Predictive Models for Blockchain, Cryptocurrency, and Derivatives Market

Abstract

This dissertation aims to conduct the empirical analysis for the financial derivative and cryptocurrency market and to develop analytical techniques based on machine learning models suitable for prediction and estimation of each field. In the financial derivative market, a Markov chain Monte Carlo (MCMC) methods employ the candidate probability distribution nearest to the target probability distribution to acquire sample distributed from the posterior density. Choice of the candidate probability distribution affects the practical convergence speed of the MCMC methodology and the fitness of the sample. In this dissertation, we propose a MCMC framework possible to samples from the candidate distribution nearest to the target probability density without the specification of the candidate distribution. We confirm that the jump diffusion models and Bayesian neural networks have the best performance in estimating and predicting given the data of the recent day for the model estimation given S&P index options in 2012. Especially, the jump diffusion model has a very high performance in terms of domain adaptation between the American option and the European option. This difference is reflected in the fact that the jump diffusion model is based on the common asset of the American option and the European option. Based on this empirical precedent study, we proposed a machine learning model called generative Bayesian neural network (GBNN) to overcome the disadvantages of the machine learning model. GBNN maximizes posterior probability through the GBNN obtains prior information from the GBNN data learned up to the previous day, and learns likelihood probability from actual trading data of learning day. We identify that the GBNN model outperform other benchmark models in terms of model prediction. Bitcoin is a successful cryptocurrency, and it has been extensively studied in fields of economics and computer science. In this dissertation, we analyze the time series of Bitcoin price with a BNN using Blockchain information in addition to macroeconomic variables. We conduct the empirical study that compares the Bayesian neural network with other linear and non-linear benchmark models on modeling and predicting the Bitcoin process. Our empirical studies show that BNN performs well in predicting Bitcoin price time series and explaining the high volatility of the Bitcoin price in Aug. 2017. In addition, we suggested the enhanced GRU model for correlation analysis between cryptocurrency markets. Assuming that the gate value obtained from the GRU model is the parameter of the VAR model, it makes possible to visualize the correlation between various alternative currencies in the cryptocurrency market. As a result, it is confirmed that there is a very significant correlation between the currencies separated from the existing currencies and the existing currencies.