Strojno učenje za klasifikaciju financijskih vremenskih nizova

master's thesis

• Title (english): Machine Learning for Financial Time Series Classification
• Author: Ivan Modrušan
• Mentor: Zvonko Kostanjčar (mentor)
• Committee member: Zvonko Kostanjčar (predsjednik povjerenstva)
• Committee member: Dražen Jurišić (član povjerenstva)
• Committee member: Stjepan Begušić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-08, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U sklopu rada su učeni modeli za predviđanje kretanja sastavnica tržišnog indeksa S&P 500. Opisani su osnovni koncepti strojnog učenja i teorijska podloga modela slučajnih šuma, potpuno povezane, konvolucijske i povratne neuronske mreže. Problem predviđanja je formiran kao klasifikacija tako da su modeli predviđali jednu od tri kategorije: rast, pad i stagnacija vrijednosti dionice. Kategorije su određene korištenjem metode trostruke granice i metode s fiksnom vremenskom udaljenosti. Osim podataka kretanja cijena dionica, modelima su dani i neki često korišteni indikatori za analizu tržišta. Podaci su zatim posloženi u prozore i modificirani tako da se modeli mogu učiti istovremeno s više dionica. Na kraju su uspoređene točnosti učenih modela i razlike povrata koji su ostvarili strategijom trgovanja. Prema rezultatima pokazano je da najbolje performanse za rješavanje ovakvog tipa problema postižu jednostavni modeli. Također, performanse modela su bile bolje kada su učeni podacima označenim metodom trostruke granice na kojima je i postignuta najveća profitabilnost i točnost korištenjem LSTM povratne neuronske mreže.

Abstract (english)

This thesis describes a way to use machine learning models for predicting the movement of stocks in the S&P 500 market index. The basic concepts of machine learning as well as theoretical basis of random forest, fully connected, convolutional and recurrent neural networks are described. The prediction problem was formed as a classification so that the models predicted one of three categories: growth, decline, and stagnation in the price of stock. Categories were determined using the triple barrier method and the fixed time horizon method. In addition to data of stock price movements, the models are given some frequently used indicators for market analysis. The data is then arranged in windows and modified so that models can be learned simultaneously with multiple stocks. Finally, the accuracy of the learned models and the differences in returns achieved by the trading strategy were compared. According to the results, it is shown that the best performance for solving this type of problem is achieved by simple models. Also, the performance of the models was better when they were fitted using data with labels created using triple barrier method. The highest profitability and accuracy was achieved using LSTM recurrent neural network.

• Keywords: strojno učenje; financije; klasifikacija; S&P 500
• Keywords (english): machine learning; finance; classification; S&P 500
• Language: croatian
• URN:NBN: urn:nbn:hr:168:428995
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 09:34:22