Purdue University

Purdue e-Pubs
Computer Science Technical Reports Department of Computer Science

1997

Financial Prediction and Trading Strategies Using

Neurofuzzy Approaches
K. N. Pantazopoulos

L. H. Tsoukalas

N. G. Bourbakis

M. J. Brün

Elias N. Houstis
Purdue University, [email protected]

Report Number:
97-002

Pantazopoulos, K. N.; Tsoukalas, L. H.; Bourbakis, N. G.; Brün, M. J.; and Houstis, Elias N., "Financial Prediction and Trading
Strategies Using Neurofuzzy Approaches" (1997). Computer Science Technical Reports. Paper 1342.
http://docs.lib.purdue.edu/cstech/1342

This document has been made available through Purdue e-Pubs, a service of the Purdue University Libraries. Please contact [email protected] for
additional information.
 FINANCIAL PREDICTION AND TRADING
STRATEGIES USING NEUROFUZZY APPROACHES

K. N. Pantazopoulos
L. H. Tsoukalas
N. G. Bourbakis
M. J. Brun
E. N. Houstis

CSD-TR 97-002
January 1997
 Financial Prediction and Trading Strategies Using
Neurofuzzy Approaches

K.N. Pantazopoulos. L.H. Tsoukalas,

N.G. Bourbakis, M.J. BrUn, and E.N. Houstis

Abstract

Neuro[uzzy approaches for predicting financial time series are investigated and shown to perfarm
well in the context of various trading strategies. The horizon of prediction is typically a few days
and trading strategies are examined using historical data. A methodology is presented where neurnl
predictors are used to anticipate the general behavior of financial indices (moving up, down, or
staying cOJlstant) in the context of stocks and options trading. The methodology is tested with
actual financial data and shows considerable promise as a decision making and planning tool.

Acknowledgments: This researchhas been partially supported by INEL University Research Consortium. The
INEL is managed by Lockheed Martin Idaho Technologies Co., for the US Department of Energy under conlraCL No.
DE-AC07-94ID3223; also by NSF grant.'! 9123502-CDA, 92022536·CCR, and 620-92-J·0069, and ARPA grant
DAAH04-94-G-OOlO. Finally. the gracious support of the Purdue Special Initiative Fellowship and the Purdue Re-
search Foundation is gratefully acknowledged. The authors are also grateful Lo A. Ho bbs for sllbslaDtial conuibll-
lions Lo Lhe stock pricing model.
1. Introduction

The prediction of financial market indicators is a topic of considerable practical interest and, if
successful, may involve substantial pecuniary rewards. Neural networks have been used for sev-
eral years in the selection of investments because of their ability to identify patterns of behavior
that are not readily observable. Much of this work has been proprietary for the obvious reason
that the users want to take advantage of the insight into the market they gained through the use of
neural network technology. Thus, conventional wisdom has it that if one succeeds in developing

a methodology that correctly predicts the ups and downs of various market indicators she or he
is unlikely to seek publication of results; conversely, if something is published it must really not
"make money." We take both views as indicative of the fact that the ultimate test of value for

financial prediction lies with the market. Be that as it may. financial markets, and economic s ys-
terns in general, present fascinating opportunities for studying complexity and testing the limits
of methodological advances and computational technologies (Anderson et al, 1988). Unlike the

natural systems of physics and chemistry. financial systems involve the actual interplay of deci-
sions and actions taken by millions of individual investors and institutions on a global scale.

Although participants in the market are constantly involved in making decisions based on their
predictions or anticipations of market phenomena, there is considerable disagreement amongst
experts as to what degree, if any. financial time series are predictable and how to predict them.
Some researchers have identified the presence of chaos in financial indicators. implying that fi-
nancial systems are characterized by non-repetitive and non-predictable fluctuations arising

through the interplay of a system's participating agents and its relation to other systems
(Radzicki. 1990). Others have argued for certain underlying regularities and patterns and hence
for a more predictable market system (Lippit, 1990), (peters, 1994). The "efficient market hy-
pothesis," currently the most widely held view of market behavior, states that no investment
"system" or technical strategy can yield average returns exceeding the average returns of the mar-

ket as a whole. Any information that would yield extra returns spreads rapidly and is "priced
out" within seconds of its introduction.

1
Although nearly everybody agrees on the complex and nonlinear nature of economic systems,
there is skepticism as to whether new approaches to nonlinear modeling, such as neural net-
works, can improve economic and financial forecasts. Some researchers claim that neural net-
works offer no major improvement over conventional linear forecasting approaches (Farber,
1988), (LeBaron, 1994). One possible explanation may be that, while empirical studies in the
natural sciences are characterized by large data sets, often numbering in the tens of thousands,
data sets in economic applications usually consist of less than one thousand observations. Con-
sequently, computational procedures designed in the former context may not be appropriate in
the latter (Ramsey, 1990). In addition, there is a great variety of neural computing paradigms,
involving various architectures, learning rates, etc., and hence, precise and informative compari-
sons may be difficult to make. Lapedes and Farber, amongst others, have offered interesting and
convincing evidence that in the context of time series prediction, neural nets can very accurately
model the non-linearities involved (Lapedes, 1988). In recent years, an increasing number of re-
search in the emerging and promising field of financial engineering is incorporating neurofuzzy
approaches (Refenes, 1996), (Trippi, 1996).

We use a neurofuzzy methodology to guide simulated trades in options based on the level of the
Standard & Poor's 500 (S&P 500) index, and find that if we use a standard option pricing model
(we use the Black & Scholes model found in Black, 1973), our trades would yield returns more
than 150 times greater than a portfolio fully invested directly in the S&P 500 index. It would not
be in line with the efficient market hypothesis, however, to expect such results from actual in-
vestment. The promise of high returns would raise the price of the options until normal return
rates were again reached.

Hobbs and Bourbakis have described a neurofuzzy simulator used for stock investing (Hobbs,
1995) that identifies patterns associated with whether a stock is under-priced or over-priced.
This model is expanded in Section 4 and neural predictions are used in the context of two related
but different trading strategies, the first based on stocks, and the second based on options. A

2
stock is a certificate of ownership representing one or more shares of a corporation's equity.
Usually stocks are listed, i.e., they are traded in the financial markets and world's stock ex-
changes. An option is a financial instrument, a fonn of contract between two parties, which is
also traded in the financial markets and organized exchanges, much like common stocks, com-
modities, currencies, etc. The basic difference between options and entities such as stocks, is
that options are derivative products, i.e., they are based on a primary asset such as a corpora-
tion's stock. Options can also be based on aggregatemetrics such as market indices. Market in-
dices are aggregate measures (weighted averages) of stocks. The "price" of a market index
fluctuates over time, representing the changes in the prices of the individual stocks. The price of
the various market indices is continuously calculated and quoted by various services.

