Full length article
A complete empirical ensemble mode decomposition and support vector machine-based approach to predict Bitcoin prices

https://doi.org/10.1016/j.jbef.2020.100335Get rights and content

Abstract

Bitcoin as an asset class has received phenomenal investor attention and is considered to have similar characteristics like gold. This study aims to analyze the price behavior of bitcoin and apply machine learning algorithm for its prediction. Understanding the nature of Bitcoin price series is a multi-scale problem, and it can be best examined by analyzing its compositional characteristics. This study uses complete empirical ensemble mode decomposition (CEEMD) to analyze the nature of Bitcoin price series. Daily Bitcoin prices from 2012 to 2018 are used to perform CEEMD to identify the short term, medium term, and long-term trend in the Bitcoin price series. The study uses support vector machine (SVM) learning algorithm to find whether it can predict Bitcoin prices and finds that SVM predicts five steps ahead Bitcoin prices for the short term, medium term, long term, and overall Bitcoin price level.

Introduction

Blockchain technology and innovation in financial systems have led to the birth of an alternative mode of currency called cryptocurrencies. Cryptocurrencies have emerged as a digital currency, which offer a decentralized and transparent financial system, as opposed to the existing centralized traditional financial system. They provide an easy, fast, and seamless peer-to-peer transfer (Corbet et al., 2018a). One of the most popular cryptocurrencies is Bitcoin (Briere et al., 2015), which has emerged as a form of alternative investment for portfolio diversification as well. The frequent high fluctuations in Bitcoin prices have made it an extremely volatile investment. Price movements within an interval of ±8000% have been observed in Bitcoin (Ciaian et al., 2016) and scholars have started emphasizing the need to examine its price behavior (Dyhrberg, 2016, Katsiampa, 2017).

Examining price fluctuations in traditional asset classes like equities, bonds, and commodities such as gold have garnered huge attention from scholars in the past (Balcilar et al., 2017). With the increasing popularity and adoption of Bitcoin as an asset class, it becomes imperative to understand the price fluctuations in Bitcoin. The mechanisms behind Bitcoin price fluctuations are very complicated and have been found to influence significant herding and contagion effect on other cryptocurrencies (da Gama Silva et al., 2019, Stavroyiannis and Babalos, 2019). Despite receiving tremendous research attention, Bitcoins are still surrounded by a lot of skepticism, as they are not backed by any state neither do their fundamental value reside on any underlying asset (Bariviera et al., 2017). Moreover, the price of Bitcoin today has increased by 5000% since its initial price in 2011 (Urquhart, 2016). Exploring the statistical properties of Bitcoin prices within the gamut of market efficiency (Fama, 1970), which is also known as the efficient market hypothesis (EMH), has become a debatable issue among scholars.

Bitcoin has become very speculative and sensitive to any fake or even real news dealing with economic or political events (Ramirez et al., 2018). Scholarly works in the field of finance, economics, and machine learning are shedding light on understanding price behavior of Bitcoins. However, compared to the traditional asset classes, the understanding of the potential of Bitcoins as an asset class has been relatively less explored. Hence, it becomes imperative to explore the nature of Bitcoin price fluctuations for reducing the downside investment risk in Bitcoins along with developing a better understanding of its asset returns potential. This study is a contribution in this direction and holds implications to develop trading strategies for Bitcoins.

Market efficiency implies that all the available information is fully reflected in the market prices, and hence, markets are subjected to a random walk model. While EMH makes room for short-term memory persistence in the market, it rejects the presence of long-term memory persistence in the financial time series of tradable assets. If there is long-term memory, it will imply that traders can develop a profitable trading strategy, as markets do not follow a random walk model (Briere et al., 2015). By exploring the linear and nonlinear stochastic movements in Bitcoins, scholars have attempted to understand the irregularity in Bitcoin prices, which can lead to better prediction of its price movement (Ramirez et al., 2018).

