Thesis (MBA)--Stellenbosch University, 2002. / ENGLISH ABSTRACT: The dream of finding the ultimate tool for forecasting market instruments like share
prices has long eluded investors throughout the world. Various forecasting techniques
have been examined with a view to helping the investor or analyst to gain a better
understanding of price behaviour in the open market. These techniques have been
based mainly on traditional statistical analysis of data to forecast price behaviour.
Though used by almost all serious investors, these techniques have yielded limited
success as investment instruments. The reason for this is that most of these methods
explored linear relationships between variables in the forecasting model, while in fact,
most relationships found between variables in the share market are non-linear.
Neural networks present a unique opportunity for the investor to overcome this
problem. Neural networks are mathematical models of the human brain and have the
ability to map complex nonlinear relationships between data sets.
This study focuses on developing a neural network model to predict the price changes
of the ALSI index on the JSE one and five days into the future. The results of the
neural network model were then compared to forecasting results obtained by using a
traditional statistical forecasting technique namely ARIMA modelling. The study
found that the neural network models did not significantly perform better than the
ARIMA models.
A further test was done to determine the performance of the five-day forecasting
model when analysing different time windows within the given data set. The test
indicated that the model did perform better when using the inputs of certain time
frames. This indicates that the neural network model needs to be updated regularly to
ensure optimum model performance.
The results of the neural network models were also used in a trading simulation to
determine whether these results could be applied successfully to trading the ALSI
index on the JSE. Unfortunately, the results of the trading simulation showed that using the neural network results as trading strategy yielded poorer results than using a
buy/hold investment strategy.
It can therefore be concluded that, although the neural network models performed
relatively well relative to traditional forecasting techniques in forecasting the ALSI
index, the forecasts were still not accurate enough to be useful as inputs in a trading
strategy. / AFRIKAANSE OPSOMMING: Die droom om die perfekte vooruitskattingsinstrument te vind om die prysgedrag van
verskillende markinstrumente vooruit te skat, ontwyk al generasies lank die meeste
beleggers. Verskillende tegnieke is al ondersoek om die belegger te help om ’n beter
gevoel van prysveranderinge in die vrye mark te verkry. Die meeste van hierdie
tegnieke het gefokus op tradisionele statistiese vooruitskattingstegnieke.
Alhoewel hierdie tegnieke wêreldwyd deur investeerders gebruik word, was hierdie
metodes se sukses as investeringsinstrument beperk. Die rede vir hierdie beperkte
sukses lê in die feit dat hierdie tegnieke slegs die lineêre verwantskappe tussen
veranderlikes gebruik het om voorspellings te maak, terwyl die meeste verwantskappe
wat tussen veranderlikes in die vrye mark bestaan, nie-lineêr is.
Neurale netwerke bied ’n unieke geleentheid vir beleggers om bogenoemde probleme
te oorkom. Neurale netwerke is wiskundige modelle wat op die werking van die
menslike brein gebaseer is en besit die vermoë om komplekse nie-lineêre
verwantskappe tussen datastelle te herken.
Hierdie studie fokus op die ontwikkeling van ’n neurale netwerk(e) om die
prysverandering van die ALSI indeks op die JEB een en vyf dae in die toekoms
vooruit te skat. Die resultate van die neurale netwerk model is verder vergelyk met
die resultate van tradisionele statistiese vooruitskattingstegnieke soos byvoorbeeld
ARIMA tegnieke. Die studie het gevind dat die neurale netwerk modelle nie
beduidend beter gevaar het as die ARIMA modelle in die vooruitskatting van die
ALSI indeks in beide die een- en vyfdag vooruitskattings nie.
’n Verdere toets is gedoen om die toepaslikheid van die gekose vyfdagmodel op
verskillende tydvensters van die tydreeks te bepaal. Die toets het aangetoon dat die
model baie meer akkuraat is vir sekere tydvensters as vir ander tydvensters. Dit dui
dus daarop dat die neurale netwerk model gereeld heropgelei behoort te word om
seker te maak dat die model optimaal presteer gegewe die spesifieke insetdata. Die resultate van die neurale netwerk model is ook gebruik in ’n simulasie om te
bepaal of die resultate die belegger kan help om beter investeringsbesluite rakende die
ALSI indeks op die JEB te maak. Ongelukkig het die simulasie resultate gewys dat ’n
beleggingstrategie gebaseer op die neurale netwerk resultate swakker opbrengste
gerealiseer het as ’n koop/hou beleggingstrategie.
Ten slotte het die studie getoon dat alhoewel die neurale netwerk modelle relatief
goed in vergelyking met tradisionele statistiese modelle gevaar het in die
vooruitskatting van die ALSI indeks, hierdie vooruitskattings nie akkuraat genoeg is
om as inset tot ’n investeringstrategie gebruik te word nie.
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:sun/oai:scholar.sun.ac.za:10019.1/53086 |
| Date | 12 1900 |
| Creators | Van Niekerk, J. P. de T |
| Contributors | Gevers, W., Stellenbosch University. Faculty of Economic and Management Sciences. Graduate School of Business. |
| Publisher | Stellenbosch : Stellenbosch University |
| Source Sets | South African National ETD Portal |
| Language | en_ZA |
| Detected Language | Unknown |
| Type | Thesis |
| Format | 93 p. |
| Rights | Stellenbosch University |
Page generated in 0.0028 seconds