Analysis of relationships between tweets and stock market trends

1 Introduction

With the advances in natural language processing and machine learning techniques the task of forecasting the stock market using textual information has grown popularity as more data becomes available. The first works [14, 43] used news as a source of information with promising results . On recent years the works [4, 45] use the social media provided its growth in popularity. Social media embeds not only news but also stock market events, opinions and insights from investors.

We collected Twitter data to create a corpus of messages related to nine IT companies. We choose Twitter because it is a popular platform and also because the messages or tweets are tagged with the company’s stock symbol by the users themselves.

In this study we analyze the tweets generated during different trends, our hypothesis is that it is possible to identify tweets depending on the positive, negative or neutral market trends. We used two different text representations, Bag of Words (BOW) and Word Embeddings (WE), for the latter word2vec, a model trained on news, is used. Previous works compare textual representations in the task of sentiment analysis using a manually tagged set, however we tag a tweet automatically by the trend in which it appeared. Our results are contrasted against a baseline that considers the distribution of the corpus. Not many related works consider how the distribution of the data may influence the results [22].

We also study whether tweet volume is correlated with the transaction volume or returns indicators and whether it can be helpful on the task of forecasting. We consider also negative relationships using a measure of local trend association [2]. Previous studies do not consider the trading hours, which are important since after the market closes new topics are generated that may not be related with the previous price but with next day’s price.

The rest of the paper is structured as follows. In Section 2, we present the related works. In Section 3, we describe the corpus. In Sections 3.1 and 3.2, we present experiments with tweet volume and in Section 3.3, experiments with the content of the tweets. Section 4 includes conclusions and future work.

2 Related work

There are two main paradigms in forecasting the stock market: technical [18] and fundamental [1] analysis. The first approach analyzes the historical prices of the time series themselves using methods such as auto-regression, ARIMA, state-space models, neural networks and others. The second approach relies on financial indicators such as the overall economy, unemployment, welfare indicators, etc.

There have been numerous approaches to integrate automatically other information than in fundamental or technical analysis in the stock market prediction.

One of the first research works that investigates such strategies is done by Sankaran [36]. In his work, traders are asked to assign weights to different attributes they call “unmeasurable factors”, including government policies or political news that may influence the exchange rates. In this approach the knowledge of experts is taken from the news they read and other internal information they posses, but the news are not directly processed.

Leung [17] creates a system that requests the user to manually download news before the market opens to predict the Hang Seng Index, generating decision rules from the news articles. Later, in [43], Leung’s approach is extended by adding other global indexes such as DJIA and FTSE and improving the processing techniques. This last approach resulted in greater performance forecasting the DJIA, it was attributed to the fact that most news sources they used were from the U.S.

Lavrenko et al. [14] collect news automatically, they consider prediction as a classification problem. The text from the news is aligned with the future financial time series trend. He analyzed different companies, one of his interesting findings is that the same word may have different effects for different companies. The set up of prediction as a classification problem has been widely used in the literature [11, 47].

Lee et al. [16] create a corpus of financial reports that ranges form 2002 to 2012. They tested several machine learning methods, the method reported with the best results was Random Forest. In this article financial and text features are combined, both extracted from the reports. Twenty one financial features are used, the most helpful being the earnings surprise (the difference between expected and reported earnings for a given company). It was found that adding text features to financial features improved the accuracy from 50.1 to 55.5.

In Zhang’s dissertation [47] the corpus from [16] was used. He implemented three different neural networks architectures. Although they did not considered the financial features but only language features the accuracy he reported was better than Random Forest and Support Vector Machine classifiers. He selected only 15 from the 1500 companies in the corpus.

On recent years it is common to use information from social media, considering that it is a channel for communication changing the ways users interact with the news [15]. Social media includes information from different sources. In [13] it was found that 85% of the topics in social media are related to news.

Twitter is a micro-blogging social media platform launched in 2006. Its mission is “To give everyone the power to create and share ideas and information instantly, without barriers”. It has 313 million active users1 and any topic can be tackled in this platform including discussions related to the stock market. When a company stock is being discussed it is clearly referenced by preceding its stock symbol with the dollar sign e.g. $MSFT stands for the Microsoft’s stock symbol.

Some of the advantages of twitter are the inclusion of different sources of information and the summarization of this information, since only messages shorter than 140 characters are permitted. One of the disadvantages is that we can find noisy tweets in the form of adds or fake news. Wolfram [42] points out that his results could have been improved handling better the noise. Zhang et al. [46] only consider re-tweets with the rationale that if it is re-tweeted then it is more relevant to users.

Social media can be explored in different manners. De Choudhury et al. [8] correlate the magnitude of individual stocks’s price change with features of user interaction such as the number of comments on a post, the numbers of replies to them, the elapsed time between each comment and other several features. Another interesting approach by Ruiz et al. [35] is the representation of social media as a graph and the analysis of correlations between the features of the graph and the price and volume of company stocks. An interesting challenge in processing social media data is tackling multimodal information, which is the subject of the research field of multimodal sentiment analysis [6, 44].

