Modelling Bitcoin returns using ARIMA and ANN models

Gadirli, Farid (2018) Modelling Bitcoin returns using ARIMA and ANN models. [Dissertation (University of Nottingham only)]

[img] PDF - Registered users only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Download (3MB)

Abstract

Recent years witnessed significant investor interest in cryptocurrencies particularly since price of Bitcoin skyrocketed in December of 2017. This dissertation intends to evaluate the possibility of predicting the return of cryptocurrency in the case of Bitcoin using econometric and Artificial Intelligence models and identify whether they could offer investors accurate results for making a profit. Bitcoin was selected due to being most popular among all cryptocurrencies and having the highest time span of data which is suitable for making time series analysis. For this thesis, the last daily prices of Bitcoin over the period of 2010-2018 have been used. As forecasting plays an essential role in making an investment and trading decisions, the literature related to the prediction of cryptocurrency and stock markets has been reviewed and it was decided to used ARIMA and ANN models which have good capabilities for making time series predictions. In particular, the recurrent dynamic type of Artificial Neural Networks called Nonlinear Autoregressive Neural Network was used. In addition to this, the volatility of Bitcoin returns was modelled using the GARCH family models. The univariate forecasting for both in-sample and out-of-sample was performed. The accuracy of the selected models was assessed by using statistical and economic accuracy measures. Also, for getting robust results and identifying the best models different configurations of both ARIMA and ANN models were tested, in particular, the different number of lags for p and q, and various number of delays, neurons and training algorithms. The results showed that ARIMA (4,0,4) and ARIMA (7,0,15) were models with the lowest errors while ARIMA (4,0,4) had a prevalence in most cases with the absolute advantage in making ten days ahead forecasts, which demonstrated a 70% and a 60% accuracy in the predictions of changes in the direction and sign of Bitcoin returns respectively. On the other hand, the NAR models delivered even better outcomes overall yielding accuracies in the direction and sign of Bitcoin returns of up to 76% and 56%; 80% and 80%, and 100% for 34 step ahead; ten step ahead and three day ahead predictions respectively. NAR models 2-10, 7-15 and 12-10 were chosen as the best ones due to displaying robust results in various training algorithms and forecast horizons. Moreover, the comparison between ARIMA and NAR models revealed that the latter one outperforms the ARIMA models in most cases, whereas the ARIMA (4,0,4) model illustrated quite competitive results despite being surpassed by NAR 7-15 and 12-10 for ten and 34 days ahead forecasts respectively. Furthermore, according to the results of volatility modelling, a T-GARCH model was selected as the most appropriate one. Also, the significant positive leverage term and the news impact curve showed that the positive shocks have more impact on future volatility than the negative one which can be related to the safe-haven properties of Bitcoins and the lack of trust to financial institutions following the recent financial crashes. Besides, as the ARCHM term was significant, the existence of the correlation between the conditional variance and return can be concluded.

Item Type: Dissertation (University of Nottingham only)
Depositing User: GADIRLI, FARID
Date Deposited: 11 Mar 2022 16:12
Last Modified: 11 Mar 2022 16:12
URI: https://eprints.nottingham.ac.uk/id/eprint/53591

Actions (Archive Staff Only)

Edit View Edit View