how many A D A L I N E processors are to type o f network. The best result was 68°70 ac-
be used in the intermediate layer. We tested curacy, but it was surprisingly enough achieved
 E. SchOneburg / Stock price prediction 23

very often for the M E R C E D E S stock, which was Determining optimal learning coefficients was
not very successfully predicted by the other net- difficult. The final tests were carried out with lear-
work types. ning coefficients cl = 0.01 and c2 = 0.05, as they

Table 4
Prediction accuracy (°7o) with the M A D A L I N E network

ADALINE PE's BASF COMMERZBANK MERCEDES

3 47 68 58 58
5 58 58 63 60
7 53 63 58 58
9 58 63 47 56
11 58 74 63 65
13 58 68 58 62
15 68 58 53 60
17 68 74 63 69
19 58 74 47 60
21 53 63 47 54

58 66 56 60

Table 5
Relationship between n u m b e r of A D A L I N E processors, basic linearization and prediction time for C O M M E R Z B A N K stock with the
M A D A L I N E network

COMMERZ- Prediction period in days

BANK
10 20 30 40 50 58

(a) 11 80 70 63 60 58 57 65
(a) 17 70 75 63 63 56 57 64
(a) 19 80 70 60 63 56 57 64
(d) 7 70 65 60 55 56 57 61
(d) 11 80 65 53 50 56 57 60

76 69 60 58 56 57 63

Table 6
Prediction accuracy (o7o) with the P E R C E P T R O N network

cl = 0.01, c2 = 0.05 BASF COMMERZBANK MERCEDES

2500 3000 3500 2500 3000 3500 2500 3000 3500

Linearization (a) 58 58 58 47 47 47 63 63 63
Linearization (b) 42 42 42 53 58 63 58 42 47
Linearization (c) 58 58 58 47 47 47 68 63 63
Linearization (d) 47 47 47 58 58 58 68 68 68
Linearization (e) 68 52 52 42 42 42 63 68 68
24 E. Sch6neburg / Stock price prediction
achieved the best results. Linearizations (a)-(e)
were used, and stock prices for 43 days learned
2500 to 3500 times. Afterwards, 19 predictions
were assessed; the results are presented in Table 6.
2.4. The B A C K - P R O P A G A T I O N network
To train the BACK-PROPAGATION network,
the input data had to be transformed (scaled) to
the real interval [0.1, 0.9] first. Finding the best
topology for the BACK-PROPAGATION net-
work required much work. The best results pro-
duced the topology with the corresponding
transfer functions as shown in Fig. 5.
As learning rules, we used the cum-delta rule as
well as the delta rule; however, only the connec-
tions to the output layer learned with the cum-delta
rule. This led to slightly better results than with the
exclusive use of one of the two rules. The summa-
tion function was the simple summation and the
output function was direct. We used the following
figures for the learning coefficients: cl (learning
rate) = 0.6 and c2 (momentum) = 0.9.
A partial result of our test series for 60 days of
training and 40 recall figures (predictions) can be
seen in Figs. 6 and 7. It can be seen that the
predicted figures exhibit a time shift of almost one
day.
output
hidden la~,er~3 hidd~'eenlayer 4
I I
hidden layer 1 hidden layer 2
input
J (10 neurons)
sigmoid
Z
sigmoid sine
I I
sigmoid sine
~linear j
Fig. 5. Topologyand transfer functionsof the BACK-PROPAGATIONnetwork.
There are two reasons for this time shift. Firstly,
the network was trained wiht a "window" of 10
days being placed on the learning data to predict
the eleventh day. The "window" was then shifted
one day to the right and the new set of data was
then learned.
Secondly, and interestingly, the network itself
has apparently discovered the following heuristic:
> > take for the next-day prediction the stock
price of the last day and modify it only slightly
< < . This is evident in the fact that the prediction
varies only slightly from the previous day's stock
price, and creates the impression of a time shift of
the prediction in relation to the actual stock price.
This apparent time shift does also occur when
using exponential smoothing of first or second
order for prediction purposes instead of neural
networks. The network therefore discovered the
importance of the last day's stock price for the
prediction which is the essence of the exponential
smooting algorithm. It used the 10 day data win-
dow only to a minor extent as a support for the
prediction.
The simple heuristic detected by the network
often leads to quite good results, e.g. in weather
forecasting, since as a rule a weather change is less
probable than a steady weather pattern. This is
also true for stock prices. The critical point with
- bias (1 neuron)
(5 + 5 neurons)
(5 + 5 neurons)
 E. Sch6neburg / Stock price prediction 25