While few studies have shown that Bitcoins are subjected to market inefficiency as well as gradually moving toward becoming efficient (Kristoufek, 2018, Zargar and Kumar, 2019), none of these studies have examined the irregular movements in Bitcoin price series as a multiscale problem. A multiscale situation arises when a time series consists of components, which are at disparate scales and require analysis from different levels of resolution (e.g., hourly, weekly). The dynamics of the series at these resolutions also vary, thereby affecting the inferences of the underlying process. Hence, multiscale series warrants decomposition and aggregation of the original series at different time resolution. This study intends to investigate the multiscale nature of the Bitcoin price series. The exploration of multiscale movement features of Bitcoin price series can also lead to better prediction of Bitcoin prices.

The application of computationally intensive techniques using machine learning algorithms is extensively utilized to develop profitable stock trading strategies. Multiple studies have been conducted to examine whether machine learning tools can be leveraged to anticipate stock predictions in a consistent manner (Nardo et al., 2016, Patel et al., 2015). Studies have also explored machine learning algorithms to predict Bitcoin prices (Ciaian et al., 2016, McNally et al., 2018). However, there are limited studies on the multiscale property of Bitcoin prices to assess the prediction of the price series (Pichl and Kaizoji, 2017). Being a highly volatile currency and subject to lot of ambiguity, it is an interesting area of research to explore whether machine learning algorithms can be used to effectively predict Bitcoin prices.

The objective of this study is twofold. Firstly, it examines the nature of Bitcoin price series and assesses their multiscale properties. The decomposition of Bitcoin price series by using an algorithm is effective, if it enables the identification of different sub-series belonging to different scales. While the variants of auto-regressive conditional heteroscedasticity (ARCH) models and wavelet decomposition are widely used to identify multiscale properties of volatile series, scholars have subjected them to many shortcomings. Moreover, the ARCH models are highly model-specific and the decomposition scales are not adjustable in wavelet analysis (Wen et al., 2017). The empirical mode decomposition (EMD) techniques generate a more adaptable decomposition for nonstationary series (Huang et al., 1998). Hence, this study proposes to assess the multiscale properties of Bitcoin price series using EMD (Lahmiri, 2016a, Lahmiri, 2016b) and extended versions of EMD algorithms (Torres et al., 2011) like empirical ensemble mode decomposition (EEMD) and complete empirical ensemble mode decomposition (CEEMD).

Secondly, it uses support vector machine (SVM) learning algorithm to predict the Bitcoin prices. The widespread application of SVM learning, such as addressing real-life engineering issues (Ahmadi and Mahmoudi, 2016, Ahmadi, 2016, Ahmadi and Chen, 2019) substantiates its ability to analyze complex time-series data. Comparing the myriad financial time series prediction methods, scholars have shown SVM based regression models as a promising alternative because it gives more optimal results (Lahmiri, 2013, Peng et al., 2018).

Section snippets

Background

With over 90% of the investor being actively involved in cryptocurrency trading (Hasso et al., 2019), cryptocurrencies have received phenomenal attention and growth during the last few years. Being the most popular cryptocurrency, different studies have explored the nature of Bitcoin price series from multiple perspectives. One set of studies have investigated the market efficiency of Bitcoin prices and another set of studies have tried to examine the relation of Bitcoin prices with other asset

Methodology

The volatile price series like gold price and crude oil prices are too complicated to process because of its nonlinear and nonstationary characteristics. Bitcoin prices like other asset classes are characterized by non-linear dynamics, which include non-periodicity, randomness, and presence of spectrum of scaling components in the series (El Alaoui et al., 2019). The complexity of the dynamics of such series are best analyzed using the non-linear models. Research has shown improved forecasting

Data analysis

For the study, daily Bitcoin price data from 322012 to 3102018 is sourced from https://in.investing.com. Fig. 1 shows the Bitcoin price series and the log of Bitcoin prices’ graph. From the graph, it can be seen that though the Bitcoin prices have shown an upward trend, the series has been highly volatile and nonlinear in nature.

