Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange
Thesis (MEng)--Stellenbosch University, 2023.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2023
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867613889773961216 |
|---|---|
| access_status_str | Open Access |
| author | Richards, Nicole |
| author2 | Burger, Eldon |
| author_browse | Burger, Eldon Richards, Nicole |
| author_facet | Burger, Eldon Richards, Nicole |
| author_sort | Richards, Nicole |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2023. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/126965 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA en_ZA |
| last_indexed | 2026-06-10T12:43:19.203Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2023 |
| publishDateRange | 2023 |
| publishDateSort | 2023 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/126965 Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange Richards, Nicole Burger, Eldon Thorsten, Schmidt-Dumont Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering. Reinforcement learning -- Johannesburg (South Africa) Automatic differentiation -- Johannesburg (South Africa) Stock exchanges -- Automation -- Johannesburg (South Africa) Thesis (MEng)--Stellenbosch University, 2023. ENGLISH ABSTRACT: Given the rapidly advancing capabilities of modern computers, there has been a considerable increase in interest in algorithmic trading. From reducing trading latency by purchasing highly sought-after property as close to stock exchanges as possible, to conducting research on the capabilities of the most recent artificial intelligence algorithms on the stock markets, market participants and academic researchers are all looking for innovative methods to achieve maximum returns with minimum risk. In this thesis, a reinforcement learning approach is adopted to determine the capabilities of such algorithms, along with a suitable neural network architecture, to make day trading decisions. A reinforcement learning agent learns through a process of trial-and-error. The agent performs various actions in order to determine which actions yield rewards, which reinforce desirable behaviour. The capacity of reinforcement learning algorithms to improve their decision-making over time through self-learning makes them naturally suitable for algorithmic trading. During the experiments conducted for this thesis, the reinforcement learning agents were trained and tested on the Johannesburg Stock Exchange. A day trading approach is taken, which means positions are never kept overnight, in an e ort to reduce trading risk. Furthermore, the available computational power is proposed as a limiting factor to determine the capabilities of such a reinforcement learning trading approach in a personal capacity when no special computational facilities are available. Various experiments are conducted to determine the impact of di erent reward functions, neural network architectures, and reinforcement learning algorithms on the reinforcement learning agents' day trading performance in terms of pro tability and risk. Additionally, hyperparameter optimisation is performed, which yields improved performance across all reinforcement learning agents. Certain inputs and reward functions are better suited to the default reinforcement learning algorithm parameters, and hyperparameter optimisation of the reinforcement learning algorithm parameters is required to make fair comparisons between the di erent reward functions and algorithms. When the number of stocks available on the Johannesburg Stock Exchange environment is increased, the nal account value that the reinforcement learning agent achieved also increased. The computational power limited the number of stocks that could be available within the Johannesburg Stock Exchange environment, but the results of this thesis serve as a proof-of-concept of reinforcement learning for algorithmic trading on the Johannesburg Stock Exchange. AFRIKAANS OPSOMMING: Daar is 'n aansienlike toename in belangstelling in algoritmiese handel. Hierdie belangstelling word gedryf deur die verbetering van die rekenaartegnologie. Besighede spandeer miljoene op hoogs gesogte besigheidseiendom wat so na as moontlik aan aandelebeurse is om handelsvertraging te verminder, en navorsing oor die mees onlangse kunsmatige intelligensie-algoritmes en hul vermo ens op die aandelebeurs neem jaarliks toe. Beide markdeelnemers en akademiese navorsers is op soek na innoverende metodes om maksimum opbrengste met minimum risiko te bereik. In hierdie tesis word 'n versterkingsleer benadering ge mplementeer om die vermo ens van versterkingsleer algoritmes, tesame met 'n geskikte neurale netwerkargitektuur, om daghandelbesluite te neem, bepaal. 'n Versterkingsleeragent leer deur 'n proses van probeer-en-fouteer. Die agent voer verskeie aksies uit om te bepaal watter aksies belonings oplewer, wat gewenste gedrag versterk. Die kapasiteit van versterkingsleer algoritmes om hul besluitneming oor tyd deur sel eer te verbeter, maak dit geskik vir algoritmiese handel. Tydens die eksperimente van die tesis word die versterkingsleeragent op die Johannesburgse aandelebeurs opgelei en getoets. 'n Daghandelbenadering word gevolg wat beteken dat aandele nooit oornag gehou word nie om handelsrisiko's te verminder. Verder, word die beskikbare berekeningskrag as 'n beperkende faktor voorgestel om die vermo ens van so 'n versterkingsleer benadering te bepaal wanneer geen spesiale berekeningsfasiliteite beskikbaar is nie. Navorsing en eksperimente word uitgevoer om die impak van verskillende beloningsfunksies, neurale netwerkargitekture en versterkingsleer algoritmes op die versterkingsleeragente se daghandelprestasie, in terme van winsgewendheid en risiko, te bepaal. Hiperparameter optimering word toegepas om verbeterde prestasie oor alle versterkingsleermiddels te verseker. Sekere beloningsfunksies en insette na die versterkingsleer agente is beter geskik vir sekere versterkingsleer algoritmeparameters, en hiperparameter optimering van die versterkingsleer algoritmeparameters is nodig om regverdige vergelykings tussen die verskillende beloningsfunksies en algoritmes te maak. Wanneer die aantal aandele beskikbaar op die Johannesburgse aandelebeurs verhoog word, verhoog die nale rekeningwaarde wat die versterkingsleeragent behaal. Die rekenkrag het die aantal aandele wat aan die versterkingsleeragente beskikbaar kon wees, beperk. Die tesis dien as `n konseptuele bewys van versterkingsleer vir algoritmiese handel op die Johannesburgse aandelebeurs. Masters 2023-02-12T08:49:41Z 2023-05-18T06:57:54Z 2023-02-12T08:49:41Z 2023-05-18T06:57:54Z 2023-02 Thesis http://hdl.handle.net/10019.1/126965 en_ZA en_ZA Stellenbosch University xii, 76 pages : illustrations. application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Reinforcement learning -- Johannesburg (South Africa) Automatic differentiation -- Johannesburg (South Africa) Stock exchanges -- Automation -- Johannesburg (South Africa) Richards, Nicole Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange |
| title | Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange |
| title_full | Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange |
| title_fullStr | Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange |
| title_full_unstemmed | Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange |
| title_short | Reinforcement learning for algorithmic day trading on the Johannesburg Stock Exchange |
| title_sort | reinforcement learning for algorithmic day trading on the johannesburg stock exchange |
| topic | Reinforcement learning -- Johannesburg (South Africa) Automatic differentiation -- Johannesburg (South Africa) Stock exchanges -- Automation -- Johannesburg (South Africa) |
| url | http://hdl.handle.net/10019.1/126965 |
| work_keys_str_mv | AT richardsnicole reinforcementlearningforalgorithmicdaytradingonthejohannesburgstockexchange |