A popular approach to forecast the stock market is the use if sentiment analysis techniques. This is usually done for predicting large indicators such as DJIA or S&P [4, 45], but it is also used for the stocks of individual companies [7, 38].

Tetlock [40] uses a psychosocial dictionary, the General Inquirer’s Harvard IV-4, to measure the sentiment of a Wall Street column. It was found out that pessimism causes downward trends and that either high or low pessimism can predict high trading volume. Bollen et al. [4] predicted the daily up and down changes in the closing values of the DJIA using different mood indicators such as happy, alert or calm.

The results of forecasting using sentiment analysis at company level are different from those at market level. Oliveira et al. [23] did not find predictive power for the returns even after trying with five different sentiment lexicons, nor did they find a significance difference among the lexicons. De Fortuny et al. [9], also at company level, compare BOW against sentiment analysis, they found out that the latter performed worse than the former and even worse than random. They attribute this behavior to the fact that sentiment lexicons are usually extracted from different contexts, such as book or movie reviews. Even more, in finance sometimes the nouns and verbs are more informative for decision making than the adjectives. Lee et al. [16], predicting individual company stocks, used generic and specialized sentiment lexicons, none of such approaches boosted the accuracy results.

From the evidence in the literature it seems that sentiment analysis is more useful at the global market level than at the company stocks level.

Besides returns, the transaction volume is another variable than has been studied, it can be defined as the number of transactions performed in a given period of time. Transaction volume is useful to study because in technical analysis it determines the importance of the changes in price and the risk involved in a transaction.

Ruiz et al. [35] tested correlation at different lags between volume and graph features generated from Twitter interaction. The greatest correlation was found at lag zero. Mao et al. [20] found weak correlation between price and tweets mentioning the stock symbol AAPL, and moderate correlation to the daily volume traded. Sprenger et al. [39] and Oliveira et al. [23] found that the message volume is useful to predict the next-day trading volume using regression analysis.

3 Experimental setup

The tweets were collected from March 23, 2017 to July 3, 2017 using the Twitter API tweepy [34]. We only considered tweets with a single company in the message, this operation reduced the corpus up to 25% of its original size. The rationale for this filter was the observation that tweets with more than one stock symbol mentioned were either advertisements or contained a relationship between such entities that is outside of our scope. By filtering out such tweets it is also avoided to use the same tweet for different forecasts. A total of 141’007 tweets were used in the analysis. The stock symbols chosen based on their popularity are amzn (Amazon.com, Inc.), aapl (Apple Inc.), fb (Facebook Inc.), goog (Alphabet Inc.), msft (Microsoft Corporation), snap (Snap Inc.), twtr (Twitter Inc.), yhoo (Yahoo Inc.)2 and znga (Zynga Inc.). Other stock symbols were considered but not enough tweets were collected (less than 100). The distribution of the tweets3 in the corpus is shown on Fig. 1. The average of daily tweets is shown in Table  1. Below we show a sample of the tweets.

OPEN IN VIEWER

Fig.1

Distribution of tweets through the corpus.

The daily stock prices at open and close market times and the volume were collected from Google Finance.4

BOW is a traditional approach that has been widely used in the literature, it is a subclass of n-gram representation in which a word represents one dimension in a document vector. A word is represented by the number of times that it appeared in the document.

Word embeddings is the representation of words as vectors, it is a recently proposed methodology by Mikolov [21]. We used the Word2Vec vectors5 obtained from a trained distributed representation on news with three million words, each word is represented by a 300-dimensional array. The semantics captured by this model allows operations among words, a popular example of such powerful operation of the trained model is: “king” - “man” + “woman” = “queen”.

3.1 Correlating tweets with volume and returns

It was tested a lagged correlation between the daily number of tweets per company against the company’s open price, close price, the difference between close and open prices, the absolute value of such difference, transaction volume, returns Ret t = ( p t - p t - 1 p t - 1 ) ,(1) and logarithmic returns LogRet t = log ( p t p t - 1 ) ,(2) where p _t is the price at close market time t and p_t-1 is the price at open market time t - 1.

The time series of the tweets is created with the days for which at least one tweet was collected. The days with zero tweets are not considered as point in the correlation even if it is a working day, this is because not for all companies we started downloading at the same time.

In Table 2 we show the results for |ρ|>0.5, where ρ is the Pearson’s correlation coefficient. Although we compared with all the indicators mentioned above volume was the indicator that appeared the most. The column Days indicates the total number of points of the time series, for each lag a point is lost. High correlation with a lag 1 indicates that the number of tweets may cause the price or transaction volume fluctuation of the next day. High correlation with lag -1 indicates that the price or transaction volume may cause the next day’s tweet volume.