Commerzbank Prognosis
using B a c k p r o p a g a t i o n

280

20

260 ~

250 ~ ~ ~

230

220 i rrllllrlllllr iiiiiill]l iii;rrr FII r r r l l l r ~ p l l l l r r i i ~ l r r r l l l ~

18,4.89 02.6.89 11,6.89 19.5.89 31.5.89 09.6.89 20.6,89 29.6.89 13.7.89

Real D a t a 4-- Predicted Data

Fig. 6. COMMERZBANK stock prediction with the BACK-PROPAGATION network.

Mercedes Prognosis
using B a c k p r o p a g a t i o n

580

5,o

560 ~ -

550 +.--~ ÷ ~ - -

510 + ~

5 0 0 T I I [ [ I I I ] I I r ~ I I I P ] I = ~ I r I I I I I r I I r i i i r i

18.5.89 26.6.89 02.6.89 09.6.89 16.6.89 23.6.89 30.6.89 07.7.89 13.7.89

Real D a t a ÷ Predicted Data

Fig. 7. MERCEDES stock prediction with the BACK-PROPAGATION network.

this simple heuristic is, however, the > > slight 3. Further tests
variation < < from the previous day's price as a
recommendation for the following day. The quali- The results presented above were encouraging
ty with which this > > slight variation < < is and have led us to further tests. We are currently
recognized and characterized is the quality with concentrating on the following areas:
which the prediction can be improved compared (1) The selection of other, perhaps more suitable
with the obvious heuristic. input information for the networks (e.g. variation
26 E. SchOneburg / Stock price prediction
quotients from previous day's prices as input or in-
formation about stock price behavior on the basis
o f reference stocks from the same field of
business--key word: field dependencies, etc.). We
are also currently examining the possibility of
designing neural networks as constraint-satisfac-
tion networks for fundamental analytical pur-
poses.
We have estimated that the amount of work
necessary to design such a network would be about
the same as the amount of work required for the
acquisition of knowledge about stock behavior to
be used in an expert system for this task. We
believe that this would be very time-consuming and
expensive.
(2) The study of further network types in terms
of their suitability for price predictions (among
others, counter propagation, Hopfield, Kohonen,
ARTx). The first analyses with Hopfield networks
required considerable effort for a suitable trans-
formation and preparation o f the input and the
definition of an energy function. Other tests using
feature extraction with Kohonen-like networks
have not yet produced results worthy to mention.
Nevertheless we presume that competitive self-
organizing networks are most suitable for predic-
tions. We do, however, intend to systematically ex-
amine other types of networks in terms of their
suitability, as we are not sure how much complex
networks are better as compared to simpler ones.
It could be that the increase in prediction capacity
of the more complex networks is relatively small,
in which case the work involved would probably
no longer be worth in practical use.
(3) The discovery of > > best < < network
types and topologies and the determination o f rele-
vant network parameters (e.g. the number of
layers and neurons, learning coefficients, connec-
tions of layers, threshold values, learning rules and
strategies, etc.).
(4) The improvement of the theoretical means at
our disposal as aids in designing and analyzing net-
works.
(5) The development of distinct, problem-ori-
ented network simulators in sequential and parallel
> > C < < . For the simulations described above,
some 5000 lines of C code were necessary. For the
analysis of the performance of complex networks
we are currently testing neurocomputers and
transferring our simulator to transputer hardware.
(6) We have estimated the statistical relevance of
our results. Since we had an accuracy of 63 °70 using
285 predictions in total the probability of this out-
come is as low as 1/100 000 which is not too bad.
On the 10 to 20 days basis the results were even bet-
ter as we had an average of 71,3°70 correct predic-
tions on 300 predictions in total.
The statistical relevance of these results is yet to
be confirmed. The test series were too short to per-
mit statistical assessment. In 1990 we will carry out
comprehensive statistical analyses over a con-
tinuous prediction period of 10 months.
(7) In order to become less dependent on the net-
work types and the various parameters, we are cur-
rently exploring the possibility of determining the
weighting between the neurons by means of a
genetic algorithm and an evolution theoretical ap-
proach. We will be reporting on our progress in
suitable time.
4. Problems
We cannot deny that we have been faced with
complex problems in some of our attempts to
determine the suitability o f neural networks for
prediction of stock prices; for some of these, there
is no solution in sight, not even in theory.
Most of the time spent in the project was related
to the difficulties of finding the > > best < < net-
work topology and setting the corresponding net-
work parameters (see point (3) in Section 3). As a
rule even slight parameter changes caused major
variations in the behavior o f almost all networks.
There is no theory available yet which could be us-
ed as a guideline to find > > best < < networks.
There is a real practical need for a theory of
equivalence classes of neural networks and the
development of a partial order on these classes
defining the relation of complexity of networks
 E. Sch6neburg / Stock price prediction 27