From the Bitcoin price and logarithm price graph, it can be observed that the fluctuations in prices are rapid and varying in level. Hence, such price series are

Results and discussion

To better understand the forecasting performance, the CEEMDAN-SVM is compared with EMD-SVM. Table 3, Table 4 give the MSE and MAPE of CEEMDAN series and EMD series.

In light of the fact that MAPE is a more stable error metric than MSE, the prediction error of CEEMDAN is less than the prediction error of EMD for MAPE. This is because the CEEMDAN extracts more effective features from the original data. Relatively, EMD-SVM is better in the case of short-term forecasting with regard to MSE measure

Conclusion

Cryptocurrencies have created disruption in the financial markets and pose huge challenges for monetary policy makers. This has also raised questions related to understanding the qualities of Bitcoin as an asset. Being a highly volatile currency and a speculative asset, it is imperative to understand the price behavior of Bitcoins. The study applies the CEEMDAN analysis to examine the multiscale nature of Bitcoin prices. The application of CEEMDAN sheds light on the distinct nature and price

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Divya Aggarwal has done her Ph.D. in the area of behavioral finance from XLRI - Xavier School of Management, Jamshedpur, India. Her research interests include financial decision making and behavioral biases. She has worked with Mckinsey Knowledge Centre, KPMG India and Avendus capital before starting an academic career. She is a recipient of UK Case Centre scholarship and has won academic awards of national and international repute. She won the best paper award at 11th ISDSI Conference by

References (70)

  • CaporaleG. et al.

    Persistence in the cryptocurrency market

    Res. Int. Bus. Finance

    (2018)
  • CorbetS. et al.

    Exploring the dynamic relationships between cryptocurrencies and other financial assets

    Econom. Lett.

    (2018)
  • da Gama SilvaP.V. et al.

    Herding behavior and contagion in the cryptocurrency market

    J. Behav. Exp. Finance

    (2019)
  • DyhrbergA.H.

    Bitcoin, gold and the dollar –A GARCH volatility analysis

    Finance Res. Lett.

    (2016)
  • GiudiciP. et al.

    What determines bitcoin exchange prices? A network VAR approach

    Finance Res. Lett.

    (2019)
  • HassoT. et al.

    Who trades cryptocurrencies, how do they trade it, and how do they perform? Evidence from brokerage accounts

    J. Behav. Exp. Finance

    (2019)
  • HayesA.S.

    Cryptocurrency value formation: An empirical study leading to a cost of production model for valuing bitcoin

    Telemat. Inform.

    (2017)
  • HsuM.-W. et al.

    Bridging the divide in financial market forecasting: machine learners vs. financial economists

    Expert Syst. Appl.

    (2016)
  • JiQ. et al.

    Network causality structures among Bitcoin and other financial assets: A directed acyclic graph approach

    Q. Rev. Econ. Finance

    (2018)
  • KatsiampaP.

    Volatility estimation for Bitcoin: A comparison of GARCH models

    Econ. Lett.

    (2017)
  • KhuntiaS. et al.

    Adaptive market hypothesis and evolving predictability of bitcoin

    Econ. Lett.

    (2018)
  • KristjanpollerW. et al.

    Asymmetric multifractal cross-correlations between the main world currencies and the main cryptocurrencies

    Physica A

    (2019)
  • KristoufekL.

    On Bitcoin markets (in) efficiency and its evolution

    Physica A

    (2018)
  • LahmiriS.

    Wavelet low-and high-frequency components as features for predicting stock prices with backpropagation neural networks

    J. King Saud Univ.-Comput. Inf. Sci.

    (2014)
  • LahmiriS.

    Long memory in international financial markets trends and short movements during 2008 financial crisis based on variational mode decomposition and detrended fluctuation analysis

    Physica A

    (2015)
  • LahmiriS.

    A variational mode decompoisition approach for analysis and forecasting of economic and financial time series

    Expert Syst. Appl.

    (2016)
  • LahmiriS.

    Intraday stock price forecasting based on variational mode decomposition

    J. Comput. Sci.

    (2016)
  • LahmiriS.

