École Nationale Supérieure de
Techniques Avancées

Résumé

Financial markets have dramatically changed with digitalisation. Trades are executed within milliseconds thank to orders matching engine based on Limit Order Book (LOB).Electronic traders now exist alongside traditional traders. Their particularities is their trading frequency which is faster than traditional ones. Electronic trading is carried out by programs. Algorithms have been design to adress issues like optimal execution problem in electronic market context. Many challenges have arisen. LOB provides a large amount of data that can be processed to analyse market. These market data are used to develop trading strategies scalabe to high frequency trading. Deep Reinforcement learning (DRL) has been proven able to outperform humans in several tasks. A natural question is whether it is possible to train a DRL agent to trade? It is a challenging task taking into account how stochastic financial markets are. During this project we focus on how DRL can be used for trading purposes. To do so, we explore ways to train a model without using real market data. We tried to address the issue of the best neural network architecture for trading.