This thesis addresses the prediction problems associated with noncausal processes in cryptocurrency markets. Chapter one provides background on Bitcoin and cryptocurrencies in general. It begins by introducing four major cryptocurrencies. Then recent developments in economic research on Bitcoin are discussed. \n \nChapter two introduces a noncausal autoregressive process with Cauchy errors in application to the exchange rates of the Bitcoin electronic currency against the US Dollar. The dynamics of the daily Bitcoin/USD exchange rate series display episodes of local trends, which are modelled and interpreted as speculative bubbles. The structure of the Bitcoin market is described to give context for the presence of multiple bubbles in the exchange rate. The bubbles may result from the speculative component in the on-line trading. The Bitcoin/USD exchange rates are modelled and predicted. The mixed causal-noncausal autoregressive model is shown to better fit the data than the traditional purely causal model. A forecasting exercise using the noncausal model is then presented. \n \nChapter three examines the performance of nonlinear forecasts of noncausal processes from closed-form functional predictive density estimators. To examine the performance, time series are simulated with different conditional means and non-Gaussian distributions. The processes considered have the mixed causal-noncausal MAR(1,1)dynamics and both finite and infinite variance. The forecasts are assessed based on the forecast error behaviour and the goodness of fit of the estimated predictive density. The persistence in the noncausal component directly relates to the magnitude of the bubble effects in the time series and is found to have a meaningful impact on how forecastable the process is. To better predict bubbles the joint density of the forecast at horizon two is shown to be an effective graphical method to detect the outset of a bubble.
Abstract In recent years, the scrutiny of bitcoin and other cryptocurrencies as legal and regulated components of financial systems has been increasing. Bitcoin is currently one of the largest cryptocurrencies in terms of capital market share. Therefore, this study proposes that sentiment analysis can be used as a computational tool to predict the prices of bitcoin and other cryptocurrencies for different time intervals. A key characteristic of the cryptocurrency market is that the fluctuation of currency prices depends on people's perceptions and opinions, not institutional money regulation. Therefore, analysing the relationship between social media and web search is crucial for cryptocurrency price prediction. This study uses Twitter and Google Trends to forecast the short‐term prices of the primary cryptocurrencies, as these social media platforms are used to influence purchasing decisions. The study adopts and interpolates a unique multimodel approach to analyse the impact of social media on cryptocurrency prices. Our results prove that people's psychological and behavioural attitudes have a significant impact on the highly speculative cryptocurrency prices.
This paper sets forth a framework for deep reinforcement learning as applied to market making (DRLMM) for cryptocurrencies. Two advanced policy gradient-based algorithms were selected as agents to interact with an environment that represents the observation space through limit order book data, and order flow arrival statistics. Within the experiment, a forward-feed neural network is used as the function approximator and two reward functions are compared. The performance of each combination of agent and reward function is evaluated by daily and average trade returns. Using this DRLMM framework, this paper demonstrates the effectiveness of deep reinforcement learning in solving stochastic inventory control challenges market makers face.
Zehra Kadiroğlu, Beyza Nur Akılotu, Abdulkadir Şengür
Bitcoin is a digital currency that uses cryptographic rules to regulation and generation units of currencies. In 2009, it was introduced by a person (or a group of people) using the name Satoshi Nakamoto before putting on the market as open source software. Bitcoin is a new payment system and investment tool for the purchase, storage and use of money as digitally. However, when people want to invest in bitcoin, the biggest factor they have to consider is how the price will change in the future. Recent developments in the field of machine learning have shown that some algorithms can provide appropriate solutions to predict future prices of crypto currencies. In this study, a detailed literature review is made about the studies using machine learning methods to estimate the future price of bitcoin. Basic and current information about bitcoin was included in this study. In addition, the theory of machine learning methods used is explained. The findings of the literature review show that machine learning methods can be successful in predicting bitcoin price.
