Predicting bitcoin price movements using sentiment analysis: a machine learning approach

Purpose Cryptocurrencies such as Bitcoin (BTC) attracted a lot of attention in recent months due to their unprecedented price fluctuations. This paper aims to propose a new method for predicting the direction of BTC price using linear discriminant analysis (LDA) together with sentiment analysis. Design/methodology/approach Concretely, the authors train an LDA-based classifier that uses the current BTC price information and BTC news announcements headlines to forecast the next-day direction of BTC prices. The authors compare the results with a Support Vector Machine (SVM) model and random guess approach. The use of BTC price information and news announcements related to crypto enables us to value the importance of these different sources and types of information. Findings Relative to the LDA results, the SVM model was more accurate in predicting BTC next day’s price movement. All models yielded better forecasts of an increase in tomorrow’s BTC price compared to forecasting a decrease in the crypto price. The inclusion of news sentiment resulted in the highest forecast accuracy of 0.585 on the test data, which is superior to a random guess. The LDA (SVM) model with asset specific (news sentiment and asset specific) input features ranked first within their respective model classifiers, suggesting both BTC news sentiment and asset specific are prized factors in predicting tomorrow’s price direction. Originality/value To the best of the authors’ knowledge, this is the first study to analyze the potential effect of crypto-related sentiment and BTC specific news on BTC’s price using LDA and sentiment analysis.

Impact of news sentiment and topics on IPO underpricing: US evidence

Purpose The goal of the study is to examine the effects of news sentiment and topics dominating in the news field prior to the initial public offering (IPO) on the IPO underpricing. Design/methodology/approach The authors’ approach has several steps. The first is textual analysis. To detect the dominating topics in the news, the authors use Latent Dirichlet allocation. The authors use bidirectional encoder representations from transformers (BERT) pretrained on financial news corpus to evaluate the tonality of articles. The second is evaluation of feature importance. To this end, a linear regression with robust estimators and Classification and Regression Tree and Random Forest are used. The third is data. The text data consists of 345,731 news articles from Thomson Reuters related to the USA in the date range from 1 January 2011 to 31 May 2018. The data contains all the possible topics from the website, excluding anything related to sports. The sample of 386 initial public offerings completed in the USA from 1 January 2011 to 31 May 2018 was collected from Bloomberg Database. Findings The authors found that sentiment of the media regarding the companies going public influences IPO underpricing. Some topics, namely, the climate change and environmental policies and the trade war between the US and China, have influence on IPO underpricing if they appear in the media prior to the IPO day. Originality/value The puzzle of IPO underpricing is studied from the point of Narrative Economics theory for the first time. While most of the works cover only some specific news segment, we use Thomson Reuters news aggregator, which uses such sources The New York Post, CNN, Fox, Atlantic, The Washington Post ? Buzzfeed. To evaluate the sentiment of the articles, a state-of-the-art approach BERT is used. The hypothesis that some common narratives or topics in the public discussion may impose influence on such example of biased behaviour like IPO underpricing has also found confirmation.

Combined deep learning classifiers for stock market prediction: integrating stock price and news sentiments

PurposeStock market forecasters are focusing to create a positive approach for predicting the stock price. The fundamental principle of an effective stock market prediction is not only to produce the maximum outcomes but also to reduce the unreliable stock price estimate. In the stock market, sentiment analysis enables people for making educated decisions regarding the investment in a business. Moreover, the stock analysis identifies the business of an organization or a company. In fact, the prediction of stock prices is more complex due to high volatile nature that varies a large range of investor sentiment, economic and political factors, changes in leadership and other factors. This prediction often becomes ineffective, while considering only the historical data or textural information. Attempts are made to make the prediction more precise with the news sentiment along with the stock price information.Design/methodology/approachThis paper introduces a prediction framework via sentiment analysis. Thereby, the stock data and news sentiment data are also considered. From the stock data, technical indicator-based features like moving average convergence divergence (MACD), relative strength index (RSI) and moving average (MA) are extracted. At the same time, the news data are processed to determine the sentiments by certain processes like (1) pre-processing, where keyword extraction and sentiment categorization process takes place; (2) keyword extraction, where WordNet and sentiment categorization process is done; (3) feature extraction, where Proposed holoentropy based features is extracted. (4) Classification, deep neural network is used that returns the sentiment output. To make the system more accurate on predicting the sentiment, the training of NN is carried out by self-improved whale optimization algorithm (SIWOA). Finally, optimized deep belief network (DBN) is used to predict the stock that considers the features of stock data and sentiment results from news data. Here, the weights of DBN are tuned by the new SIWOA.FindingsThe performance of the adopted scheme is computed over the existing models in terms of certain measures. The stock dataset includes two companies such as Reliance Communications and Relaxo Footwear. In addition, each company consists of three datasets (a) in daily option, set start day 1-1-2019 and end day 1-12-2020, (b) in monthly option, set start Jan 2000 and end Dec 2020 and (c) in yearly option, set year 2000. Moreover, the adopted NN + DBN + SIWOA model was computed over the traditional classifiers like LSTM, NN + RF, NN + MLP and NN + SVM; also, it was compared over the existing optimization algorithms like NN + DBN + MFO, NN + DBN + CSA, NN + DBN + WOA and NN + DBN + PSO, correspondingly. Further, the performance was calculated based on the learning percentage that ranges from 60, 70, 80 and 90 in terms of certain measures like MAE, MSE and RMSE for six datasets. On observing the graph, the MAE of the adopted NN + DBN + SIWOA model was 91.67, 80, 91.11 and 93.33% superior to the existing classifiers like LSTM, NN + RF, NN + MLP and NN + SVM, respectively for dataset 1. The proposed NN + DBN + SIWOA method holds minimum MAE value of (∼0.21) at learning percentage 80 for dataset 1; whereas, the traditional models holds the value for NN + DBN + CSA (∼1.20), NN + DBN + MFO (∼1.21), NN + DBN + PSO (∼0.23) and NN + DBN + WOA (∼0.25), respectively. From the table, it was clear that the RMSRE of the proposed NN + DBN + SIWOA model was 3.14, 1.08, 1.38 and 15.28% better than the existing classifiers like LSTM, NN + RF, NN + MLP and NN + SVM, respectively, for dataset 6. In addition, he MSE of the adopted NN + DBN + SIWOA method attain lower values (∼54944.41) for dataset 2 than other existing schemes like NN + DBN + CSA(∼9.43), NN + DBN + MFO (∼56728.68), NN + DBN + PSO (∼2.95) and NN + DBN + WOA (∼56767.88), respectively.Originality/valueThis paper has introduced a prediction framework via sentiment analysis. Thereby, along with the stock data and news sentiment data were also considered. From the stock data, technical indicator based features like MACD, RSI and MA are extracted. Therefore, the proposed work was said to be much appropriate for stock market prediction.

IMPACT OF COVID-19 ON CRYPTOCURRENCIES: EVIDENCE ON INFORMATION TRANSMISSION THROUGH ECONOMIC AND FINANCIAL MARKET SENTIMENTS

This paper investigates the relationship between the COVID-19 crisis and the two leading cryptocurrencies, Bitcoin and Ethereum, from 31 December 2019 to 18 August 2020. We also use an economic news sentiment index and financial market sentiment index to explore the possible mechanisms through which COVID-19 impacts cryptocurrency. We employ a VAR Granger Causality framework and Wavelet Coherence Analysis and find the cryptocurrency market was impacted in the early phase of the sample period through economic news and financial market sentiments, but this effect diminished after June 2020.