Table of Contents
II. Stock Sentiment Literature
A. Bitcoin Framework
A.1. Bitcoin Sentiment Literature
III. Data and Methodology
A. Thomson Reuters MarketPsych Indices
A.1. Constructing Sentiment Measures
IV. Empirical Results
A. Bitcoin Sentiment Measures and Future Returns
B. Drivers of Bitcoin Tradevolume
In this paper, I analyze the relation between daily bitcoin returns and sentiment, using a dataset reaching from 2013 to 2018. I find that daily bitcoin returns are not only affected contemporaneous by the bitcoin-sentiment measures, but also in the next three days - while established stock-market sentiment measures provide no explanatory power. Additionally, the negative emotions show return-reversal patterns as often observed in sentiment-induced mispricing literature, resulting in higher returns the next two days, after affecting returns negatively today. I further find that trading volume affects returns positively today, in the next four days, and that it is connected with stock-market measures. For what I wisdom, this is also the first academic research that uses the recently introduced Thomson Reuters MarketPsych Indices on cryptocurrencies.
Keywords: Bitcoin, Returns, Reversa, Sentiment, TRMI
JEL Classification Codes: G02, G12
In the short era of nine years Bitcoin - the first digital peer-to-peer cryptocurrency - made its way into mainstream media and consciousness by potentially disrupting the financial mechanisms as we know them. Unlike with traditional fiat money (i.e. USD, EUR etc.), no banks or other financial institutions are needed. Neither for issuing money, nor for transferring or validating transactions. The blockchain technology keeps track of every single transaction in a public distributed ledger which encompasses the double spending problem - leaving financial intermediaries out of the process - and paving the way for industry-shaping blockchain innovations. While this sounds like the next financial revolution, Bitcoins market capitalization with 116.6 billion USD and the whole cryptocurrency market capitalization of 274 billion USD is still far from being adopted as a mainstream asset class compared to the equity market capitalization of 69 trillion USD in equity markets. Mainly the lacking regulation and dramatic fluctuations in Bitcoins price keeps institutional investors and the broad mass away, e.g. after closing at the all-time high of 19.497 USD on 17th December 2017, one bitcoin is trading at 6.100 USD at the 27th of June 2018 - losing 70% of its value in six months. Price dynamics like this fueled researchers’ attention on Bitcoin. Approaches trying to put on a value for Bitcoin by supply-demand assumptions (Jenssen, 2014) or cost-of-production models (Garcia et al., 2014, Hayes, 2015) were accompanied by mainstream media coverage. For example CNBC regularly covers cryptocurrency specific contents and interviewed finance-famous celebrities like noble prize winner Robert Shiller, investor legend Warren Buffet or the CEO of JPMorgan Chase Jamie Dillon regarding Bitcoin. They compare Bitcoin with the tulip mania of 1637, coming to the conclusion that bitcoins are worth nothing at all and forecast a bad ending. This led researchers to examine on Bitcoin bubble behavior (Cheung, Roca and Su (2015), Cheah and Fry (2015)) and in the direction of investor sentiment as an important part of Bitcoin pricing. Because of the missing of a fundamental value, the Bitcoin market is mainly driven by speculators, trend chasers, short-term investors and noise traders (Kristoufek, 2013). Meaning the price of Bitcoin should depend a lot on market sentiment, tone, visibility, media coverage and social media activity.
Sentiment measures like Google Trends and Wikipedia page views (Kristoufek, 2013), Twitter posts (Kaminski, 2014) or news articles (Polasik et al., 2015) in fact have a significant effect on bitcoin returns, although most of the analysis only had data for months or took place in Bitcoins inefficient market stage before the second half of 2013 (Urquhart, 2016).
In this paper I contribute to the prior research with a much richer dataset with daily observations over five years, taking supply-demand, macroeconomic factors and sentiment measures in consideration. For what I wisdom, this is also the first academic research using the recently introduced Thomson Reuters MarketPsych Indices (TRMI) for cryptocurrencies as sentiment measures, which scores are calculated with over two million cryptocurrency-specific entries daily - combining social media platforms, websites and popular newspapers as sources. To the best of my knowledge this research is also the first one to examine on bitcoin return reversals regarding sentiment measures (Kuo Chen et al. (2017) used past returns as sentiment measure).
