High Frequency Trading Based on Technical Analysis and Deep Learning in the Cryptocurrency Markets

OData support
Nagy Gábor
Department of Telecommunications and Media Informatics

I reviewed some of the most cited academic literature about technical analysis-based trading and their effectiveness in different markets and time periods. I also reviewed broadly the literature about feed forward neural networks and recurrent neural networks and their application in financial time series forecasting. I showcased some of the most popular theories explaining the usefulness of technical trading in stark contrast with the Efficient Market Hypothesis. I employed various goodness of fit and performance metrics in my for high frequency trading software and ended up using the Sharpe ratio and the Welch’s t-test statistics besides the obvious candidate performance metrics: the holding period mean and total return. I improved the definition of some of the most popular visual charting patterns and implemented them in Cython and Python Numpy for fast performance. First with a reasonable selection of technical indicators and visual chart patterns I analyzed high frequency intraday cryptocurrency price series and in certain cases I found statistically significantly high test and validation period conditional mean returns even in the single best predictive pattern/indicator trading rule case. However, the Long Short-Term Memory (LSTM) based deep recurrent neural network (DRNN) combined with the GARCH model exceeded my expectations. The DRNN-based trading system I developed promises statistically significant high return of almost 300% during the mostly bear market from 201710-18 to 2019-02-01.


Please sign in to download the files of this thesis.