| Abstract | Stroj s potpornim vektorima pripada danas samom vrhu klasifikacijskih algoritama
strojnog učenja s uspješnom primjenom u rješavanju najraznovrsnijih problema. Kao prvi
algoritam proistekao iz statističke teorije učenja, svoju izvrsnu sposobnost generalizacije
zahvaljuje implementaciji principa strukturne minimizacije rizika, baziranog na simultanoj
minimizaciji empirijskog rizika i VC dimenzije, odnosno kapaciteta klase funkcija koje učeći
stroj implementira.
U radu se istražuje mogućnost predviđanja smjera kretanja cijena na tržištima
vrijednosnica primjenom stroja s potpornim vektorima pri čemu su kao ulazne varijable
korišteni tehnički indikatori, dok je izlaznu varijablu predstavljao predznak prinosa na
određeni dan u budućnosti. S obzirom da, osim o samome algoritmu, uspješnost klasifikacije
ovisi i o ostalim elementima sustava, ispitivan je utjecaj odabira značajki, različitih
kombinacija parametara, neravnoteže u podacima te duljine niza na rezultate klasifikacije.
Uspoređivane su različite evaluacijske mjere, a rezultati predviđanja testirani su i u simulatoru
trgovanja gdje se pokazalo da se bolji rezultat može dobiti kombinacijom više klasifikatora na
način da svaki od njih uči rješavati svoj zadatak.
Od tri burzovna indeksa, iako odabrana za eksperiment na temelju testova
predvidljivosti vremenskog niza, samo su kod jednoga u konačnici postignuti zadovoljavajući
rezultati s obzirom da se kao osnovna prepreka boljim rezultatima pokazala nedovoljna
prediktivna moć odabranih tehničkih indikatora. |
| Abstract (english) | Support vector machine today belongs to the very top of the machine learning
classification algorithm, with successful application in resolution of all kinds of problems. As
the first algorithm was the result of the statistical learning theory, it owes its excellent
generalization performance to the implementation of the structural risk minimalization
principle, based on simultaneous minimalization of the empirical risk and VC-dimension, i.e.
the capacity of the class of functions implemented by the learning machine.
In this paper, the possibility of prediction of the stock market price movement was
researched by means of implementation of the support vector machine, where technical
indicators were used as input variables, while the sign of the excess of returns on a specific
date in the future represented the output variable. Given that, apart from the algorithm itself,
the classification performance also depends on other system elements, the impact of the
choice of features on the classification results, i.e. different parameter combinations, data
imbalance, as well as the series length, were also examined. Different evaluation measures
were compared and the prediction results were also tested in a market trading simulator,
where it was shown that a better result could be obtained by the combination of multiple
classifiers, in the manner that each one learns how to solve its own task.
Out of the three stock market indexes, although chosen for the experiment on the basis
of the time series predictability tests, ultimately in only one of them satisfactory results were
achieved, given that insufficient predictive power of selected technical indicators proved to
represent the main obstacle in obtaining better results. |
