A1670
Title: Edge in cryptocurrency trading: Deep learning, varied data sampling, and target labeling strategies
Authors: Przemyslaw Gradzki - University of Warsaw (Poland) [presenting]
Piotr Wojcik - University of Warsaw (Poland)
Abstract: The aim is to delve into a comprehensive examination of data sampling and target labelling techniques for the development of algorithmic trading strategies tailored to the most liquid cryptocurrencies. Within the realm of academic discourse, the prevailing data sampling method revolves around the utilization of time bars, which entail systematically spaced observations (e.g., hourly or daily intervals) derived from the ever-active 24/7 market environment. This investigation scrutinizes the trading efficacy of this conventional approach in contrast to the more information-centric strategies, such as volume/dollar bars and the custom filter. Furthermore, a comparison is provided between the most commonly employed target labelling approach, which entails forecasting the value or directional movement (upward or downward) of the next time bar, and the triple barrier method. Each of these methodologies offers its own theoretical advantages over established techniques, and this research undertakes an empirical evaluation of their superiority within the framework of crafting a trading strategy. Notably, state-of-the-art deep learning architectures are employed, including convolutional neural networks (CNN), long short-term memory networks (LSTM), and the transformer model.
