Nowcasting the Bitcoin Market with Twitter SignalsLast Update: September 2014. An early version of this paper was presented at Collective Intelligence 2014, MIT, Cambridge, USA, June 10-12 2014.

# Nowcasting the Bitcoin Market with Twitter Signals

## Abstract

This paper analyzes correlations and causalities between Bitcoin market indicators and Twitter posts containing emotional signals on Bitcoin. Within a timeframe of 104 days (November 23 2013 - March 7 2014), about 160,000 Twitter posts containing ”bitcoinâ and a positive, negative or uncertainty related term were collected and further analyzed. For instance, the terms ”happy”, ”love”, ”fun”, ”good”, ”bad”, ”sad” and ”unhappy” represent positive and negative emotional signals, while ”hope”, ”fear” and ”worry” are considered as indicators of uncertainty. The static (daily) Pearson correlation results show a significant positive correlation between emotional tweets and the close price, trading volume and intraday price spread of Bitcoin. However, a dynamic Granger causality analysis does not confirm a statistically significant causal effect of emotional Tweets on Bitcoin market values. To the contrary, the analyzed data shows that a higher Bitcoin trading volume Granger causes more signals of uncertainty within a 24 to 72-hour timeframe. This result leads to the interpretation that emotional sentiments rather mirror the market than that they make it predictable. Finally, the conclusion of this paper is that the microblogging platform Twitter is Bitcoinâs virtual trading floor, emotionally reflecting its trading dynamics.

\acmVolume

2 \acmNumber3 \acmArticle1 \articleSeq1 \acmYear2013 \acmMonth5 \SetAlFnt\algofont \SetAlCapFnt\algofont \SetAlCapNameFnt\algofont \SetAlCapHSkip0pt \IncMargin-

## 1 Abstract & Keywords

This paper analyzes correlations and causalities between Bitcoin market indicators and Twitter posts containing emotional signals on Bitcoin. Within a timeframe of 104 days (November 23 2013 - March 7 2014), about 160,000 Twitter posts containing ”bitcoin” and a positive, negative or uncertainty related term were collected and further analyzed. For instance, the terms ”happy”, ”love”, ”fun”, ”good”, ”bad”, ”sad” and ”unhappy” represent positive and negative emotional signals, while ”hope”, ”fear” and ”worry” are considered as indicators of uncertainty. The static (daily) Pearson correlation results show a significant positive correlation between emotional tweets and the close price, trading volume and intraday price spread of Bitcoin. However, a dynamic Granger causality analysis does not confirm a causal effect of emotional Tweets on Bitcoin market values. To the contrary, the analyzed data shows that a higher Bitcoin trading volume Granger causes more signals of uncertainty within a 24 to 72-hour timeframe. This result leads to the interpretation that emotional sentiments rather mirror the market than that they make it predictable. Finally, the conclusion of this paper is that the microblogging platform Twitter is Bitcoinâs virtual trading floor, emotionally reflecting its trading dynamics.2

Keywords: Bitcoin, Twitter, Emotions, Sentiments, Prediction, Market Mirror

## 4 Correlation Analysis

shows that the biggest Bitcoin exchange platforms in terms of trading volume seem to be most sensitive for negative tweets and signals of uncertainty () on Bitcoin. Also, a day with a low amount of negative tweets correlates with a higher close price. It can complementary be noted that (in three out of 4 cases) a higher ratio of positive to negative tweets (=more positive than negative tweets) is accompanied by a higher close price. 8 To the contrary, and alone do not seem to correlate with the close price. Now, the previously introduced sentiment signals will sequentially be tested with further market indicators for BitStamp, the current biggest exchange market:9

Following results can be summarized from :

1. Emotions on Twitter, especially and , positively correlate with the BitStamp trading volume (cf. ).

2. The sum of emotions and signals of uncertainty also fuel intraday price volatilities, such as the (), reflecting the difference between the highest and lowest intraday trading price of BitStamp on a given day.

3. The more negative emotions and signals of uncertainty appear on a given trading day, the more likely is a lower close price. Mixed with signals of uncertainty, negative sentiments may be interpreted as a sign of dissatisfaction or pessimism by traders (and their observers). As illustrates, negative emotions especially seem to appear on trading days with a decreasing close price.

Current data suggests that the term of ’Nowcasting’ (predicting the present), [Giannone et al. (2008), Choi and Varian (2012)] might be applicable to Bitcoin intraday market development. However, the specific cause and effect relationship in terms of predictability will be tested in . So far, correlations in this analysis follow the simple and assumption that the relation between variables is linear, which is hardly satisfied in the often appearing random walk of financial market movements. In order to find out more about the prediction value of Twitter sentiments for the development of the BitStamp market, another calculation will be conducted. For this purpose, we will project the BitStamp close price each -2 and +2 days into the past and future respectively, while applying a moving average for related sentiment data.

