Agent-based learning for pattern matching in high-frequency trade data
Previousresearchofsequentialinvestmentstrategiesforportfolioselectionhaveshownthatthereare strategies that exist that can beat the best stock in the market. In this dissertation, an algorithm is presented that uses a nearest neighbour approach similar to the one used by Gy¨orﬁ et al [20, 21, 22]. Theapproachishoweverextendedtoincludezero-costportfoliosandusesaquadraticapproximation, instead of an optimisation step, to determine how capital should be allocated in the portfolio based on the neighbours that have been found. A portfolio that results in an increase in the investor’s capitalandcomparesfavourablytocertainbenchmarks,suchasthebeststock,indicatesthatthereare patternsinthetimeseriesdata. Otherfeaturesofthealgorithmpresentedistoallowforthedatatobe clustered by a selection of stocks or partitioned based on time. The algorithm is tested on synthetic datasetsthatdepictdifferentmarkettypesandisshowntoaccuratelydeterminetrendsinthedata. The algorithm is then tested on real data from the New York Stock Exchange (NYSE) and data from the JohannesburgStockExchange(JSE).Theresultsofthealgorithmfromtherealdatasetsarecompared to implemented versions of past strategies from the literature and compares favourably.
A dissertation submitted in fulﬁlment of the requirements for the degree of Master of Science to the Faculty of Science, University of the Witwatersrand, 2017
Loonat, Fayyaaz (2017) Agent-based learning for pattern matching in high-frequency trade data, University of the Witwatersrand, Johannesburg, http://hdl.handle.net/10539/23537