Although, the growth in the cryptocurrency market slowed down after the meteoric rise in late 2017, the market is still enjoying steady capital inflow. This has made the study of market dynamics between the cryptocurrencies and equity market indispensable. In this paper, the study of the volatility spillovers and correlation between the two has been undertaken by considering five Asian stock indices and four cryptocurrencies ranging from November 2014 to December 2018, to cover three phenomena- Leverage effect, Volatility spillovers and Time varying correlation using EGARCH, Diagonal BEKK and DCC tests respectively. Firstly, the EGARCH test reveals the absence of leverage effect in the returns of cryptocurrenices. Secondly, the multivariate GARCH test shows, out of all the cryptocurrencies taken, the past innovations in Bitcoin affect the future volatility of the equity market returns the most. Lastly, the DCC model reveals evidence of time varying correlation between the markets and Bitcoin. Keywords: Cryptocurrencies; Asian equity market; Volatility spillovers; Dynamic conditional correlation JEL Classifications: G12, G14, G17, C15, C32 DOI: https://doi.org/10.32479/ijefi.8624
Van Minh Hao, Nguyen Huynh Huy, Bo Dao, Thanh-Tan Mai · 5 authors
Predicting cryptocurrency price movements is a challenging task due to the highly stochastic nature of the market. This paper exploits features from social media, combining with the historical price to build an accurate model for predicting trending of the Bitcoin, the most popular cryptocurrency these days. The novelty of this work is introducing a new feature called 'interaction' which helps improve the model's performance significantly. Our approach shows a very promising result, which outperforms other recent works by a large margin.
Asli Demir, Beyza Nur Akılotu, Zehra Kadiroğlu, Abdulkadir Şengür
Bitcoin is a new currency that is recognized as a creative payment network. The operating system functions independently of a central structure or bank. Bitcoin is managed by the developed network. Bitcoin's open source code structure allows it to be uncontrolled and uncontrollable by an individual. The use of bitcoin requires virtual wallet. Verification of all payments is secured using public key encryption. Fluctuate in Bitcoin prices at a high rate, making it difficult to predict. For this reason, new cryptology methods related to bitcoin price estimation and security are tried in the literature. In our study, bitcoin price estimation was made by using machine learning methods using KAGGLE Bitcoin Dataset 2010-2019 data set. The methods used are long-short term memory networks, support vector machines, artificial neural networks, Naive Bayes, decision trees and the nearest neighbor algorithm. Obtained accuracy rates are 97.2%, 91.8%, 86.6%, 85%, 81.2% respectively.
The research on forecasting Ethereum STORJ token has not been widely studied compared to forecasting Bitcoin. The objective of this paper is threefold: apply existing Bitcoin price forecasting models to the Ethereum STORJ token price; evaluate the dynamics of the model predictive utility across three time horizons (h=5 days, h=20 days and h=50 days); and determine if Ethereum STORJ token clustering coefficients impact the effectiveness of the forecasting model. We choose Bitcoin forecasting models of: ARIMA, ARMA-GARCH, VAR, alpha-Sutte Indicator and NNAR. Model effectiveness is analyzed using RMSE, MAE and MAPE. We find that VAR outperforms all models in the short and mid-term horizons (h=5 and h=20 days) and NNAR outperforms all models in the long-term horizon (h=50 days). Non-linearity, the intrinsic value that Neural Network has, may strongly effect the forecast accuracy result. When adding the clustering coefficient to ARIMA, we find that the variable is significant but only marginally improves the forecasting of the Ethereum STORJ token. The VAR model, which includes the clustering coefficient is shown to better forecast Ethereum STORJ prices in the short and mid-term forecasting horizons. The NNAR model is a better model for the long-term forecasting horizon.