In the case of the stock trading strategy, a neural predictor is designed based on the assumption

that «the price ofa stock should generally change proportionately with (or against) certain mar-
ket indices, e.g., the Dow Jones Industrial Average (DITA), or the S&P 500." The neural net-
work, after analyzing these market variables, predicts the true price for a particular stock x, and
decides whether it is under- or over-priced. Although the overall market might swing up one day,
and down the next, the neural network model should be right when the market stays constant; the
up and down swings should offset each other on the other days. Under this assumption, a
neurofuzzy model is developed, reflecting the fact that some fluctuations in market indices might
be more important with respect to the price of stock x than others. It involves a neural network,
where each neuron's running average error from its previous predictions can be used to calculate a
reliability index for the neuron. At the last layer, the network makes a fuzzy or disproportionate
weighted average of all the neuron's predictions based on a Gaussian activation function of the
output neuron's average error. The network traces several market variables (MVs) based on a
first order gradient of their running averages. The model presented here. is a fuzzy neural net-
work with two hidden layers which analyze consecutive sets of market variables.

3
In the case of the option trading, we discuss strategies involving options on the S&P 500 index,

i. e., options that are derived from the S&P 500, although our results can be extended for strategies
based on other options in a straightforward manner. The option trading strategy is based on neu-
ral predictors of the volatility of the S&P 500 which anticipate lip, down or same type of move-
ments of the index. Based on the superior knowledge provided by the volatility predictions, a
profitable trading strategy that involves option is made possible. The model presented here is a
fuzzy recurrent neural network with one hidden layer that analyzes the stream of daily prices of
the S&P 500. The S&P 500 options are among the most widely traded options and are listed in
the Chicago Board of Trade. Options in general come in many different flavors and variations. In
this paper we will be concerned only with basic options, calls and puts, and combinations of
them, such as straddles. A call is a financial contract giving the right, but not the obligation, to the
contract holder to buy a specified asset (the underlying asset) for a specified price (the strike
price) on, or before, a specified date (the expiration date). A put is similar, only it gives the right
to sell the underlying asset for the strike price on., or before, the expiration date. A straddle, is
one of the many packaged options that can be created when calls and puts are combined linearly;
in the case of a straddle, one call option and one put option, both on the same underlying asset,
with the same expiration date, and the same strike price, are packaged together.

The rest of the paper is organized in six sections. Section 2 discusses the neurofuzzy approaches
pertinent to the time series prediction problem and their practical ramifications. Section 3 pro-
vides an introduction to the prediction problem as it is encountered in financial engineering ap-
plications and a brief survey of some neural network models used in the market today. ~ection 4
discusses a neurofuzzy methodology for stock trading strategies and Section 5 presents an appli-
cation of a neurofuzzy methodology to trading strategies of options based on predictions of vola-
tility. Section 6 discusses the results of the neurofuzzy approach taken.

4
2. Neurofuzzy Approaches in Prediction

Neurofuzzy approaches represent an integration of neural networks and fuzzy logic that have
capabilities beyond either of these technologies individually (see (Tsoukalas, 1997)). In applica-

tions of financial time series prediction., establishing crisp criteria for decision making is rather
difficult to achieve; neurofuzzy approaches are shown to be capable of providing improved deci-
sion support in situations where crisp estimates are either meaningless or unavailable.

Although about 80% of all neural network applications utilize networks in which there are no
feedbacks from the output of one layer to the inputs of the same layer or earlier layers of neu-
rons, there are situations (e.g., when dynamic behavior is involved) where it is advantageous to

use feedback. When the output of a neuron is fed back into a neuron in an earlier layer, the out-
put of that neuron is a function of both the inputs from the previous layer at time t and its own

output that existed at an earlier time, i.e., at time (t-At) where At is the time for one cycle of cal-

culation. Whereas feed forward neural networks appear to have no memory since the output at
any instant is dependent entirely on the inputs and the weights at that instant, neural networks
that contain such feedback, called recu"ent neural networks, exhibit characteristics similar to
short-term memory; because the output of the network depends on both current and prior in-
puts. Although virtually all neural networks that contain feedback could be considered as recur-
rent networks, the networks considered here use backpropagation for training.

The complexity introduced by feedback connections, even for elementary systems, is readily ap-
parent during training when it may take hours and even days of computer time for a network 1O
converge. Yet, for some networks the increase in complexity is often compensated for because the
feedback drastically reduces the number of cycles needed to train a neural network. Feedback can
often be used advantageously to speed up the training of a neural network and to avoid local min-
ima. Indeed, it is sometimes possible to train a neural network after feedback has been added
whereas it may not have been previously possible to train it to the desired low level of error.
However, capturing dynamic behavior in a model is the most common justification for the use of

feedback, and recurrent neural networks are almost always used for dynamic signals such as the

5
 Output Yk

Desired Output

Recurrent
Feedbacks

rk-15

Time

Figure 1. Neural network for time-senes prediction.

financial indicators investigated in this work. Neural networks can be used to predict future
values in a time series based on current and historical values. Such predictions are, in a sense, a
form of inferential measurement. Because of their ability in prediction, there has been an
extraordinary amount of interest in the use of neural networks to predict stock market behavior
(Refenes, 1996). This popularity continues in spite of the fact that the predictions for neural
networks cannot be explained or verified. Perhaps the main reasons for the continuing popularity
in the field are that neural networks do not require a system model and that they are relatively
insensitive to unusual data patterns.

