Living in a data-intensive environment is a natural consequence to the continuous innovations and technological advancements, that created countless opportunities for addressing domain-specific challenges following the Data Science approach. The main objective of this thesis is to present applied Data Science approaches in FinTech, focusing on proposing innovative descriptive and predictive models for studying and exploring Bitcoin Price Dynamics and Bitcoin Price Prediction. With reference to the research area of Bitcoin Price Dynamics, two models are proposed. The first model is a Network Vector Autoregressive model that explains the dynamics of Bitcoin prices, based on a correlation network Vector Autoregressive process that models interconnections between Bitcoin prices from different exchange markets and classical assets prices. The empirical findings show that Bitcoin prices from different markets are highly interrelated, as in an efficiently integrated market, with prices from larger and/or more connected exchange markets driving other prices. The results confirm that Bitcoin prices are unrelated with classical market prices, thus, supporting the diversification benefit property of Bitcoin. The proposed model can predict Bitcoin prices with an error rate of about 11% of the average price. The second proposed model is a Hidden Markov Model that explains the observed time dynamics of Bitcoin prices from different exchange markets, by means of the latent time dynamics of a predefined number of hidden states, to model regime switches between different price vectors, going from "bear'' to "stable'' and "bear'' times. Structured with three hidden states and a diagonal variance-covariance matrix, the model proves that the first hidden state is concentrated in the initial time period where Bitcoin was relatively new and its prices were barely increasing, the second hidden state is mostly concentrated in a period where Bitcoin prices were steadily increasing, while the third hidden state is mostly concentrated in the last period where Bitcoin prices witnessed a high rate of volatility. Moreover, the model shows a good predictive performance when implemented on an out of sample dataset, compared to the same model structured with a full variance-covariance matrix. The third and final proposed model, falls within the area of Bitcoin Price Prediction. A Hybrid Hidden Markov Model and Genetic Algorithm Optimized Long Short Term Memory Network is proposed, aiming at predicting Bitcoin prices accurately, by introducing new features that are not usually considered in the literature. Moreover, to compare the performance of the proposed model to other models, a more traditional ARIMA model has been implemented, as well as a conventional Genetic Algorithm-optimized Long Short Term Memory Network. With a mean squared error of 33.888, a root mean squared error of 5.821 and a mean absolute error of 2.510, the proposed model achieves the lowest errors among all the implemented models, which proves its effectiveness in predicting Bitcoin prices.
Applied Data Science Approaches in FinTech: Innovative Models for Bitcoin Price Dynamics
ABU HASHISH, IMAN HISHAM JAMIL
2020
Abstract
Living in a data-intensive environment is a natural consequence to the continuous innovations and technological advancements, that created countless opportunities for addressing domain-specific challenges following the Data Science approach. The main objective of this thesis is to present applied Data Science approaches in FinTech, focusing on proposing innovative descriptive and predictive models for studying and exploring Bitcoin Price Dynamics and Bitcoin Price Prediction. With reference to the research area of Bitcoin Price Dynamics, two models are proposed. The first model is a Network Vector Autoregressive model that explains the dynamics of Bitcoin prices, based on a correlation network Vector Autoregressive process that models interconnections between Bitcoin prices from different exchange markets and classical assets prices. The empirical findings show that Bitcoin prices from different markets are highly interrelated, as in an efficiently integrated market, with prices from larger and/or more connected exchange markets driving other prices. The results confirm that Bitcoin prices are unrelated with classical market prices, thus, supporting the diversification benefit property of Bitcoin. The proposed model can predict Bitcoin prices with an error rate of about 11% of the average price. The second proposed model is a Hidden Markov Model that explains the observed time dynamics of Bitcoin prices from different exchange markets, by means of the latent time dynamics of a predefined number of hidden states, to model regime switches between different price vectors, going from "bear'' to "stable'' and "bear'' times. Structured with three hidden states and a diagonal variance-covariance matrix, the model proves that the first hidden state is concentrated in the initial time period where Bitcoin was relatively new and its prices were barely increasing, the second hidden state is mostly concentrated in a period where Bitcoin prices were steadily increasing, while the third hidden state is mostly concentrated in the last period where Bitcoin prices witnessed a high rate of volatility. Moreover, the model shows a good predictive performance when implemented on an out of sample dataset, compared to the same model structured with a full variance-covariance matrix. The third and final proposed model, falls within the area of Bitcoin Price Prediction. A Hybrid Hidden Markov Model and Genetic Algorithm Optimized Long Short Term Memory Network is proposed, aiming at predicting Bitcoin prices accurately, by introducing new features that are not usually considered in the literature. Moreover, to compare the performance of the proposed model to other models, a more traditional ARIMA model has been implemented, as well as a conventional Genetic Algorithm-optimized Long Short Term Memory Network. With a mean squared error of 33.888, a root mean squared error of 5.821 and a mean absolute error of 2.510, the proposed model achieves the lowest errors among all the implemented models, which proves its effectiveness in predicting Bitcoin prices.File | Dimensione | Formato | |
---|---|---|---|
AbuHashish - Data Science Approaches in FinTech.pdf
accesso aperto
Dimensione
2.38 MB
Formato
Adobe PDF
|
2.38 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/84844
URN:NBN:IT:UNIPV-84844