Following shows that emotions negatively correlate with the future close price. Especially negative tweets are negatively correlated with the BitStamp close price within a 48-hour timespan (cf. ). Further, the sum of emotions and positive tweets both show a negative correlation with regard to the future close price. Further data suggests that signals of uncertainty (hope, fear, worry) do not only amplify trading volumes (cf. ) but also the close price. The higher the amount of uncertainty signals, the lower the BitStamp close price within 2days. However, we can assume that tweets are not always ” ” as timestamps might suggest and thus a time lag in retweets may make an interpretation less robust.

demonstrates the correlation between the BitStamp (intraday) trading volume as well as the intraday price spread (High - Low) within a 48-hour timespan. As far as data enables an interpretation, a high amount of emotions and signals of uncertainty (in the present) correlates with high trading volumes within the last 24 hours. Especially, a high amount of negative signals is a key influencer for trading volume within the past 24 hours. Again, a more balanced (= lower) ratio of positive and negative sentiments also contributes to a higher trading volume, while for intraday price spreads (), the current data does not support any significant influence by Twitter sentiments.

## 5 Granger-Causality Analysis

”Correlation does not imply causation” is a long-known phrase in science. Thus, in order to go beyond correlations and develop a better understanding with regard to the causalities, we apply a Granger causality analysis [Granger (1969)] to the daily time series of Twitter sentiments and the Bitcoin market movement. Granger causality is a statistical concept of causality that can be used to determine if one time series is useful in forecasting another. A scalar is said to Granger-cause scalar if is better predicted by using the past values of than by solely relying on past values of . If causes and does not cause , it is said that unidirectional causality exists from to . If does not cause and does not cause , then and are statistically independent. If causes and causes , it is said that feedback exists between and .

The calculation of Granger causality requires that the time series have to be covariance stationary, so an Augmented Dickey-Fuller [Dickey and Fuller (1979)] test has been done first, in which the () of non-stationarity was rejected at the 0.05 confidence level for all Twitter and Bitcoin time series variables. All evaluated data is stationary. 10 To test whether Twitter emotions âGranger-causeâ the changes in the Bitcoin (BitStamp) market, two linear regression models were applied as shown in equations (1) and (2). The first model (1) only uses lagged values of Bitcoin market data to predict while second model includes the lagged value of Twitter emotions, which are denoted by . In the given model, we applied a lag of 1, 2 and 3.

 Yt=ct+p∑i=1βiYt−i+et (1)
 Yt=ct+p∑i=1αiYt−i+p∑i=1βiXt−i+ut (2)
 withH0=β1=β2=...=βp=0 (3)

After establishing the linear regression equations, is defined as

where and are the two sum of squares residuals of equations and and is the number of observations.

If the statistic is greater than a certain critical value for an distribution, then we reject the null hypothesis that does not Granger-cause , which means Granger-causes . As , the question whether âGranger-causesâ can be solved by checking the value of .

According to , there is no significant bivariate Granger causality correlation for (Twitter Bitcoin) with regard to all market indicators. Twitter have no lagging effect on the Bitcoin market. However, there are bivariate Granger causality correlations for (Bitcoin Twitter), which means that Bitcoin (BitStamp) market movements induce reactions on Twitter. In particular, indicates that BitStamp trading volume Granger causes signals of uncertainty within a 24 to 72-hour timeframe ( , cf. ). This could be interpreted as such that high trading volumes (on average) come along with a high number of transactions. The higher the amount of transactions, the more people on Twitter may articulate their uncertainties by expressing signals of hope, fear or worry.

It is noteworthy that Granger causality does not imply ”true causality” [Granger (2004)] and differs from ”causation” in the classical philosophical sense. For example, if both and are influenced by a common third variable with different lags, might erroneously be believed to Granger-cause .

## 6 Summary

The main research question of this paper was how far virtual emotions might influence a virtual and decentralized financial market like Bitcoin. Summarizing, we can draw following results:

1. Static intraday measurements suggest a moderate correlation of Twitter sentiments with Bitcoin close price and volume. Also, a lagged correlation analysis showed that the sum of emotional sentiments and especially negative signals positively correlate with the intraday trading volume within the last 48 hours. This can be translated as follows: When the trading volume was (and is) high, emotions fly high on Twitter. As such, Twitter may be interpreted as a place that reflects the ”speculative momentum”.