Besides support for almost no correlation of stock-market returns and established stock-market sentiment measures with bitcoin returns as in Makrochiriti and Moratis (2016), Bianchi (2017) and Kuo Chen at al. (2017), I also extend the prior research on sentiment induced return reversals. Bitcoin sentiments not only affect returns today, but also the future returns up to four days. The main finding of the paper is the negative contemporaneous effect of negative emotions on daily bitcoin returns, which reverse in the next two days, driving returns up. A brief analyses on the drivers of Bitcoins trading volume offers potentially connections to stock-market and economic measures.
The rest of the paper proceeds as follows. Section 2 provides an overview on stock and Bitcoin related sentiment literature. Section 3 describes the data and methodology used. Section 4 shows the empirical results and interpretations of the findings. Section 5 concludes and provides areas for future research.
II. Stock Sentiment Literature
Today, in the academic research, it is generally accepted that noise traders relying on sentiment can drive prices away from fundamental values. Barberis, Shleifer, and Vishny (1998) and De Long et al. (1990) show in simple models that irrational noise traders with erroneous beliefs affect asset prices and that arbitrageurs are not willing to bet against them due to limits to arbitrage. Further, De Long et al. define the term “noise trader risk” meaning that the unpredictability of future opinions of the noise traders can push prices even further up (if noise traders become even more optimistic) or down (even more pessimistic respectively) - deterring arbitrageurs from taking short positions, even when there is no fundamental risk. Important research in the area of limits to arbitrage was added by Shleifer and Vishny (1997). It is also of relevance for this paper, as bitcoins were subject to short-selling impediments and therefore prone to build bubbles under high investor sentiment, as rational investors could not bet against overpricing at least until late 2017 when futures on bitcoin prices went live.
Investor sentiment is broadly defined by Baker and Wurgler (2006) as a belief about future cash flows and investment risks that is not justified by the facts at hand. In order to research the connection between sentiment and asset prices, one has to build measures of subjective investors’ sentiment or objective textual sentiment.
In the context of textual sentiment several sources and methods have been used to examine whether the public opinion relating a topic is positive, neutral or negative. The sources used in the literature are corporate disclosures, news stories and analyst reports, and internet board postings. Antweiler and Frank (2004) examined messages from the “yahoo! Finance message board”. Via machine learning algorithms they show that a higher degree of bullish messages leads to an increase in trading volume. Sprenger and Welpe (2010) used the same approach regarding stock related tweets on Twitter. They observe a strong correlation between abnormal returns and the bullishness of tweets. Tetlock (2007) applied a Harvard IV-4 dictionary based approach on the tone of the Wall Street Journal’s column “Abreast of the market” from 1984 to 1999. A more pessimistic tone leads to negative returns the next day and a reversal over the next week. Garcia (2013) extends and confirms the results of Tetlocks study with data from 1905 to 2005 on two New York Times columns. Loughran and McDonald (2011) proposed a new word list for financial aspects to reduce the misclassification of negative words, after they found that nearly three-fourths of the negative words in the Harvard IV-4 dictionary are not negative in financial terms. Consequently, a term weighting scheme lead to even higher explanatory power for the word lists although they cannot predict stock returns. Kearney and Liu (2014) provide an overview and comparison of methods and sources being used for textual sentiment until 2013. Based on the Loughran and McDonald word list, Chen et al. (2014) analyze the tone of seekingalpha.com opinions and find statistically and economically significant predictions for future stock returns and earnings surprises. Hillert and Ungeheuer (2018) even show that visibility itself leads to higher returns regardless of the tone. They address this phenomenon to “bad news” being the corporate governance channel, meaning that “bad news” leads to more transparency and more honest manager behavior, thus leading to long-run corporate improvements and higher returns. While the “good news” serve as the advertising channel. According to the current state of the literature in textual sentiment analyses, it is generally accepted that sentiment does affect prices or at least delivers additional information on stock returns.
Researchers used several other proxies for sentiment than from corporate disclosures, journals or social media platforms. Hirshleifer and Shumway (2003) used the weather as a proxy for sentiment, showing that on sunnier days, returns are higher. Another study from Edmans, Garcia and Norli (2005) researches the effect of international sport outcomes, observing that a loss leads to falling prices in the losing-country’s stock market – particularly for small stocks. Kaplanski and Levy (2010) find that aviation disasters lead to sharp decline in returns and fully reversal after ten days. Surveys from households to capture investor sentiment were used by Charoenrook (2003), Brown and Cliff (2005) and Lemmon and Portniaguina (2006), coming to the conclusion, that stock prices fall after periods of optimism. Baker and Wurgler (2006) constructed a composite sentiment index, containing six underlying sentiment proxies, coming to similar results. Stambaugh, Yu, Yuan (2012) use the same sentiment index as Baker and Wurgler, analyzing 11 well-documented anomalies with regard to sentiment and short-sale impediments. They find that the anomaly-mispricings are stronger and predict significant negative returns, following high sentiment months. Huang et al. (2015) further improve the Baker and Wurgler index by sorting out approximation errors or noise in the proxies with a partial least-square method, leading to a higher R² and improving the predictive power.
In the digitalization era, Google search frequencies tend to reflect retail investors interest and sentiment. Research in this field has been undertaken by Da et al. (2011), providing evidence of a strong correlation between search frequency and trading volume. Another sentiment index from Da, Engelberg and Gao (2014) called FEARS index uses Google search queries on words like “unemployment”, “recession”, “find a job” etc. This way they can build an honest household sentiment measure and find that spikes in the FEARS index (meaning a rise in fear, uncertainty etc.) results in low returns the next day. However this effect is reversed in the following two days, which is consistent with temporary sentiment-induced mispricing. Whether the FEARS index has also an effect on bitcoin returns or not will be analyzed in Chapter IV.
The most recent studies of media sentiment have incorporated intraday sentiment analysis and psychological sentiment analysis. Renault (2017) constructed a lexicon for the jargon of online investors on the social media platform StockTwits – this approach significantly outperformed the dictionary based methods. He then compared the intraday sentiment with intraday stock returns on an S&P 500 ETF and finds that the first half-hour sentiment change predicts the last half-hour S&P 500 ETF returns. In particular, this effect seems to come from inexperienced noise traders. Sun, Najand and Shen (2016) come to similar results using TRMI intraday sentiment data, observing the effect for the last two hours of a trading day. Shen (2017) finds evidence for forecasting the next five days’ commodity returns using TRMI psychological sentiments (optimism, fear and joy).
Shen and Chen (2018) research company-specific psychological sentiments (optimism, joy, anger, gloom, fear and stress) on Dow Jones Industrial Average listed stocks via the TRMI data. They find significant predictive power for the large-cap stocks returns, stronger sentiment effects under high media attention and that a single-dimension sentiment cannot substitute the psychological sentiments.
Today it is generally accepted that investor sentiment affects stock prices. Summarizing the sentiment analysis on stocks, it is evident that market sentiment can drive prices away from fundamentals. Sentiment measures like the CBOE Volatility Index, Michigan Consumer Sentiment Index or TRMI are established indices in academic research.
A. Bitcoin Framework
Before I summarize the sentiment literature on cryptocurrencies - specifically bitcoin -, it is important to mention some of the specific characteristics of Bitcoin. Bitcoin was first mentioned in a white paper by Satoshi Nakamoto (2009) as a decentralized peer-to-peer cryptocurrency. It has low transaction fees, is fast, flexible, transparent and decentralized (Kuo Chen et al. (2017). The most important innovation behind Bitcoin is the distributed ledger technology, which encompasses the problem of double-spending and therefore works without financial intermediaries. Transactions are public, unwoundable and verified into blocks by nodes of the network, which solve an algorithmic task to do so. For the technological background and functioning of the blockchain transactions in detail see Jenssen (2014) or Boehme et al. (2015). For a good overview on the cryptocurrency ecosystem see Hileman and Rauchs (2017). Academic research regarding bitcoin is rapidly emerging. Researchers focused on the question whether Bitcoin is a “real” currency or behaves more like a speculative asset. Work in this direction has been undertaken by Baur et al. (2017), Barieviera et al. (2017), Desan (2012) and Yermack (2015). All concluding that Bitcoin is not “money”, due to its weak properties as a medium of exchange, unit of account or store of value. For definitions on money and the regulatory framework regarding fiat currencies and Bitcoin see Allen (2017). Bitcoin is therefore commonly viewed as a speculative asset.
Because valuing an asset is a crucial part for investing, researchers tried to put on a fundamental value for Bitcoin. Jenssen (2014) attributes the value of a bitcoin to computer-power used for mining bitcoins, as it is resource-intensive, as well as to demand and supply arguments since the supply of bitcoins is fixed at 21 million coins. Garcia et al. (2014) also used the approach of the producing costs, leading to a lower boundary for the fundamental value of Bitcoin. Elaborating on this idea, Hayes (2015) formalizes a cost-of-production model for valuing cryptocurrencies. With a OLS regression containing only three variables, namely computational power (difficulty to mine), rate of coin production and which algorithm is used, the model yields to a significant R² of 0.84 for the value formation. Interestingly the variable “total supply” is not a significant driver of the price. Nonetheless Hayes proposed that there might variables missing, like a risk-premium for example.
The majority of academics did not try to address Bitcoin a fundamental value, but investigated on the behavior of Bitcoins price dynamics. Bitcoins price is characterized by extreme volatility and bubble behavior as it is common with assets laxity in regulation, overtrading and overestimated growth prospects (Cheung, Roca and Su, 2015). Applying the Phillips-Shi-Yu (2013) model, which has proven to detect bubbles in stock and currency markets, to Bitcoin data from 2010-2014 Cheung, Roca and Su (2015) find 33 bubbles in the Bitcoin market. Besides three longer-lived bubbles reaching from 66-109 days due to certain events like a hacking-attack to the exchange Mt. Gox, the others only lasted for a few days. In the end known economists like Robert J. Shiller (2017) stated the fundamental value is zero and any attempt to give Bitcoin a value is at best ambiguous. Cheah and Fry (2015) are showing exactly this in their model. Their estimated long-term fundamental value variable is not statistically different from zero in bubble price rises. Other research in the field of speculative bubbles regarding Bitcoin were conducted by Bouchaud and Donier (2015), Corbet, Lucey and Yarovaya (2017) and Blau (2018).
Summarizing the prior, Bitcoin is usually labelled a complete speculative asset. Because of the missing of a fundamental value, also the so-called fundamentalists are missing in the Bitcoin market, as a fair price is unknown. This leaves it to speculators, trend chasers, short-term investors and noise traders (Kristoufek, 2013). Meaning the price of Bitcoin should depend a lot on market sentiment, tone, visibility, media coverage and social media activity.
A.1. Bitcoin Sentiment Literature
Recent research in the field of Bitcoin related sentiment has been undertaken – in contrast to the vast stock sentiment literature, sentiment related research in this new asset class relies mainly on textual sentiment analysis from online-sources like websites and social media platforms. Kristoufek (2013, 2015) shows a significant bi-directional relationship between both; Google search queries, Wikipedia daily views and the price of Bitcoin. Showing that not only do search queries affect the price, but also the price affects the search queries, therefore leading to frequent bubbles in the market. Garcia et al. (2014) also used Google queries and Wikipedia views, but additionally obtained tweets on Twitter and reshared posts on facebook as a word-of-mouth variable, coming to comparable results. With a vector autoregression model (VAR), they identify a social- and an adoption-cycle feedback loop. The social-cycle runs as follows: Growing popularity of Bitcoin leads to higher search volumes, which increases the social media activity regarding Bitcoin, encouraging individuals to purchase bitcoins and thus leading to price increases. Price increases eventually reinforce the search volumes. Kaminski (2014) examines emotional tweets (positive, negative and uncertainty) on Bitcoin, finding a positive correlation to the closing price, trading volume and intraday spread of bitcoins. Additionally, he runs Granger causality tests and shows that the tweets are not predicting market values, but reflecting the trading dynamics.
A common textual analyses, often used on stocks, via the LexisNexis database on articles regarding Bitcoins tone has been conducted by Polasik et al. (2015). They find a strong and highly significant effect of popularity (either measured by Articles or Google search queries) and also a significant positive effect for the tone on the monthly returns – arguably it is not advisable to apply monthly measures in such a volatile market as the cryptocurrency-market with immense daily fluctuations and outliers as we can see from the daily returns distribution in Fig. 1.
[Insert Fig. 1. about here]
Bilanakos et al. (2015) use a machine-learning algorithm, namely Support-Vector-Machines to research the mood on twitter. In short-run regressions they find positive effects for the twitter-sentiment ratio and also for Wikipedia views on the Bitcoin price. Mai et al. (2016) are comparing Twitter and Bitcointalk.org Forum sentiment via the Loughran and McDonald dictionary with Vector Error Correction Models (VECMs). The analysis offers a better understanding of how sentiment and information is spread along the two platforms and also distinguishes between different users; namely the “silent majority” and the “vocal minority”. The data favors Forum over Twitter sentiment as an important indicator for future bitcoin returns and the posts by the “silent majority” over the “vocal minority”. The VAR model by Makrichoriti and Moratis (2016) sheds light onto Bitcoins relationship to overall economic variables (as a systemic risk factor) and U.S. market sentiment. The results suggest that Bitcoin is mostly affected by investor sentiment and not by the systemic risk factor. Wang and Vergne (2017) find counterintuitive results. They report a significant negative relationship between public interest (so-called buzz) as well as no significant effect for bad press on weekly returns.
In this paper, I contribute to the research with a much richer dataset from 12/27/2013 until 02/28/2018, covering the immense Bitcoin price run-up to the all-time high of $19.497 and drop to $7.621 within 2 months (See Fig. 2).
[Insert Fig. 2. about here]
Also the media attention had its peak in December 2017 with all-time highs in Google Trends (see Fig. 3) and news magazine coverage as well.
First, I will construct two sentiment indices - one measuring positive and another one measuring negative emotions for the Bitcoin data. For what I wisdom, this is the first academic research using the newly introduced TRMI for cryptocurrencies. The focus of the first regressions will be on Bitcoins relationship with macroeconomic factors, stock-market returns and established stock-market sentiment measures. Afterwards, I will examine on return reversals as in Da, Engelberg and Gao (2014) and if the built sentiment measures also affect bitcoin returns in the future. With regards to the visibility theory of Ungeheuer and Hillert (2018) the buzz for Bitcoin will be analyzed, both on returns and on transaction volume.
[Insert Fig. 3. about here]
III. Data and Methodology
In this research, I use data regarding daily bitcoin prices and transaction volume from Coinmarketcap.com. The reason for choosing Coinmarketcap along various other sources (see Table 1) is due to the methodology. Coinmarketcap weights volumes and prices from all major exchanges regarding their contribution to the whole Bitcoin market. Also it is most often used in prior research papers and is the market leader in this area. For the sentiment data MarketPsych LLC provided the TRMI cryptocurrency data. Blockchain related data like hash rate, mining difficulty and supply growth are obtained via quandl.com. The data sources reach from 12/27/2013 to 02/28/2018. The starting point of the data is chosen because it is the first day of Coinmarketcap providing volumes, also research found that bitcoin prices were inefficient before the second half of 2013 (Urquhart, 2016, 2018). In addition, I have been provided with not yet public released data on the FEARS Index from 2012 to 2016. Take note that this data was derived by research assistants of Prof. Da, thus have not been fully inspected and are therefore subject to change. The data range leads to 1525 observations for the majority of the variables. For a detailed overview on variables used in this research see the descriptive statistics in Table III in Subsection A.2.
[Insert Table I about here]
Because prior research came to the conclusion, that Bitcoin has nearly no correlation with stock market or underlying economy measures (Makrichoriti and Moratis (2016) and Kuo Chen et al. (2017)), a bitcoin-specific pricing model as basis is used. Nonetheless, my model controls for macroeconomic factors. The base model for explaining daily bitcoin returns is derived from Wang and Vergne (2017).
The reason for choosing their model as a basis is because of the detailed methodology description and robustness, as well as data availability to extend their dataset. They use weekly data and focus on liquidity, technical development, supply growth, negative publicity and public interest as explanatory variables for five cryptocurrencies returns. Because I don’t need to compare the technical development between different cryptocurrencies, the technical development variable is left out. As well as the liquidity measure, since Bitcoin is the most tradable cryptocurrency anyways. The negative publicity measure is only news based (negative mentions in the factiva database regarding Bitcoin) and public interest is measured by the alexa web traffic rank and bing search queries. Instead I will use TRMI data, which combines news and social media sources, which should lead to more comprehensive results. The public interest measure is substituted by the daily U.S. Google Trends search queries for the same reason. In addition, I added several blockchain-related supply-demand factors as Kristoufek (2015) suggests after finding significant supply-demand price drivers for Bitcoin. Also the weekly time trend which controls for stock market measures in Wang and Vergne (2017) is called the macroeconomic trend in my model. For the detailed measures descriptions see Subsection A.2.
 To date it is not clear who the person or group behind the synonym Satoshi Nakamoto is.
 For offering computational power to confirm each transaction in the blockchain, the nodes or miners are rewarded. The term mining relates to gold miners, searching for the next block in the blockchain.
- Quote paper
- Andreas Bialek (Author), 2018, Bitcoin Pricing. An Empirical Analysis, Munich, GRIN Verlag, https://www.grin.com/document/459668