Open Access

Hans-Juergen Griesbacher

• Cryptocurrency brokers offer price quotations to investors for a set period. The investor can decide whether to accept the offer while the quote remains open. The broker bears a risk of the price changing while the quote is open. To mitigate this risk, brokers add an individual risk premium to the price. This thesis investigates how to improve the risk premium’s precision by using an artificial neural network (ANN) instead of the classical general autoregressive conditional heteroscedasticity (GARCH) model. The thesis opens by describing financial risks, conventional traditional approaches, and the theory of machine learning, in particular neural networks. Subsequently, short-term volatility in the market is explored in terms of volatility prediction with the classical GARCH and ANN approach with a focus on an investor’s purchase process. With the predicted volatility, the risk premium is calculated with the expected short-fall method. This thesis explores BTC-USD, ETH-USD, and ETH-BTC. The risk premium is calculated for a 10-minute quotation period with data from the cryptocurrency exchange Kraken from October 2019 to April 2020. The analysis shows that the ANN approach delivers a more precise volatility prediction and risk premium calculation due to a lower mean square error deviation compared to the GARCH model. However, the ANN has a low explanation of power in the test period. Different volatility cluster phenomena in the short-term data compared to the classical daily basis are also identified.