2. The Granger causality analysis shows that there is no statistical significance for Twitter signals as a predictor of Bitcoin with regard to the close price, intraday spread or intraday return. To the contrary, results in indicate that the Bitcoin trading volume Granger causes signals of uncertainty within a 24 to 72-hour timeframe. Higher trading volumes Granger cause more signals of uncertainty.

3. Summarizing the results, the microblogging platform Twitter may be interpreted as a virtual trading floor that emotionally reflects Bitcoin’s market movement.

Keeping up that picture, the imagination of classic open-outcry trading floors comes to mind, where traders shouted and made use of hand signals on the pit. Measuring sound noise on stock exchanges, \citeNcoval-2001 suggests that the communication and processing of highly subtle and complex non-transaction signals (noise) by traders in such an environment plays a central role in determining equilibrium supply and demand conditions. The authors further conclude in 2001 that while trading volumes migrate to electronic exchanges, information from face-to-face interaction might be lost. Now, 13 years after their publication, we might conclude: Maybe, the noise on trading floors is back; just in a different form and space.

While the current data of only 104 days already looks promising, a longitudinal analysis of about 6 months might provide a better quality of scientific expressiveness, especially in view of the fact that we currently observe a very volatile market with an observation of 1612 tweets per day on average. Particularly, events such as the breakdown of Mt.Gox can be considered as (both internal and external) market shocks that essentially influence the considered data and statistical methods. Equal attention should be paid to data and sentiment quality, which is very limited in our current methodology. While a better linguistic might significantly improve the quality of data, emotional contagion on Twitter should also be considered as another important factor [Hu et al. (2013), Coviello et al. (2014)]. For example, a TwitterRank [Weng et al. (2010)] telling more about a user’s emotional influence and authencity might contribute to better data quality on the weight of nodes in the communication. Notable in this context is a study by \citeNhernandez-2014 of about 50,000 messages from more than 6,000 users on Twitter with focus on Bitcoin. The researcher’s analysis shows a consistent pattern that people interested in Bitcoin are far less likely to emphasize social relations than typical users of the site. Specifically, Bitcoin followers are less likely to mention emotions (beyond family, friends, religion, sex, and) and have significantly less social connection to other users on the site. If this assumption is true, it can hardly be estimated which effect it might entail for this study.

Finally, we cordially invite fellow researchers to keep a close eye on Twitter and the Bitcoin market and to improve the outlined approach.

### Footnotes

1. Last Update: September 2014. An early version of this paper was presented at Collective Intelligence 2014, MIT, Cambridge, USA, June 10-12 2014.
2. The authors likes to thank Peter Gloor for his feedback on an early version of this paper.
3. For an explanation of the term ’mining’ see \citeNnyt-2013 or \citeNkrolldaveyfelten-2013
4. http://Bitcoincharts.com, March 31st, 2014.
5. N=104
6. http://tweetarchivist.com
7. Together, BitStamp (34%), Bitfinex (26%), BTC-e (16%) and BTC China (10%) account for about 86% of the overall Bitcoin market volume. Cf. http://Bitcoincharts.com/charts/volumepie/ (March 31st, 2014)
8. It may be noted that the number of positive tweets is much higher than that of negative ones, more than 10 times higher on average. So far, the assumption by [Zhang et al. (2011)] that people prefer optimistic to pessimistic words can be confirmed.
9. by March 31 2014
10. As there was a trend observable for the Bitcoin market development and emotions on Twitter, an analysis with consideration of constant and trend was conducted.

