A1892
Title: Neural architecture search for bitcoin market prediction
Authors: Georg Velev - Humboldt University Berlin (Germany)
Stefan Lessmann - Humboldt-University of Berlin (Germany) [presenting]
Abstract: The design of algorithms for the prediction of the Bitcoin market and for trading with Bitcoins has received a lot of attention both from researchers in academia and in the financial industry because Bitcoin represents the cryptocurrency with the highest market value among all crypto assets. Numerous studies dealing with cryptocurrency forecasting have applied long short-term memory recurrent neural networks due to their ability to learn temporal dependencies from time series data. Recently, transformer-based networks have shown promising results on various tasks including time series forecasting. Neural architecture search (NAS) facilitates the automated design of neural-based architectures, which are tailored to a specific task. NAS models have been reported in the literature to outperform hand-crafted architectures on machine-learning tasks such as image and text classification. Therefore, NAS is applied using a policy gradient for algorithmic trading with Bitcoins. On the micro level, the focus is on the search for novel recurrent cells and Transformer-based cells. On the macro level, the remaining hyperparameters that were set to fixed values during the micro search are optimized. The results achieved are benchmarked by the best-performing neural architectures with long short-term memory neural networks.