    Minute-ahead stock price forecasting based on singular spectrum analysis and support vector regression

    Appl. Math. Comput.

    (2018)
  • LahmiriS. et al.

    Multi-fluctuation nonlinear patterns of European financial markets based on adaptive filtering with application to family business, green, Islamic, common stocks, and comparison with Bitcoin, NASDAQ, and VIX

    Physica A

    (2020)
  • MallquiD.C. et al.

    Predicting the direction, maximum, minimum and closing prices of daily Bitcoin exchange rate using machine learning techniques

    Appl. Soft Comput.

    (2019)
  • MingL. et al.

    The double nature of the price of gold -A quantitative analysis based on Ensemble Empirical Mode Decomposition

    Resour. Policy

    (2016)
  • NadarajahS. et al.

    On the inefficiency of Bitcoin

    Econ. Lett.

    (2017)
  • PatelJ. et al.

    Predicting stock and stock price index movement using trend deterministic data preparation and machine learning techniques

    Expert Syst. Appl.

    (2015)
  • PengY. et al.

    The best of two worlds: Forecasting high frequency volatility for cryptocurrencies and traditional currencies with Support Vector Regression

    Expert Syst. Appl.

    (2018)
  • StavroyiannisS. et al.

    Herding behavior in cryptocurrencies revisited: Novel evidence from a TVP model

    J. Behav. Exp. Finance

    (2019)
  • Cited by (39)

    • Empirical mode decomposition based techniques for imaging of shallow delamination in concrete using impact echo

      2023, Mechanical Systems and Signal Processing
      Citation Excerpt :

      A similar investigation on early detection of machinery faults found CEEMDAN effective in denoising of vibration signals through the reconstruction of defect-bearing IMFs [44]. Besides the establishment of the method in machinery faulty detection, CEEMDAN is also applied in weather forecasting [45], market prediction [46], and medical studies [47]. Unfortunately, two major issues still remains unaddressed with this method that inhibits the near-perfect decomposition of components.

    • How well do investor sentiment and ensemble learning predict Bitcoin prices?

      2023, Research in International Business and Finance
      Citation Excerpt :

      Indeed, by adopting nonlinear interactions, the predictive power of regression models can be enhanced in the financial domain. In this context, SVR and SVM have successfully been applied to the cryptocurrency market (Aggarwal et al., 2020; Mallqui and Fernandes, 2019; Peng et al., 2018; Poongodi et al., 2020). However, the limited performance of SVR occurs because SVR is implemented using approximation algorithms due to computational complexity.

    View all citing articles on Scopus

    Divya Aggarwal has done her Ph.D. in the area of behavioral finance from XLRI - Xavier School of Management, Jamshedpur, India. Her research interests include financial decision making and behavioral biases. She has worked with Mckinsey Knowledge Centre, KPMG India and Avendus capital before starting an academic career. She is a recipient of UK Case Centre scholarship and has won academic awards of national and international repute. She won the best paper award at 11th ISDSI Conference by Decision Science Institute. In 2014, she won the Global Peter Drucker essay challenge under the student’s category and was invited to Vienna, Austria.

    Shabana Chandrasekaran is an Assistant Professor of Decision Sciences area at Xavier Institute of Management, Bhubaneshwar, India. She received her Ph.D. in Information Systems from XLRI - Xavier School of Management, Jamshedpur, India, and was awarded the XLRI medal for outstanding student of the year. She facilitates courses on Business Analytics & Data Mining, Business Statistics, and Research Methods. Her research interests include machine learning, text mining, and natural language processing and social media analytics. Her research work has been published in refereed journals and international conferences.

    Balamurugan Annamalai is an Assistant Professor in the area of Marketing at Indian Institute of Management (IIM), Sambalpur, India. He holds a doctorate (FPM) in marketing from XLRI - Xavier School of Management, Jamshedpur, India. His research interest is in the area of customer engagement on social media. He is an engineering graduate in computer science from IIT (BHU), Varanasi, and a management postgraduate from IIM Calcutta. He has consulting and industry experience in brand management and marketing research.

    View full text