### References

1. George A. Akerlof and Robert J. Shiller. 2009. Animal Spirits: How Human Psychology Drives the Economy, and Why It Matters for Global Capitalism. Princeton University Press.
2. Werner Antweiler and Murray Z. Frank. 2004. Is all that talk just noise? The information content of Internet stock message boards. Journal of Finance 59, 3 (Juna 2004), 1259–1295.
3. Hohan Bollen, Huina Mao, and Xiao-Jun Zeng. 2011. Twitter mood predicts the stock market. Journal of Computational Science 2, 1 (March 2011), 1–8.
4. Hyunyoung Choi and Hal Varian. 2012. Predicting the Present with Google Trends. Economic Record 88 (June 2012), 2–9.
5. Munmun De Choudhury, Hari Sundaram, Ajita John, and Dorée D. Seligmann. 2008. Can blog communication dynamics be correlated with stock market activity?. In In HT ’08: Proceedings of the nineteenth ACM conference on Hypertext and hypermedia (2008). 55–60.
6. Joshua D Coval and Tyler Shumway. 2001. Is sound just noise? The Journal of Finance 56, 5 (2001), 1887–1910.
7. Lorenzo Coviello, Yunkyu Sohn, Adam DI Kramer, Cameron Marlow, Massimo Franceschetti, Nicholas A Christakis, and James H Fowler. 2014. Detecting Emotional Contagion in Massive Social Networks. PloS one 9, 3 (2014), e90315.
8. David Dickey and Wayne Fuller. 1979. Distribution of the Estimators for Autoregressive Time Series With a Unit Root. J. Amer. Statist. Assoc. 74, 366 (1979), 427–431.
9. Domenico Giannone, Lucrezia Reichlin, and David Small. 2008. Nowcasting: The real-time informational content of macroeconomic data. Journal of Monetary Economics 55, 4 (2008), 665–676.
10. Eric Gilbert and Karrie Karahalios. 2010. Widespread worry and the stock market. In Proceedings of the International Conference on Weblogs and Social.
11. Peter A. Gloor, Jonas S. Krauss, Stefan Nann, Kai Fischbach, and Detlef Schoder. 2009. Web Science 2.0: Identifying Trends Through Semantic Social Network Analysis. In International Conference on Computational Science and Engineering, Vol. 4.
12. Clive WJ Granger. 1969. Investigating causal relations by econometric models and cross-spectral methods. Econometrica: Journal of the Econometric Society (1969), 424–438.
13. Clive WJ Granger. 2004. Time series analysis, cointegration, and applications. American Economic Review (2004), 421–425.
14. Ivan Hernandez, Masooda Bashir, Gahyun Jeon, and Jeremiah Bohr. 2014. Are Bitcoin Users Less Sociable? An Analysis of Users’ Language and Social Connections on Twitter. In HCI International 2014-Posters’ Extended Abstracts. Springer, 26–31.
15. Xia Hu, Lei Tang, Jiliang Tang, and Huan Liu. 2013. Exploiting social relations for sentiment analysis in microblogging. In Proceedings of the sixth ACM international conference on Web search and data mining. ACM, 537–546.
16. Alejandro Jaimes, Eduardo Ruiz, Vagelis Hristidis, Carlos Castillo, and Aristides Gionis. 2012. Correlating Financial Time Series with Micro-Blogging Data. (2012).
17. Joshua A. Kroll, Ian C. Davey, and Edward W. Felten. 2013. The Economics of Bitcoin Mining, The Economics of Bitcoin Mining, or Bitcoin in the Presence of Adversaries. (November 2013).
18. Satoshi Nakamoto. 2009. Bitcoin: A Peer-to-Peer Electronic Cash System. (2009).
19. New York Times. 2013. Into the Bitcoin Mines. (2013).
20. Chong Oh and Olivia Sheng. 2011. Investigating Predictive Power of Stock Micro Blog Sentiment in Forecasting Future Stock Price Directional Movement. Proceedings of the International Conference on Information Systems, ICIS 2011 (January 2011).
21. Rick Scott and George Loewenstein. 2010. Handbook of Emotions. Guilford Press, Chapter The Role of Emotion in Economic Behaviour, 138–158.
22. Jianfeng Si, Arjun Mukherjee, Bing Liu, Qing Li, Huayi Li, and Xiaotie Deng. 2013. Exploiting Topic based Twitter Sentiment for Stock Prediction. Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, ACL 2013 2 (2013), 24–29.
23. Timm O. Sprenger, Andranik Tumasjan, Philipp G. Sandner, and Isabell M. Welpe. 2013. Tweets and trades: The information content of stock microblogs. European Financial Management (May 2013).
24. The Economist. 2013. Bitcoin under pressure. (November 2013).
25. The Wall Street Journal. 2013. Beware the Risks of the Bitcoin: Winklevii Outline the Downside. (July 2013).
26. Jianshu Weng, Ee-Peng Lim, Jing Jiang, and Qi He. 2010. Twitterrank: finding topic-sensitive influential twitterers. In Proceedings of the third ACM international conference on Web search and data mining. ACM, 261–270.
27. Shaozhi Ye and Felix S. Wu. 2010. Measuring Message Propagation and Social Influence on Twitter.com. In Proceedings of the Second international conference on Social informatics. 216–231.
28. Xue Zhang, Hauke Fuehres, and Peter A. Gloor. 2011. Predicting Stock Market Indicators Through Twitter “I hope it is not as bad as I fear”. Procedia - Social and Behavioral Sciences 26 (2011), 55–62.
You are adding the first comment!
How to quickly get a good reply:
• Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
• Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
• Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters