Nov. 22, 2023, 11:43 p.m.

In Przemka Kanarek competition for the best master's thesis in computer science, two equivalent second prizes were awarded. One of them went to Michał Dąbrowski for the thesis Statistical Arbitrage in Polish Equities Market using deep learning techniques. Michał's supervisor was Marek Adamczyk.

Michał describes his work as follows.

The thesis focuses on presenting an analytical stock trading strategy. This strategy is a variation of the concept of pair trading, which involves taking opposite positions in highly correlated assets. For example, imagine a package consisting of 1 PLN worth of shares bought from Tauron and 1 PLN worth of shares sold from Enea. It can be expected that the value of such a portfolio should oscillate around 0 due to the similar nature of both companies and their competitiveness in the energy market. Any significant deviations from the "equilibrium" state provide an opportunity for profit in trading such a constructed portfolio. The profit is then independent of the current trend in the stock prices; it only exploits quickly correcting differences in their relative values. Assuming that one component is described by the other, pair trading can be seen as conducting transactions on the residuals of a linear model.
In the presented variation, the second component of the pair is replaced by a portfolio of shares from multiple companies aiming for the best replication of the first component. This version does not require the search for correlated assets; additionally, it allows for a more precise extraction of the residuals process in terms of its oscillation around the value of 0. The replication process, i.e., the construction of a linear model describing the first component, is presented in three independent ways: using existing stock market indices; using artificially defined indices based on dimensionality reduction techniques, and with the help of recurrent neural networks. The first two methods have already been used in the literature, while the third constitutes the author's contribution to the work.
The above-mentioned pair trading variation with an additional division into 3 approaches was tested on the Polish stock market from 2017 to 2020. The results confirmed the effectiveness of already known approaches based on existing and artificially created indices on a much smaller stock market than the previously considered American market. An additional achievement of the work turned out to be repeatable, positive results obtained from the author's technique based on recurrent neural networks. This can be considered as a promising prognosis for the further development of this approach in future works.
Michał DąbrowskiMarek Adamczyk