Although backpropagation neural networks are usually used for time-series prediction, it is
possible to use any neural network capable of mapping an input vector into an output vector.
Typically, the input of a single time series into a neural network is made as shown in Figure 1.
The fluctuating variable is sampled at an appropriate rate to avoid aliasing, and sequential
samples are introduced into the input layer in a manner similar to that used in a transverse filter.
At every time increment. a new sample value is introduced into the rightmost input neuron, and a
sample value in the leftmost input neuron is discarded. The main difference compared to the

6
transverse filter is that the sample preceding those in the input is introduced into the single
output neuron as the desired output. In this way, the network is trained to predict the value of
the time series one time increment ahead, based on the previously sampled values. The network
can be trained to predict more than one time increment ahead, but the accuracy of the prediction
decreases when predictions are further into the future. Since such systems are often used in real
time, or with data from historic records, the amount of training data is usually very large. Even
so, it is important to periodically check the training to assure that over-training does not occur.
Furthermore, selection of training data is important because taking every n rh sample can introduce
bias. This can be overcome by using unequal time intervals between input samples (one method
to accomplish this is by sampling at a high rate and then using different number of time intervals
between the different input samples).

Although it is possible to predict multiple outputs, it is best to predict only one value because
the network minimizes the square error with respect to all neurons in the output layer.
Minimizing square error with respect to a single output gives a more precise result. If multiple
time predictions are needed, individual networks should be used for each prediction (Tsoukalas,
1997) in the manner shown in Figure 2.

Pre-processing Post-processing

Figure 2. Different neural networks predict different features of a time series.

7
Generally, large scale deterministic components, such as trends and seasonal variations, should be
eliminated from inputs. The reason is that the network will attempt to learn the trend and use it
in the prediction. This may be appropriate if the number of input neurons is sufficient for the
input data to span a complete cycle (e.g., an annual cycle). If trends are important, they can be
removed and then added back in later. This allows the network to concentrate on the important
details necessary for an accurate prediction.

The standard method of removing a trend is to use a least squares fit ofthe data to a straight line,
although nonlinear fitting may be appropriate in some cases (e.g., cyclic fluctuations). An
alternate method of removing trends and seasonal variations is to pass the data through a high-
pass filter with a low cutoff frequency. There are alternative techniques in which a low pass
filter is used to leave only the slowly varying trend which then are subtracted from the original
signal, with the difference being the value sent to the neural network input layer.

One of the interesting variations of the above technique for prediction is to use differences
between successive sample values as inputs to the neural network. This effectively eliminates
constant trends and slowly changing trends by converting them to a constant offset. Even
seasonal trends are usually removed in this way. Using differences in predicting is generally
useful in stock price predictions, especially if the difference is scaled relativel y to the total price
of the stock, which is effectively the percent price change.

3. Prediction in Financial Engineering Applications

The vast majority of financial and economic activity involves the analysis and prediction of nu-
merous variables. Variables that are encountered in this context are either leading or trailing eco-
nomic and financial indicators of some behavior or pattern, as well as prices and indices. In the
case ofIeading variables, the primary reason for the need of prediction and forecasting is that it
can be used to take preemptive actions. More than any other context, the financial markets offer
a bold example of how a reliable prediction can be capitalized upon. Consequently, it is natural

8
for financial engineers to seek for the best possible prediction systems. Financial engineering is
the means for implementing financial innovation which is an integral part of the activities of con-
temporary financial institutions. Finance theoreticians argue for the long standing "efficient mar-
ket hypothesis (EMH)" as a basis of denouncing any technique that attempts to find useful
information for the future behavior of stock prices by using historical data. However, the as-
sumptions underlying the El\1H in many cases are not realistic. Studies indicating that some
form of the EMH is indeed observed in the financial markets, as well as others, that conclude to
the contrary, add to the confusion. Irrespective of the theoretical foundations, technical analysis
is used widely as a tool facilitating various financial activities. Among the most challenging fman-
cial activities that can make direct use of prediction systems is portfolio management and trading.
Portfolio management aims primarily at maximizing profits and minimizing risks on a bundle of
assets such as stocks and bonds, by selecting which assets to trade of hold.

Portfolio management systems using neural networks have been developed by a number of pea-
pie and reported to achieve good results. Wilson, for example, reports on a hybrid system based
on several price prediction models including technicaL adaptive, and statistical models (Wilson,
1994). The first neural network layer used by Wilson's system, selects which of these models is
working best for each stock followed. Then the selected model makes a recommendation on
whether to buy, seH or hold a particular stock. The final layer of the system decides which
stocks to buy or sell. depending on the maximum risk that the portfolio is allowed to undertake,
as it is measured by the beta correlation coefficient of the Capital Asset Pricing Model (Black,
1972). Although the system's performance is better than S&P 500. the results show a great vola-
tility in the value of the portfolio over the 250 weeks test period, including a dip below the S&P
500 about hal£VVay through the tested time frame. Nonetheless, the results are quite impressive,
and the concept of dynamically managing the portfolio's risk factor very promising.

Lucid Capital Network has reported on some of their preliminary tests, on neural network mod-
els and applications to the futures market (Konstenius, 1994). They used a single layer network
with up to 1000 nodes, sigmoidal pre- and post-processing, and a beta learning rate of 5% to

9
10%. The study amplifies the importance and effectiveness of some sound pre-processing tech-
niques; it recommends normalizing the data so as to provide the network strictly with day-to-day
changes, rather than the actual prices of equities. In addition it recommends a "holiday manage-
ment" processing: filling gaps in the database or the holiday intervals with the previous day's
data. The concern over holiday gaps is a significant source of noise for most neural network plat-
forms.

A feed forward analog neural network has been developed by Lowe for portfolio management
(Lowe, 1994). The marketing analysis is based on the fundamental assumption that the market is
not totally efficient, i.e., some stock prices are not "true." Additionally, and to some degree con-
trarily, this approach incorporates "Sharpe's theory" (Sharpe, 1964) as well, whose formulas are
based upon the assumption that the market is always in equilibrium. Lowe's model was tested in
both US and European markets. Although the test seems very short (less than a year), Lowe's
portfolio has earned an impressive 15% return - even during the volatile bear market where his
data was extracted from.

