master's thesis
Neural networks and decision trees for classification and prediction stock prices

Marijana Škrobak (2017)
Josip Juraj Strossmayer University of Osijek
Faculty of Economics in Osijek
Chair of Quantitative Methods and Informatics
Metadata
TitleNeuronske mreže i stabla odlučivanja za klasifikaciju i predviđanje cijene dionica
AuthorMarijana Škrobak
Mentor(s)Marijana Zekić Sušac (thesis advisor)
Abstract
Cilj ovog rada je izraditi modele stabla odlučivanja i neuronske mreže za problem predviđanja i klasifikaciju kretanja dionica Zvijezda d.d., pojasniti glavne pojmove modela neuronske mreže i stabla odlučivanja, te pojam dionice. U radu će se na konkretnom primjeru obrađenom u programskom paketu Statistica dobiti i opisati rezultat kretanja dionica Zvijezda d.d. odnosno njihov rast i pad. Ovo istraživanje će koristiti studenti i ostali zainteresirani za područje uporabe metoda neuronskih mreža i stabala odlučivanja u poslovanju. Dobivenim rezultatima pokazalo se da je uspješniji model klasifikacije koji je napravljen pomoću tangens hiperbolne funkcije i ukupna stopa klasifikacije mu je 100%. Veća je za 7,14% u odnosu na rezultate dobivene modelom stabla odlučivanja. Problem klasifikacije je osim ukupne stope klasifikacije pokazao i rezultate analize osjetljivosti izlazne varijable na ulaznu. Dobivenim rezultatima došlo se do saznanja da najveći utjecaj ulazne varijable na izlaznu ima varijabla „Zadnja“ s koeficijentom osjetljivosti 32339602, dok najmanji utjecaj ima ulazna varijabla „Količina“ s koeficijentom osjetljivosti 146039,8. Rezultati osjetljivosti dobiveni modelom stabla odlučivanja prikazuju kako je „Broj transakcija“, varijabla koja je najmanje značajna u odnosu na izlaznu varijablu s koeficijentom osjetljivosti 0,074827. Najveći utjecaj ulazne varijable na izlaznu u modelu stabla odlučivanja ima varijabla „Promjena“ s koeficijentom osjetljivosti 1,000000.
Keywordsneural networks decision tree stocks multilayer networks input and output variables
Parallel title (English)Neural networks and decision trees for classification and prediction stock prices
Committee MembersMarijana Zekić Sušac (committee chairperson)
Josip Mesarić (committee member)
Đula Borozan (committee member)
GranterJosip Juraj Strossmayer University of Osijek
Faculty of Economics in Osijek
Lower level organizational unitsChair of Quantitative Methods and Informatics
PlaceOsijek
StateCroatia
Scientific field, discipline, subdisciplineSOCIAL SCIENCES
Economics
Business Informatics
Study programme typeuniversity
Study levelgraduate
Study programmeBusiness economy; specializations in: Business Informatics
Study specializationBusiness Informatics
Academic title abbreviationmag.oec.
Genremaster's thesis
Language Croatian
Defense date2017-02-15
Parallel abstract (English)
The aim of this study is to create models of a decision tree and neural networks for problem prediction and classification movement of stock Zvijezda d.d., to explain the main concept of the model neural networks and decision trees, and the concept of stocks. The study will be created on concrete examples made in the software package Statistica, where it will get and describe the results of the movement of the stock Zvijezda d.d. and it's rise and fall. This research will be used by students and others interested in the application methods of neural networks and decision trees in business. Results showed that the successful model classification is made using a hyperbolic tangent function and the overall classification rate is 100%. An increase of 7.14%, compared to the results obtained from the model decision tree. The problem of classification with the overall rate of classification showed the results of the sensitivity analysis of output variables to the input. The results found that the greatest impact of the input variables to the output variable have the "Last" with a coefficient of sensitivity 32339602, while the smallest effect has the input variable "Quantity" and the coefficient of sensitivity 146039.8. The results of sensitivity analyses obtained a model decision tree that shows that "The number of transactions" variable is the least significant, with relation to the output variable with a coefficient of sensitivity 0.074827. The greatest impact of the input variables to the output in the decision tree model has a variable "Change" with a coefficient of sensitivity 1.000000.
Parallel keywords (Croatian)Neuronska mreža stablo odlučivanja dionice višeslojna mreža ulazne i izlazne varijable
Resource typetext
Access conditionAccess restricted to students and staff of home institution
Terms of usehttp://rightsstatements.org/vocab/InC/1.0/
URN:NBNhttps://urn.nsk.hr/urn:nbn:hr:145:824008
CommitterGordana Kradijan