OPEN IN VIEWER

Table 2

Higher results for lagged correlation between tweet volume and financial indicators using a lag from -3 to 3

Symbol	Indicator	Correlation	Lag	Days	Symbol	Indicator	Correlation	Lag	Days
Before trading hours					During trading hours
amzn	Volume	0.6526	0	45	snap	Volume	0.9567	0	52
fb	Open	0.5170	-1	44	snap	Volume	0.6765	1	51
fb	Open	0.5509	0	45	twtr	Volume	0.5147	0	52
fb	Close	0.5140	-1	44	yhoo	Volume	0.5955	0	50
fb	Close	0.5210	0	45	yhoo	Volume	0.5603	1	49
snap	Volume	0.9141	0	40	yhoo	Volume	0.7739	3	47
snap	Volume	0.5546	1	39	After trading hours
During trading hours					snap	Volume	0.6846	0	52
amzn	Volume	0.6375	0	70	snap	Volume	0.9175	1	51
aapl	Volume	0.6056	0	67	snap	Volume	0.5748	2	50
snap	Volume	0.5408	-1	51	znga	Volume	0.5229	0	54

The time series shown on the Figs. 1 and  2 are z-normalized using z ( x i ) = x i - μ ( X ) σ ( X ) ,(3) where x _i  ∈ X are the points in the time series X, μ (X) and σ (X) are the mean and standard deviation of X respectively.

OPEN IN VIEWER

Fig.2

(a) Correlation between Apple’s tweet volume and transaction volume lag = 0, ρ = 0.61; (b) Correlation between Amazon’s tweet volume and transaction volume lag = 0, ρ = 0.64; (c) Correlation between Snapchat’s tweet volume and transaction volume lag = 0, ρ = 0.96; (d) Correlation between Twitter’s tweet volume and transaction volume lag = 0, ρ = 0.51; (e) Correlation between Yahoo’s tweet volume and transaction volume lag = 0, ρ = 0.60; (f) Correlation between Yahoo’s tweet volume and transaction volume lag = 3, ρ = 0.77. For Figures (a)–(f), the tweet volume is measured during trading hours.

OPEN IN VIEWER

Fig.3

Associations between (a) Snapchat’s tweet volume and open price lag = 0, ρ = -0.17, LTAM = -0.74; (b) Yahoo’s tweet volume and open price lag = -2, ρ = -0.22, LTAM = -0.58; (c) Snapchat’s tweet volume and close price lag = 0, ρ = -0.19, LTAM = -0.61; (d) Yahoo’s tweet volume and open price lag = -1, ρ = 0.14, LTAM = -0.57; (e) Amazon’s tweet volume and returns lag = 1, ρ = -0.34, LTAM = -0.55; (f) Snapchat’s tweet volume and close price lag = 1, ρ = -0.19, LTAM = -0.75. For Figures (a)–(f), the window of the MAT is k = 2.

We can observe from Table 2 that tweet volume and trading volume showed more correlation at lag 0. This is consistent with [35], where the greatest correlation on average was ρ = 0.4728 at lag 0 and between tweets and trading volume. However they use the total number of tweets during the day, the daily change in close prices and daily traded volume. We also tried the total number of tweets during the day achieving similar results. Another interesting observation is the case of the company Snapchat, the result indicates that the volume of tweets the night before and the morning before are highly correlated with the transaction volume during the following trading hours.

We show that the tweet messages generated during a positive, negative and neutral trends of stock prices can be distinguished by state of the art classifiers. This means that the topics on social media depend on the price change of a company’s stock. We measure our results considering how the balance of the dataset may influence the results.

The results obtained using Word Embeddings representation of tweets significantly outperform the results obtained by Bag of Words. This may be caused by the significant reduction in the dimensionality.

For more than half of the considered companies, we found moderate correlation between tweet volume and trading volume, which is consistent with the findings in [35] and with the Efficient Market Hypothesis [19]. We showed that for two of the companies tweet volume can be used for prediction of transaction volume with lag of one or three days. Regarding the consideration of the trading hours we found for one company that the volume of tweets generated the day before after the close time is correlated with the volume during trading hours the next day.

We also used the MAP transform and the LTAM to find associations between the tweet volume different indicators. We found inverse relationships that were not discovered by correlation analysis. For this reason, LTAM can be used as complementary measure to the correlation coefficient.

As our future work, we will increase the corpus size, try different techniques on the textual representations (e.g. add dimensionality reduction for BOW model) and add semantic analysis of the tweets. We would also combine our text approaches with time series specific forecasting techniques such as regression or ARIMA, this combination has shown good results in the literature [23, 37]. We also aim to use deep learning-based approaches with hierarchical architectures to enhance scalability and increase accuracy of our method [25–29].