Different Data Engineering methods are compared for preparing cryptocurrency market data for a supervised classification model to predict and act on profitable trading decisions. Market data consisting of records of every single transaction is hereby utilized. To manage this data and be able to run flexible computations on it, a special stream processing framework is implemented. As a result, conclusive differences between methods are found, as well as a method that proved to be profitable when run on previously unseen future data and after accounting for realistic trading costs.
Python, Pandas, Numpy, Cryptocurrency, Trading, Data Engineering, Machine Learning
The goal was to take a specific approach of constructing a model for making trading decisions, that is proposed in the project's primary source, and find out whether it has any merit on the cryptocurrency markets. In addition we wanted to compare different methods of processing data and find out which ones are the most useful in this context.
Cryptocurrency markets can be very attractive for algorithmic trading especially working with the market data itself. In a previous (IP5) project more traditional statistical methods were explored on the same market data. Here we focus on a different approach which uses a supervised classifier model at its core.
The tested approach succeeded in creating a profitable trading strategy, which did not succeed in significantly beating the corresponding market index however. Working with tick data can pose some computational challenges for projects like this. It has been found that features resulting from multiplication or division or basic values are most useful. The sampling techniques under test were also most useful for capturing mean reversion opportunities in the market.
Duration: February 2021 - August 2021
Time budget: 360 hours
Team size: 1 person
Institute for Data Science i4DS
FHNW - University of Applied Sciences Northwestern Switzerland
FHNW Bahnhofstrasse 6
5210 Windisch
Prof. Dr. Manfred Vogel (manfred.vogel@fhnw.ch)
Michel Plüss (michel.pluess@fhnw.ch)
Fedor Perejoguine
Prof. Dr. Manfred Vogel (manfred.vogel@fhnw.ch)
Michel Plüss (michel.pluess@fhnw.ch)