In the past years, cryptocurrencies have received a lot of attention in popular media. Having attracted significant speculation, prices have soared in 2017, fell in 2018 and are generally known to be very volatile. However, some of the price changes have been due to organized manipulation. Traditionally known in the world of penny stocks and made illegal in most countries, pump and dump schemes are frequent in cryptocurrencies, and mostly unregulated. In this paper, we perform quantification and detection of pump and dump schemes that are coordinated through Telegram chats and executed on Binance - one of the most popular cryptocurrency exchanges. We detail how pumps are organized on Telegram, and quantify the properties of 149 confirmed events with respect to market capitalization, trading volume, price impact and profitability. Based on this ground truth, and regular trading intervals obtained from twitter timestamps, we optimize a binary classifier in order to be able to detect additional suspicious trading activity. Our results indicate that pump and dump schemes occur frequently in cryptocurrencies with market capitalizations below $50 million, that scheme operators often organize their actions across multiple channels, that such activity tends to lead to inflated prices over longer time periods and machine learning can help to identify activity that is similar to known pump and dump schemes.
In this paper, we apply Long Short-Term Memory (LSTM) neural networks to model the token price time series data which incorporate the local topological measures of investor transaction network and market summaries. In addition, we propose a novel LSTM-based model using the leave-one-out cross-validation technique and utilizing the network motif analysis. The numerical results show that the proposed LSTM-based model could significantly improve the performance of the prediction of token price compared to the benchmark LSTM-based models and deep portfolios regardless of the training and testing data split ratio. Some concluding remarks and future research directions are provided.
Cryptocurrency prices have changed very dynamically in the market. Buyers and sellers can trade cryptocurrency on the market without time limits compared to traditional trading markets such as exchange currency markets and stock markets. Also, since it is a cryptocurrency created by an anonymous inventor, cryptocurrency price predictions are not determined by the company’s financial performance. Rather, the cryptocurrency price is related to how many investors participate in the market. In this sense, the prediction of cryptocurrency prices is very difficult and related to market participants. In this study, to better understand cryptocurrency pricing factors, we explore cryptocurrency price forecasts and use deep learning to improve forecasts. Specifically, we collected variables related to the investor’s decision. Use linear regression to select features to find important variables in cryptocurrency price prediction. In regression analysis, three models were created to identify the model that represents the best performance of cryptocurrency price prediction. Model 1 uses all variables without function selection. Model 2 uses only variables that are important in feature selection. Model 3 uses only variables that are not important for feature selection. Our test results show that Model 2 outperforms other models. We conclude that using the appropriate variables can improve cryptocurrency price predictions.
This study attempts to analyze patterns in cryptocurrency markets using a special type of deep neural networks, namely a convolutional autoencoder. The method extracts the dominant features of market behavior and classifies the 40 studied cryptocurrencies into several classes for twelve 6-month periods starting from 15th May 2013. Transitions from one class to another with time are related to the maturement of cryptocurrencies. In speculative cryptocurrency markets, these findings have potential implications for investment and trading strategies.
We implement hidden Markov models (HMMs) and hidden semi-Markov models (HSMMs) on Bitcoin/US dollar (BTC/USD) with the aim of market phase detection. We make analogous comparisons to Standard and Poor’s 500 (S and P 500), a benchmark traditional stock index and a protagonist of several studies in finance. Popular labels given to market phases are “bull”, “bear”, “correction”, and “rally”. In the first part, we fit HMMs and HSMMs and look at the evolution of hidden state parameters and state persistence parameters over time to ensure that states are correctly classified in terms of market phase labels. We conclude that our modelling approaches yield positive results in both BTC/USD and the S and P 500, and both are best modelled via four-state HSMMs. However, the two assets show different regime volatility and persistence patterns—BTC/USD has volatile bull and bear states and generally weak state persistence, while the S and P 500 shows lower volatility on the bull states and stronger state persistence. In the second part, we put our models to the test of detecting different market phases by devising investment strategies that aim to be more profitable on unseen data in comparison to a buy-and-hold approach. In both cases, for select investment strategies, four-state HSMMs are also the most profitable and significantly outperform the buy-and-hold strategy.
