Experimental study of online learning algorithms for market making
Πειραματική μελέτη άμεσων αλγορίθμων μάθησης για δημιουργία αγοράς
View/ Open
Keywords
Άμεση μάθηση ; Online learning ; Market making ; High frequency trading ; Limit-order-bookAbstract
The key role of a Market Maker in stock exchange is provision of liquidity via issuing of
buy and sell orders for a security; at the same time he is willing to profit from the changes
of the stock price. Nowadays, in modern stock exchanges, most of the trading occurs in
computers and thus the environment has become more demanding in terms of the speed
and complexity of transactions. For that reason, the Market Maker must use high tech
software and efficient Algorithms to adapt to these changes. In this thesis, we developed a
limit order book simulation with its basic functions using Python. We further developed a
"random order generator" which feeds the Limit Order Book with pseudorandom buy and
sell orders. In order to create the generator, we analysed the Algorithmic Trading Challenge
dataset from Kaggle which consists of real stock market data. Finally, we designed an Agent
that uses a class of Online learning algorithms, the ε n Greedy, the UCB1 and the EXP3.
These Algorithms are using a class of window based strategies, inspired from the literature.
In our experiments we compared these algorithms with each other, focusing on the Agent’s
profit and their regret in terms of the overall best window strategy.