Ahmad and Fatrni (Ahmad, 1994) have designed a "quadric neural network (QNN)" to predict
financial time series. The QNN was based on the Gabor-Kolmogorov polynomial model. The
fundamental idea is that all statistical and stochastic functions must be studied within the frame-
work of self-defined logical, statistical and physical concepts. They provide formulas that s ys-
tematically lead to the development of several functions, linear and non-linear, which ultimately
predict the financial time series. The model analyzes all the available data derived from a single
stream of a stock's day to day prices. Much of these derived data is arrived at by varying the
time frame used by the network, and dynamically updating the network's weights through a
feedback function ofthe least mean square (LMS) algorithm. Although this approach of system-
atic formulas is clearly defined and the network was shown to converge for a hypothetical series,
it appears that substantial testing would be required to demonstrate in practice Kolmogorov's
theory and its fundamental idea of accurately predicting a series simply by analyzing its single
stream of historical pricing.

10
4. A Neurofuzzy Methodology for Stock Trading Strategies

Stock trading is the activity of buying, selling, or sorting (selling without owning) the stock of a

publicly traded corporation. In this section, a feed-forward neural network fuzzy model (FNFM)

used for stock trading is presented where each neuron uses a number of expressions to calculate

its prediction, measure its accumulated error (reliability factor) and update its parameters.

The first step in the FNFM is the calculation of the relative change of the market variables, so

that different market variables can be compared in terms of relative growth. This transformation

is described in Equation 1 (Kung, 1993):

dMV. =IOg( MV. J (1)

MV('-A)"
where, MVrv denotes the market variable number v at time t, and d MV the log (relative change of

MVrv )' The log function is used (any base would work) so that the dMV values will be evenly

distributed as shown Figure 3.

Relative Change:

No Change
I 10 I
-------,:-+----.,,0+----+-=-,
.. -101
-------..
dMV

Figure 3. Distribution of the dMV

A lO-fold increase (d MV = 1), for example, followed by 1/1O-fold increase (d MV = -1) would
average to a d MV = 0, or a I-fold increase (i.e., no change).

Calculation of the Market Variable's Ideal Weight (Sensitivity)

Each dMV has a certain sensitivity factor to stock x, the stock of interest. This sensitivity factor,
or weight (W""J, indicates whether theMV moves with or against x, and to what degree. Equa.
tion 2 describes this relation.

11
 · dMV
From EquatlOn2 we see that Wml :::: 0;. Thus, W...,j indicates a particular MV's sensitivity lO
dMV.
stock x as shown in Figure 4.

MV less MVrnore MVrnore MV less

.-w....
sensitive

-I I
sensitive

0
sensitive

+11
sensitive

+W....
~

Inversely Proportional Directly Proportional

Figure 4. Sensitivity of stock x with respect to a market variable MV.

Learning Rate (BETA)

Before each neuron calculates its "prediction," its weight is revised by taking a weighted average of
its previous weight and the ideal weight (i.e., the teaehervariable) which should have been used in
the last iteration. Equation 3 indicates the dependence of this weighted average on the learning rate
BETA, i.e.,

Wval :::: BETA· Wva/ + (1- BETA)· w~:fa/ (3)

where W:a; is the previous (old) weight and wjg,ea/ is the ideal weight for the last iteration. Since

differentleaming rates vary with the magnitude of a MV's previous perfonnance, several BETA
values are used independently. Thus, a single market variable can provide short tenn predictions
as well as longer tenn ones. Each weight corresponds to a particular MV and with a particular
learning rate L. Equation 4 shows the way BETA is calculated where NumL denotes the total
number of different learning rates used for a particular market variable.
1
L+-
BETA = 2 (4)
NumL
From Equation 4 we see that the smaller the learning rateL, the lesser the importance given to his~

torica1 perfonnance.

12
The Activation Function
At each iteration (each day in this model), the FNFM calculates (NumA x NumV) components
from the overall network prediction, where NumV is the number of MVs used and NllmA is the
time frame. All these components are averaged based on each neuron's reliability index. The reli-
ability index, RI, is a Gaussian activation function of the average absolute error, as shown in Fig-
ure 5, and itis described by EquationS.

- C
Rl - I
*e (-errorIC z ) (5)

RI

LargeC]

mall C,
error

Figure s. Gaussian activation functions.

Since CJ is effectively normalized by the weight average, any arbitrary value, e.g. 10, will work.
C2 , however, defines the steepness of the curve, and thus defines how much more important a
component, e.g. with 1% average error, is than another component with a 10% average error. If a
20-fold will be chosen more important (for this example) a C2=O.05 is used.

Market variables (lheones used here are market indices) are the primary input data that the FNFM
analyses to gage the "true" value of stock x. Thirteen market variables are selected as network in-
put. These variables are among the most common market indices which are affected and affect the
price of the stock which we use in the evaluation of the performance of the FNFM (the common
stock of IBM). Some of the indices used are shown in Table 1.

FNFM Simulator and Parameter Settings

In running the FNFM network simulator, first the historical price database to be used is selected;
in the example described below the IBM database is selected. It contains a large table of daily
prices for the IBM stock, as well as values of the indices over the past 6 years.

13
 Table 1. Market indices used by lhe FNFM.
DJIA Dow Jones Industrial A veral!e (30 comoanies)
DJTA Dow Jones Transoortation Average (20 com anies)
DJUA Dow Jones Utility A vernge (15 companies)
DJ6S Dow Jones 65 Stock composite
DJFI Dow Jones Futures Index (Commodities)
S&P 100 Standard & Poor's 100 Index
S&P EI Standard & Poor's Electronic Instruments
S&P SC Standard & Poor's Semiconductors
S&P BE Standard & Poor's Computer Systems
S&PSOO Standard & Poor's 500 Index

The parameters NumV (number of market variables) and NumT (size of historical sample) are set
accordingly to the number of variables and number of daily prices available. The default values of
NumT, NumA (time frame size), and NumL (number of learning rates) can also be adjusted as
necessary. The parameter NumA defines the size of the input layer of the neural network, since,
depending on the size of the time frame, an appropriate number of inputs must be used. NllmL
defines the length of the second dimension for the first layer. When the neuron updates its pa-
rameters ("learns") after each pass, it takes a weighted average of the old parameters and the new
ones. The BETA value defines how to proportionately average the two values. While it is advan-
tageous for some neurons to learn quickly, it is more advantageous for other nodes to learn in a
slower pace. The approach implemented here is to take a combination of these neurons
(introducing a third dimension to the neural network) by allowing each neuron to have its own reli-
ability index. The parameter NllmL defines how many of these combinations Lo follow for each
particular neuron. The overall network has NumV * NumA * NumL neurons with NllmT passes
through them. These neurons all feed into the last layer, a single output neuron, that proportion-
atelyaverages all of the nodes' predictions based on their proven reliability.