In this study, the impacts of Bitcoin on Japan, China, Turkey and USA stock indexes were investigated. Nikkei225, SSE380, BIST100 and S&P500 were selected as the stock market indexes. The weekly data including dates between January 03, 2016 and December 16, 2018 were analyzed using EViews program and firstly the time-dependent, non-stationary data set was stabilized. Then, the stabilized data was analyzed with VAR(3) model according to Akaike information criteria. According to the Johansen cointegration test, 2nd model was selected as the most appropriate model for the study. The variables were rearranged with error correction model and then Granger causality analysis was performed. As a result of these analysis, it was determined that Bitcoin only affected BIST100 and that there were two-way causality relation between them. In addition, a one-way causality from Nikkei225 to SSE380, from SSE380 to Bitcoin, from S&P500 to Nikkei225 and from Nikkei225 to Bitcoin were observed.
Η ανά χείρας διπλωματική εργασία επικεντρώνεται στη πρόβλεψη του price return του Bitcoin μέσω αλγορίθμων Μηχανικής μάθησης. Στόχος αυτής της εργασίας ήταν η ανάπτυξη του αντίστοιχου κώδικα, χρησιμοποιώντας πραγματικά δεδομένα, που αφορούν το κρυπτονόμισμα Bitcoin, η πρόβλεψη της εξέλιξης του price return και τελικά η αξιολόγηση του κάθε αλγορίθμου, μέσω της ακρίβειας της πρόβλεψης. Αρχικά, πραγματοποιήθηκε μία βιβλιογραφική ανασκόπηση βασισμένη στο Bitcoin, στη Μηχανική μάθηση και στο συνδυασμό αυτών των δύο. Στη συνέχεια, το επόμενο βήμα αφορούσε τη μεθοδολογία και τα δεδομένα. Ήταν απαραίτητη η συλλογή, η επεξεργασία και ο διαχωρισμός των δεδομένων σε δεδομένα εκπαίδευσης και ελέγχου. Επίσης, σε αυτό το σημείο της έρευνας, προέκυψαν τα πρώτα περιγραφικά στατιστικά των δύο συνόλων δεδομένων. Λαμβάνοντας υπόψιν τη βιβλιογραφία, οι επιλεγμένοι αλγόριθμοι Μηχανικής μάθησης ήταν ο Random Forest και ο Naïve Bayes, ενώ ως οικονομική μέθοδος χρησιμοποιήθηκε η Λογιστική Παλινδρόμηση. Τόσο τα δεδομένα εκπαίδευσης, όσο και τα δεδομένα ελέγχου, εφαρμόστηκαν σε κάθε αλγόριθμο. Βασιζόμενοι στην ακρίβεια της πρόβλεψης, τα κύρια συμπεράσματα ήταν, ότι ο αλγόριθμος Random Forest είχε τη βέλτιστη απόδοση (ακρίβεια πρόβλεψης: 65%), ο αλγόριθμος Naïve Bayes είχε μία αρκετά καλή απόδοση (ακρίβεια πρόβλεψης 60%), ενώ η Λογιστική Παλινδρόμηση έφερε ικανοποιητικά αποτελέσματα (ακρίβεια πρόβλεψης 60%). Τέλος, οι υπόλοιπες παρατηρήσεις που προέκυψαν κατά τη διάρκεια της έρευνας, αναφέρονται στο τελευταίο κεφάλαιο αυτής της διπλωματικής εργασίας.
