magistrsko delo
Blaž Košenina (Author), Vitomir Štruc (Mentor), Janez Konečnik (Co-mentor)

Abstract

Trgovanje na Forex trgu je vsako leto bolj priljubljeno, saj v Sloveniji ni obdavčeno. Na Forex trgu trgujemo z valutnimi pari, na njihovo vrednost pa vpliva ogromno dejavnikov, ki jih težko predvidimo. Največji izziv Forex trga je nedvomno njegova volatilnost. V našem magistrskem delu razvijemo algoritme strojnega učenja, s katerimi lahko pridobimo uporabne informacije za trgovanje na Forex trgu. V okviru magistrskega dela izdelamo regresijski model, ki je sestavljen iz več izhodne Bayesove linearne regresije, s katerim napovedujemo začetno, končno, največjo in najmanjšo ceno v enem dnevu. Model je sposoben napovedovati cene za več dni vnaprej. Z izbranim Bayesovim modelom lahko določimo intervale zaupanja, saj kot rezultat dobimo verjetnostno porazdelitev izhodnih vrednosti. Poleg regresijskega modela razvijemo tudi inteligentnega agenta, ki s spodbujevalnim načinom učenja samodejno sprejema transakcije v simuliranem tržnem okolju Forex trga. Agenta sestavlja algoritem globokega Q-učenja (ang. Deep Q-Learning Network), ki je povzročil revolucijo na področju spodbujevalnega učenja. Predlagani rešitvi ovrednotimo na vseh evro križih med letoma 2010 in 2021. Izkaže se, da je naš regresijski model zadovoljivo deluje zgolj pri napovedovanju cen za en dan vnaprej, kjer ima malenkost slabši rezultat od naivnega modela le pri napovedovanju končne cene. Inteligentni agent doseže boljše rezultate od izhodiščnih strategij. Portfelj agenta se v 11ih letih poviša za 11,7 odstotka. Učenje agenta brez upoštevanja transakcijskih stroškov privede do manj uspešnih rezultatov. Očitno agent z vključenimi transakcijskimi stroški izdela bolj robustno in zanesljivo strategijo ter izvede manj transakcij. Algoritmi strojnega učenja so nepogrešljiv sestavni del robustne trgovalne strategije, kljub temu pa je za dobičkonosno avtonomno trgovanje potrebno veliko znanja, poskusov in procesorske moči.

Keywords

Forex trg;linearna regresija;Bayesova linearna regresija;inteligentni agent;spodbujevano učenje;DQN;magisteriji;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FE - Faculty of Electrical Engineering
Publisher: [B. Košenina]
UDC: 339:077:004(043.3)
COBISS: 69039363 Link will open in a new window
Views: 455
Downloads: 195
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Trading on the Forex market with machine learning algorithms
Secondary abstract: The foreign exchange market (Forex) is a global market, where currency pairs are traded. Trading on the Forex market has been slowly gaining popularity over the past few years and is not taxed in Slovenia. Currency pairs are traded on Forex market and their value is affected by a huge number of factors that are difficult to predict. The biggest challenge of Forex market is undoubtedly its volatility. In this master's thesis, we focus on the development of machine learning algorithms, which provide useful information for trading in the Forex market. In the first part of our master's thesis, we used multiple output Bayesian linear regressions model to predict open, close, high and low daily price. Model can predict prices several days in advance. Bayesian model outputs the probability distribution of target variables, which we can use to determine confidence intervals. In the second part, we have developed a reinforcement learning agent to trade on a simulated Forex exchange environment. The agent is built using Deep Q-Learning Network algorithm, which has revolutionized reinforcement learning. The proposed solutions are evaluated at all euro crosses between 2010 and 2021. It turns out that our regression model outperforms existing models when predicting one day in advance, where it only slightly underperforms when predicting close price compared to the naive model. On the other hand, our reinforcement learning agent achieves better results than our hand-crafted strategies. The agent's portfolio has grown by 11.7 percent in 11 years. Training our agent without transaction cost leads to poor performance. Apparently, transaction costs force an agent to locate a more resilient and reliable strategy when market change occurs. Machine learning algorithms are an essential part of a robust trading strategy, nevertheless building a profitable autonomous trading system requires a lot of knowledge, trial & error and processing power.
Secondary keywords: Forex market;linear regression;Bayesian linear regression;intelligent agent;reinforcement learning;DQN;
Type (COBISS): Master's thesis/paper
Study programme: 1000316
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za elektrotehniko
Pages: XVIII, 54 str.
ID: 13088996