To evaluate the predictive capabilities of the FNFM model, two investment strategies are consid-
ered using historical data for the IBM slock. The first strategy invests $1 each day. The graph of
the daily portfolio balance demonstrates the stability of this investment strategy (low volatility).
The second strategy, starts also with $1, and every day invests this amount plus all the available
cash balance of the portfolio. This second strategy is a "dividend and capital gains reillvesunent
strategy" which more accurately reflects a realistic trading practice. The overall performance of Ihe
network is evaluated based on two objectives: the rate of return, and the volatility of its day to day
perfonnance. To evaluate both, a simple and objective approach is to consider the daily cash bal-
ance of a portfolio that starts with an initial balance of $1 and subsequently trades the IBM stock

14
 51

Investment
Profit in
dollars

$I

'88 '91
Tmding Period (1988 - 1994) '"
Figure 6: Using FNFMwith IBM stock, reinvesting dividends, generated an annualized return
over the test period of 20.9%.

using the FNFM. Figure 6 illustrates the balance of the test portfolio for IBM with the FNFM
parameters set to NumA= 1, NumL = 3, for a period of 6 years.

5. A Neurofuzzy Methodology for Prediction-Based Option Trading

Strategies
As mentioned in Section 1, an option is a financial instrument, a form of contract between two
parties, thatis traded in the world's financial markets, much like common stocks or other entities
(e.g. commodities, currencies, etc.). We areconcemed here with basic options, calls and puts and
combinations of them, such as straddles. The interested reader can find an in-depth discussion of
options and financial derivatives in (Hull, 1997)

The payoff of an option is the value of the contract at the expiration date. Prior to the expiration
date the option is traded in the financial markets as any other asset (e.g. stock) and its price fluctu-
ates based on a number of economic and other parameters. For example an S&P 500 Dec '98 650
Call option or an S&P 500 Dec'98 650 Put could be bought today; its value at the expiration date
(by convention the third Friday of December 1998) will depend on the actual price of the underly-
ing asset, which in this case is the S&P 500 index. Observe that the value of the S&P SOD is not
known with certainty, since, it changes as the prices of the stocks contributing to the index (500
stocks ofD.S. corporations from various industrial sectors) change. Figure 7 illustrates the payoff
of call and put options for a strike price of $650.

15
 Call Value
Put Value

$650 S& SOO

$650 S&P 500
Figure 7. Payoff of a call (left) and a pilI option (right) for a strike price of $650.

The payoff value for the call is described in Equation 6

CVr = max( O,SP500 - 650), (6)

where SP500 stands for the S&P 500 price and CVr is the option value at the expiration date. The
payoff value for the put is given by Equation 7

PVr = max( 650 - SP500,O) (7)

where PVT is the option value at the expiration date. A straddle, which is one call and one put
packaged together as one option, has a payoff shown in Figure 8.

StraddleValue

Figure 8, Payoff of a straddle.

The payoff of a straddle at expiration is given in Equation 8.

SV =max(650-SP500, SP500-650) (8)

r
Prior to expiration, the value of an option V can be computed as a function of a number of parame-
ters. The task of pricing an option is referred to as option pricing. According to the widely used
Black & Scholes option pricing model (Black, 1973), the value is given by

v = V(S,t, cr,q,K,r, r -I) (9)

16
where: S is the price of the underlying asset, e.g. the S&P 500,
t is the current time,
q is the dividend yield,

ais the volatility of the underlying asset's returns,

K is the strike price,
r is the risk-free rate of return, and
T is the expiration of the option measured in years

By convention, options are written with duration in multiples of 3 months (0.25 years). K, q, T,
and r are usually inputs to the option trading strategies; K and T are constants known in advance, r
is taken to be the T_Bill l rate with matching duration as the duration of the option, and q is ap-

proximated by a constant dividend yield. S is described as a stochastic process and its value is

determined by the laws of supply and demand. The time to expiration is calculated as (T -t)/252

years 2 .

The volatility a, a measure of the variability of the time series, is taken to be constant in most

practical cases; it is found either as the volatility in the more recent past of the time series, or, as

the volatility implied from other option prices quoted in the financial markets. It is known that a

model for future volatility that assumes the volatility is constant is not completely in agreement

with the Black & Scholes modeling ("smile-effect", (Hull, 1997)). A successful trading strategy,

which uses a neurofuzzy volatility predictor, can be built based on this deficiency, i.e., of using
historic or implied volatility to price options. If options are priced today using either of the two

estimates mentioned above (historic or implied volatility), and their price fluctuates according lO
the unknown future volatility behavior, a reliable prediction of future volatility can be of tremen-
dous value; it can be used to identify over-priced or under-priced options and execute appropriate
trades. The option value is an increasing function of volatility; the higher the volatility, the higher
the option value and vice-versa Consequently, a system that would provide an estimate [or the

1 T-Bi//s are treasury bills, short temJ US government debt that is considered to be free ofrisk. An alternative
would be to use Money Market account yields or yields on Certificates ofDeposit (CODs). However, it is common
fracfice to approximate the short term risk-free rate ojreturn by the rate offeredfrom T-Bills.
In order to convert in years, we use the number oftrading days in a year and not the actual days. A widely used
approximation for the trading days is 252 per year (HuJl, 1997).

17
future volatility trend could be used to build a trailing strategy, based on options, that would pro-
vide aoove normal returns.

Trading Strategies using Options based on Volatility Prediction

If we know that future volatility is going to be increasing (up), with a reliability of some degree,
then we could buy options. Since the options are selling at current volatility estimates, their price
will increase accoriling to the predicted upward movement of volatility. If the prediction is realized,
we stand to gain, ceteris paribus, the difference ~ - Va where uR denotes the predicted vola-
R 0

tility and 00 the volatility currently used by the market. This will be realized if nothing else
changes in this period.

Recall that the option value is a function of a number of other parameters which change over time,
and in particular, the time to expiration and the underlying asset price. Intuitively, a prediction of
higher volatility implies that the asset price is going to move; it does not however specify in which
direction. Consequently, if, as a reaction to the prediction for higher volatility we buy call options
and the asset price moves down, we might lose. On the other hand if we buy put options and the
asset price goes up we might also lose. A "safer" way to react on a prediction for higher volatility
would be to buy straddles, since, whichever way the price of the underlying is going to move we
are going to profit assuming that the prediction of higher volatility is realized.