and minimal elements (i.e., networks with the Professor Eberhard Sch6neburg,
s m a l l e s t n u m b e r o f n e u r o n s n e c e s s a r y t o s o l v e cer- born in 1956, studied mathematics,
computer science, physics and
tain kinds of problems etc.). As classification philosophy at the Berlin Free Univer-
t h e o r y in l o g i c s h o w s , this is p o s s i b l e e v e n f o r c o m - sity and Technical University from
1979 to 1984. After the completion of
plex mathematical t h e o r i e s . It s h o u l d at l e a s t in his studies (degree: Dipl. Math.), he
principle be possible for subclasses of neural net- worked as software and systems
engineer for SIEMENS AG in Berlin
w o r k s . W i t h o u t a n y t h e o r y o f this k i n d at h a n d and Boca Raton, Fla, USA
o n e is a l m o s t a l w a y s f o r c e d t o p r o c e e d u s i n g a (1984 - 1986). At SIEMENS, his area
of involvement consisted of the
time-consuming and laborious trial-and-error development of intelligent self-test
strategy. and diagnosis modules for real time applications and the
automatic generation of programs.
Professor Sch6neburg accepted an offer to DORNIER to
take over the direction of the department for new technologies,
expert systems and computer security in 1986. In 1988 he found-
References ed Expert Informatik GmbI-I. This company specializes in ex-
pert systems applications in banks and in the production area
(solvency analyses, production planning), and was the first
[1] M. Hiittner, Prognoseverfahren und ihreAnwendung (De company in Germany to be asked to develop neuronal networks
Gruyter, Berlin, 1986). for a major bank; it also rates among the leaders in the develop-
[2] P. Mertens, ed., Prognoserechnung (Physica-Verlag, ment of anti-computer-virus sytems.
With the chair of artificial intelligence at the Furt-
Wiirzburg, 1981). wangen/Schwarzwald Technical College, Professor
[3] E. Sch6neburg, N. Hansen, M. Gantert and M. Reiner, Sch6neburg's scientific interest has concentrated on applica-
Kurzfristige Aktienkursprognose mit Neuronalen Netzen, tions of nonclassical logic in robot technology, in the prediction
Expert Informatik GmbH, Berlin, 1989. of semi-chaotic time series with neuronal networks (dissertation
topic) and in expert systems security. Professor Sch6neburg has
[4] B. Widrow and M.E. Hoff, Adaptive switching circuits, just completed writing a German textbook on neural networks
in: Anderson and Rosenfeld, eds., Neurocomputing (MIT containing a software simulator in > C < (to appear in 1990).
Press, Cambridge, MA, 1988) 126 ff.