Purpose Crypto-currencies, decentralized electronic currencies systems, denote a radical change in financial exchange and economy environment. Consequently, it would be attractive for designers and policy-makers in this area to make out what social media users think about them on Twitter. The purpose of this study is to investigate the social opinions about different kinds of crypto-currencies and tune the best-customized classification technique to categorize the tweets based on sentiments. Design/methodology/approach This paper utilized a lexicon-based approach for analyzing the reviews on a wide range of crypto-currencies over Twitter data to measure positive, negative or neutral sentiments; in addition, the end result of sentiments played a training role to train a supervised technique, which can predict the sentiment loading of tweets about the main crypto-currencies. Findings The findings further prove that more than 50 per cent of people have positive beliefs about crypto-currencies. Furthermore, this paper confirms that marketers can predict the sentiment of tweets about these crypto-currencies with high accuracy if they use appropriate classification techniques like support vector machine (SVM). Practical implications Considering the growing interest in crypto-currencies (Bitcoin, Cardano, Ethereum, Litcoin and Ripple), the findings of this paper have a remarkable value for enterprises in the financial area to obtain the promised benefits of social media analysis at work. In addition, this paper helps crypto-currencies vendors analyze public opinion in social media platforms. In this sense, the current paper strengthens our understanding of what happens in social media for crypto-currencies. Originality/value For managers and decision-makers, this paper suggests that the news and campaign for their crypto in Twitter would affect people’s perspectives in a good manner. Because of this fact, the firms, investing in these crypto-currencies, could apply the social media as a magnifier for their promotional activities. The findings steer the market managers to see social media as a predictor tool, which can analyze the market through understanding the opinions of users of Twitter.
Forecasting exchange rates is difficult because financial time-series data is too complicated to analyze. In traditional financial studies, economic models and statistic approaches were widely used for predicting exchange rates. Recently, machine learning and deep learning techniques have played increasingly important roles in financial technology studies. This study adopts a deep learning technique called relation networks (RNs) to predict the exchange rates of fiat currencies and cryptocurrencies. To discover the relationship among different currencies, the concept of visual question answering (VQA) is applied in RNs. We also propose a specially designed architecture for the feature extraction stage to consider both spatial and temporal relationships simultaneously. The experimental results show that the proposed approach can achieve higher prediction performance for cryptocurrencies with approximately 65% accuracy rate. We aim to improve traditional approaches and construct a model using the concept of VQA based on RNs to optimize the prediction performance between fiat currencies and cryptocurrencies.
Cryptocurrency is decentralized and electronic alternative of currency. Bitcoin is the best example of this type of currency, and after bitcoin hundreds of such currencies were launched in market. Cryptocurrency is more frequently used, as it is theft proof, accessible anywhere & anytime. By using cryptocurrency the settlement of money is instant. Portfolio management is a technique through which a person decides how to allocate the resources. It helps in making better decisions and in minimizing risks of loss. Cryptocurrency and portfolio management are mostly used in financial sectors like banking and stocks. This paper reviews about the use of cryptocurrency and portfolio management with its associated challenges.
This research presents a novel mechanism of digital asset trading system on blockchain called JSP-DATS. The JSP-DATS includes (1) a novel mechanism of trading, and (2) a novel mechanism of blockchain which will be the infrastructure of the system. The proposed novel mechanism of blockchain uses the "Random-Checker Proof of Stake" consensus model which can decrease transaction time. The blockchain of the JSP-DATS has been designed to multiple layers. This design is easy to develop, and can be used for further research. The internal mechanism of the proposed system including steps of encoding/decoding, key management, and the storage of encrypted digital assets on the blockchain has will be discussed in this paper. In addition, we have implemented the designed model using Microsoft Visual C++ and encryption libraries from the MSDN web-site to create a software prototype. The prototype is used to study and measure the speed of the proposed scheme. The results show that transaction time of the proposed scheme is lower than that in BitCoin and Ethereum blockchain. With the proposed scheme, the seller (digital asset owners) can see transactions of the trading system transparently, and they can receive their percentage share immediately. In addition, we expect that buyers will indirectly benefit from purchasing digital assets at a lower price.
Forecasting time series data is an important subject in economics, business, and finance. Traditionally, there are several techniques such as univariate Autoregressive (AR), univariate Moving Average (MA), Simple Exponential Smoothing (SES), and more notably Autoregressive Integrated Moving Average (ARIMA) with their many variations that can effectively forecast. However, with the recent advancement in the computational capacity of computers and more importantly developing more advanced machine learning algorithms and approaches such as deep learning, new algorithms have been developed to forecast time series data. This article compares different methodologies such as ARIMA, Random Forest (RF), Support Vector Machine (SVM), Long Short-Term Memory (LSTM) and WaveNets for estimating the future price of Bitcoin.