In the Black & Scholes option pricing framework, the risk that incurs from an option position (i.e.
the risk of losing the initial investment) can be eliminated by appropriate transactions. The elimi-
nation of this risk is called hedging and involves getting a position in the underlying asset and the
risk-free asset (T-Bills). The size of the position, i.e. the number of shares, is determined as the
first order partial derivative of the value function V with respect to the underlying asset. This type
of hedging strategy is referred to as "delta" hedging and prescribes that a portfolio consisting of the
option and 11 shares of the underlying would not have any risk from movements in the price of the

underlying asset. L1 is determined by the pricing model and it is also an increasing function of
volatility. Hedging can be viewed as an insurance and as such has some cost. This cost reduces
the gains realized (and the risk). In an option trading strategy that also involves selling options
apart from buying them, a reliable volatility prediction could be used to reduce the necessary
hedging, and thus reduce costs. Consequently, a trading strategy based on future volatility predic-
tion can create profits both from the change in value, but also from an appropriate reduction in the
size of hedging (i.e., smaller share number than prescribed by il).

18
Crisp interpretation of the prediction risk for volatility trends is very difficult The fuzzification of
these two variables, i.e., the risk (reliability) of the prediction and the prediction itself, provides a
convenient to use and conceptually promising approach to fully exploiting neural network predic-
tive capabilities in the decision making process. Figure 9 illustrates a two-dimensional fUZZy rela-
tion involving the two prediction variables: risk and volatility. The volatility trend is described as
up, down or same and the risk level of the decision as low, average and high. The membership
functions of the resulting fuzzy numbers are described by correspondingly subscripted It's. The
infinite crisp decision space is transformed into a coarser, more parsimonious fuzzy space, where
fuzzy logic can be used to make inferences and come up with accurate recommendations to buy,
sell or "do-nothing." The pyramidal shape of the fuzzy rule at the center of Figure 9, for example,
is interpreted as a "do-nothing" recommendation. Furthennore, the fuzzification of the decision
space reduces the complexity of the neural network.

The neurofuzzy volatility prediction system is trained to anticipate the changes in the volatility
of the returns of the S&P 500 system and provides recommendations that are used to implement
short term option trading strategies. The input to the prediction system, i.e., the financial time

series, is the daily close value of the S&P500 index. The time series used is shown in Figure 10.

fuzzy prediction

~dm.m
1.0

lJ./uSh ,
- -.:.:.. - ---;,/ volatility
:. // trend
----;-.;.,.
,
---/
:~="7-""",,",,,;..;..;.-,.;..;.=;,..v:---,/
-----~-- -----~-~~-- -----/
risk

Figure 9: Fuzzy decision universe.

19
 IndexLeveI
600 ,--~--~--~---~--~--~--~-- __------,
.60
.00
.60
'00
260
200
,60
, 00
60
°0~~"'2S0?;'''0:=::::::;':;;'=0;;0::::::.~.;;;.;;:.:::::;;.~.:C'';;-----;'''''''0",0;;--'''2'''O:';;;'--;-'-;';:COO",O;;---;-:";;:OO:.;;;o--;-,;c]eooo
NumberofSamples

Figure 10, The Standards and Poor's, S&P 500 lime series from 1928 to 1993.

Preprocessing of the Time Series Input

The S&P 500 time series that is used (17,604 daily samples) is pre-processed before used by the
volatility prediction system. The pre-processing is done as follows. Let us denote by
St, I = 1, ...,17604 the S&P 500 daily closing prices. The logarithm of these data points is
s! ::: log St. Hence, the difference of successive data points is

20
SI" =St" - S(_1 st--1 )
= log - (10)
( S,_I

where it is assumed that sO = O. Denoting by NllmA the time frame, the time series is further

transformed as

(11)

where std denotes the standard deviation, and t= NumA, ",,17604. The time series is scaled by

x =10 4 sm. (12)

t t

Scaling facilitates neural network processing and training. The result of the pre-processing is a
modified time series, shown in Figure 11.

The time series is partitioned into three sets: the training set, the test set, and a third set which is left
completely untouched. Both the test and the unseen sets are out-oC-sample sets. The training sel
and the test set are drawn from the last 6000 elements of the time series, that is, from 2/5/1970 to
10/29/1993. The training set is comprised by the odd numbered elements and the test set by the
even numbered elements.

The training set is further partitioned into three subgroups: the lip, the same, and the down, of
NllmA-sized vectors. The up set contains the input vectors which precede an increase of the index
volatility in excess of 16.5%. The same set contains input vectors which precede absolute change
less than, or equal, to 16.5% and the down set contains input vectors which precede a decrease in
volatility of more than 16.5%. The three different sets are used to train three corresponding neural
networks, with one hidden recurrent sigmoidal layer and one linear output layer. The training of
the three networks was done for 15000 epochs on the training data and an error threshold (sum
square error) of 10- 3.

21
 Volatility Changes

O."'-~--_--_-~--_-
_ _- _ - - _ - "
0.11>

_0,1:5

-'2
-Cl.21>~,---;,;;;O:;'OO"'''O:O''O'-;'''O;';;O'O--;''''OO:;;O''-'''O''O;;;OO''''''''''O;;;O'C""O:OO'O""OO~O;;;o""oJeooo
Number of Samples

Figure 11. Modified S&P 500, time series from 1928 to 1993 (after pre-processing).

Crisp Post-Processing of the Neural Network's Output

We train three individual neural networks, called NNup, NNdowll and NNsame, on appropriately
selected training subsets, up, down, and same, in order to recognize patterns that precede the fol-
lowing movements in volatility changes: a) the NNup predicts a higher than 16.5% increase over
the previous day, b) the NNsame predicts change of less than 16.5% in either direction, and c) the
NNdown predicts a reduction of 16.5% or more. We use the relativeRMS error,

(13)

22
where, k is the number of samples, xk the true output, and xk the predicted value, to study the

perfonnanceof the system on the three sets, the train set, the test set, and the completely unseen
set. We tested five different rules for combining the output of the neural networks. The output of
NNup is denoted with II, of NNsame with s, and of NNdown with d. The rules are:

• Rule 1: Use the aritbrneticmean of the three outputs, i.e. (u+s+d)/3.

• Rule 2: Use only u.
• Rule 3: Use only s.
• Rule 4: Use only d.
• Rule 5: Exclude those outputs that predict something out of the range the were trained, e.g., if
u < 16.5% discard it~ of those remaining, take the arithmetic average.

Table 2 summarizes the relative RMS on the three tested inputs.

Table 2. Relative root mean square error obtained by the three networks during testing.

Traininl! Set Test Set Out-of-SamDle Set

Rule 1 0.3920 0.3775 0.4608
Rule 2 0.3388 0.3287 0.3350
Rule 3 1.0586 1.0397 1.4032
Rule 4 0.2788 0.2667 0.2555
Rule 5 0.2918 0.2758 0.2565

The approximation of each three neural network to the financial time series is shown Figure 12.

Fuzzy Post-Processing of the Neural Network's Output

We train three neural networks, in a similar manner as discussed above, only this time the output
of the neural networks is not a crisp number but a fUZZy number. The three neural networks,
fNNup, JNNdown, and JNNsame, provide fuzzy predictions for the change in volatility fllp,
fdoWll, jsame. The membership functions of the three fuzzy outputs are triangular.

Each network gives three outputs, the left ( I), center ( c ), and right ( r) which represent the
membership function of the prediction. For the output of a particular network to be meaningful the
condition 1< c < r must be satisfied. If it is not, the output is discarded. Recall that each network
is trained to recognize a particular class of input patterns. We tested the following post-processing
rules to evaluate the performance of the fuzzy neural network over the three sets, training, testing
and unseen.
• Rule 1: lfall three outputs do notsatisfy the constraint of shape predict nothing. Else
• Rule 2: fJall three satisfy the constraint then discard the one with the most distorted member~
ship function and take the fuzzy mean of the other two. Else

23
• Rule 3: Ifiess than three satisfy the constraint then take the fuzzy average of the those that
satisfy the constraint.

In all cases that there is prediction, the prediction is the centroid of the resulting fuzzy number. In
choosing the most distorted membership function to discard, in Rule 2, we use the criterion:

• Criterion: Compare the total support of the fuzzy number. Remove the one with biggest
support. The support is computed as (r -l).

Table 3. Relative root mean square error obtained by the three networks during testing

Trainin Set Test Set Out-or-Sam Ie Set

Criterion 0.3024 0.2864 0.2781

Evaluating the prediction capabilities using option trading strategies

In order to evaluate the predictive capabilities of the neural network we use its recommendation in
an option trading strategy involving straddle "at-the-money" options on the S&P 500.

For brevity, and in order to avoid introducing intrinsic financial knowledge, we considering only a
simple trading strategy, based on straddles made up of call and put options on the SP500 index.
For pricing the options we are using the Black & Scholes pricing model. The formulas for the
prices of the call and put options used are:

P(S, t) ~ Ke -r(T-I) N(-d ) _ lie -q(T-I) N(-d ) (14)

2 t
C(S, 0 =lie -q(T-I) N(dl) - Ke -r(T-0 N(d2)

2
S a
log(X) +(r-q+"2XT-t)
where, d j = Fi
a T-t
d
2
~ d -a~
1
(15)

In the above formulas q denotes the constant dividend yield.

For finding the current market price of the options, i.e. the price which they can be purchased, we
make the assumptions listed. in Table 4. The trading strategy simulation is summarized as follows:
1. Start wi th an initial amount (cash balance) of $ 1,000.
2. At the beginning of each day, mark-to-market (i.e. compute the market value) of the port-
folio of options.
3. Liquidate the portfolio, if it contains holdings other than cash.
4. If the neurofuzzy volatility prediction system recommends "hold" do nothing.
5. If the volatility prediction system recommends "buy", purchase as many "at-the-money"
straddles as the cash balance permits.
6. Go to step 2.

24
 to

jNNup , ..\.---:
,
~'''~''-l
'~-'---'- '-----'-'
'~'-'-.
...._.
°
2400 2420 2440 "00 2480 2500 2520 2540 2560 2580 '000
4

jNNsame ,
,
2400 2420 2440 2480 2500 2520 2540 2560 2580
"00 '000
4

jNNdown

r._.
.~-~-
-'- ~----,-'
r"'-
_..... ,-
2°400 2420 2440 2400 2480 2500 21620 21640 2560 2680 '000

Figure 12. Comparing the modified S&P 500 time series with the neural network predictions
(dashed lines denote predictions, solid lines denote actual time series data): (a) The network trained
with patterns of upward movement of the index generally over-predicts the time series. (b) The
network trained with patterns of "about the same" movements of the index generally stays close to
the index. (c) The network trained with patterns of downwards movement generally under-predicts
the actual S&P 500 time series.

We simulate this trading strategy on the out-or-sample set for approximately 10,000 days. The
comparison is done between a portfolio fully invested in the S&P 500 index and the option portfo-
lio managed using the neurofuzzy volatility prediction system. We keep track of both portfolios,
and their market value at the beginning of each trading day. Returns from investing into the S&P

25
500 are shown in Figure 13. Returns from the option trading strategy using the volatility predic-
tion system for the same period and with the same initial investment are shown in Figure 14.

In the case of the S&P 500 portfolio we get at the end of the period a portfolio balance of $3,862.
In the case of the option portfolio we get at the end of the period a portfolio balance of $625,120.
The resulting performance is substantially better for the option trading strategy which uses the
neurofuzzy volatility prediction system. The higher volatility exhibited by the daily balances of the
option portfolio is in part expected and explained by the leverage of options compared to stocks.
Each option gives the right to buy or sell, depending on the type of option, 100 shares of a certain
asset. Generally, a $1 change in the price of the asset, results in a lOG-fold change in the price of
the option. Also, the choice to fully invest the available cash balance at all times contributes to the
balance volatility. The 75% reliability of the neurofuzzy predictor leaves a 25% margin for false
predictions. Note that correct prediction of direction for volatility movement says nothing about
the magnitude of the movement. That means the gains from correct predictions and losses from
incorrect predictions vary. A trade made based on a faulty prediction results in a loss, which in
order to be covered requires twice the increase. For example, a 50% loss must be followed by a
100% profit in order to reach the same balance as the one before the faulty prediction. In general,
the above trading strategy demonstrates the JX)tentiai of the methodology used to predict the volatil-
ity movements. In a realistic framework it would be combined with more sophisticated manage-
ment and hedging activities in order to reduce the variability of the returns. The success of the
option trading strategy is primarily attributed to the prediction of the future volatility. In all cases
the value of the option portfolio is more than that of the S&P 500 portfolio.

Table 4. Assumptions for the parameters used to price options.

S The price of the index is taken to be the closing price for the day; i.e. we exe-
cute all trades simultaneously at the end of the day. .

K The strike price is taken to be the "at-the-money" value, i.e. same as the price
of underlying S&P 500 index.
a The volatility that the market uses to price the options is the historic volatility
for the last 30 days.
r The risk-free interest rate is taken to be the T-Bill rate, approximately 5%.
q The constant dividend yield of the index is taken to be approximately 2.5%.
(T-t) The duration of the options used is 1 month. This in practice translates into
using the option with the closest expiration date.

26
 450'O'----~----~---~---~~---
400

350

3000
250

200

6000 8000 10000

Figure 13. The returns from investing into the S&P 500 between 1928 and 1968 (the vertical
axis indicates the portfolio balance in dollars).
o
3.5 x 10

2.5

1.5

0.5

O!--~~~~~~----J
o 2000 4000 6000 8000 10000

Figure 14. Returns from investing according to neurofuzzy predictions of the market for the pe-
riod between 1928 to 1968 (vertical axis indicates portfolio balance in millions or dollars).

6. Conclusions
Neural networks, when properly configured, offer superb pattern matching capabilities that can
be used for predicting the ups and downs of financial indicators such as the S&P 500 index. The

fuzzification of such neural predictions may lead to robust and overall successful trading strate-

gies.

27
We have presented in this paper two cases in which neurofuzzy prediction approaches are used

successfully to predict financial variables in a way that is meaningful from an investment point of

view, i.e., resulting in profitable trading of stocks and options. However, another interpretation

of our results is possible. They could be a challenge to standard option pricing models. The high

returns our simulations yielded would surely have led in reality to higher option prices - higher

than the standard models predict.

In the present paper our simulations involved standard option pricing models and neurofuzzy

approaches to strategic trading. Future work involves replacing the standard pricing models with

neurofuzzy approaches to predicting prices, in order to gain insight into the impact of

neurofuzzy approaches as they come to be used in actual markets.

7. References

Ahmad J., Fatmi, H., A Quadric Neural network System for Prediction of Time Series Data, Proc.
IEEE World Congress in Neural Networks, pp. 3667-70. Orlando, FL, 1994.
Anderson, P. W., Arrow, KJ., Pines, D., Eds., The Economy as an EvolVing Complex System,
Addison-Wesley, Redwood City, California, 1988.
Black, E, Jensen, M. and Scholes M, The Capital Asset Pricing Model: Some Empirical Tests, in
Studies ill Theory of Capital Markets, Jensen M., Ed., Preager, New York, 1972.
Black, F. and M. Scholes, The Pricing of Options and COJlXlrate Liabilities, Journal of Political
Economy, 81, pp. 637-59, May-June 1973.
Farber, J. D. and Sidorowich. J.J., Can New Approaches 10 Nonlinear Modeling Improve Eco-
nomic Forecasts? in The Economy as an Evolving Complex System, Anderson, P.W., Ar-
row, K.J., Pines, D., Eds., pp. 99-115, Addison-Wesley, Redwood City, Calirornia,1988.
Hobbs, A., Bourbakis, N.G., A Neurofuzzy Arbitrage Simulator for Stock Investing, Proceed-
ings of tile IEEE/IAFE 1995 Computational Intelligence for Financial Engineering, New
York,Apri19-11,1995.
Hull, John C., Options. Futures, and Other Derivatives, Third Edition, Prentice Hall, 1997.
Konstenius, J., Mirrored Markets, Proc. IEEE World Congress in Neural Networks, PP 3671-
75, Orlando, A.-, 1994.

28
Kung, S.Y., Digital Neural Networks, Prentice-HallIntemational, London, 1993.
Lapedes A., and Farber R. Nonlinear Signal Processing Using Neural Networks: Prediction
al/d System Modelil/g, Los Alamos National Laboratory Technical Report, LA-UR--87-2662,
June 1987.
leBaron, B., Weigend, A.S., Evaluating Neural Network Predictors by Bootstrapping, Proc. of
the Imem. Con! On Neural Information Processing (ICONIP '94), pp 1207-1212, Seoul,
Korea, 1994.
Lippit, V.D., The Golden 90's, Real World Macro: A Macroeconomics Reader from Dollars &
Sense, 7th Edition, pp 73-75, Dollars & Sense, Sommerville, MA, 1990.
Lowe, D., Novel Exploitation of Neural Network Methods in Financial Markets, Proc. IEEE
World Congress in Neural Networks, pp3623-27, Orlando, FL, 1994.
Peters, E., Fractal Market Analysis, John Wiley & Sons, New York, NY, 1994
Radzicki, M.J., Institutional Dynamics, Detenninistic Chaos, and Self-Organizing Systems, Jour-
Iud of ECOlwmic Issl/es, Vol. XXIV, No.1, pp. 57-102, March 1990.
Ramsey, J.B., Syers, C. L., and Rothman, P., The Statistical Properties of Dimension Calcula-
tions Using Small data Sets: Some Economic Applications, Imernational Economic Review,
Vol. 31. No.4, pp. 991-1020, 1990.
Refenes, P., Abu-Mustafa., Y., Moody, J.E., Weigend, A.S., Eds., Neural Networks in Finan-
cial Engineering, World Scientific, Singapore, 1996.
Sharpe, W., Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk,
Joumal of Fil/al/ce, Vol. 19, pp. 452-442., 1964.
Trippi, R, Lee,K., Artificial Intelligence in Finance & Investing, Irwin Publishing, Chicago,
1996.
Tsoukalas, L.H., Uhrig, RE., Fuzzy and Neural Approaches in Engineering, John Wiley &
Sons, New York, NY, 1997.
Wilson C.L., Self-OrganizingNeurai Networks for Trading Common Stocks, Proc. IEEE World
Congress in Neural Networks, pp. 3657-61, Orlando, FL, 1994.

29