Sentiment Analyis and Bitcoin Price Prediction

doi:10.21203/rs.3.rs-3988264/v1

Download PDF

Research Article

Sentiment Analyis and Bitcoin Price Prediction

https://doi.org/10.21203/rs.3.rs-3988264/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The emergence of Bitcoin as a decentralized digital currency has underscored the importance of developing advanced techniques for predicting its price fluctuations. This study evaluates the predictive power of Bitcoin-related Google search volumes and Twitter sentiment analysis within short time frames. By leveraging machine learning algorithms and opinion mining, we identify correlations between online behaviors and Bitcoin price movements. Our methodology encompasses data sourcing, preprocessing, exploratory analysis, feature selection using Correlation Analysis, F-regression, Shapley values, and price prediction with a Long Short-Term Memory (LSTM) model. Findings reveal that Google search data, compared to Twitter sentiment, significantly enhances model accuracy and reduces prediction errors. The study suggests future research to investigate other search engines and online news sentiment, acknowledging limitations in data quality and accessibility of historical Twitter data.

Google web search data

Tweet Sentiment

Bitcoin Price Prediction

Shapley values

Long Short-Term Memory (LSTM)

Bitcoin, heralded as the first digital currency of its kind, leverages a decentralized, peer-to-peer system underpinned by blockchain technology. This innovative framework not only facilitates transactions without traditional financial intermediaries but also offers a means to democratize financial services, significantly reducing transaction costs and accelerating the remittance process [1]. As of January 25th, 2024, Bitcoin boasts a formidable market capitalization of approximately $777 billion, a testament to its widespread adoption and the inherent volatility that characterizes its price movements[2]. The burgeoning interest in Bitcoin and the broader cryptocurrency market has coincided with a surge in the application of machine learning (ML) techniques for price prediction. This research endeavor seeks to delve into the intricacies of forecasting Bitcoin's closing price by examining a triad of influential factors: historical price data, Google search trends, and the sentiment pervading Twitter discussions. The integration of machine learning offers a nuanced approach to dissecting the temporal and sentiment-driven dynamics that influence Bitcoin's valuation. The academic discourse surrounding cryptocurrencies and blockchain technology provides a rich backdrop for this study. Limba et al.[1] offer a theoretical exploration of these technologies, shedding light on their socio-legal implications and the transformative potential they hold for the financial sector. The discourse extends to the analysis of Bitcoin's inception, anchored in the Nakamoto payment model, and its recognition by the World Economic Forum as a pivotal emerging technology[1]

In the realm of predictive analytics, Albariqi and Winarko [3] and Mittal et al. [4] have underscored the utility of neural network models, including the Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) models, in capturing the nuanced patterns of Bitcoin price movements. These studies highlight the variable performance of ML models across different time horizons and the intriguing role of sentiment data derived from social media in refining predictive models [3, 4].

Building on this foundation, our study introduces a comprehensive data mining framework that traverses the entire spectrum of predictive modeling, from the initial problem definition to the intricacies of modeling. This framework not only encapsulates trend analysis and opinion mining but also emphasizes the criticality of feature selection and model evaluation in the context of time-series data. Our research endeavors to enrich the existing literature by providing a detailed examination of the predictive power of Google search trends and Twitter sentiment in the context of Bitcoin price fluctuations. By employing a combination of traditional and innovative feature selection methodologies, such as Correlation Analysis, F-regression, and the application of Shapley values, this study aims to uncover novel insights into the factors driving Bitcoin's market dynamics [5–7].

Research problem

The research problem addressed in this study revolves around understanding the degree to which sentiment and web search data can influence Bitcoin's market prices, specifically in predicting time series metrics like the daily closing price. The investigation extends to examining the associations between various Bitcoin pricing indicators (Open, Close, High, Low) and enriched features developed through natural language processing, sentiment analysis, and analysis of web search trends. This inquiry aims to uncover the potential predictive power of combining traditional financial indicators with novel data sources from online sentiment and search behaviors, thereby advancing the methodologies employed in financial time series forecasting within the context of cryptocurrency markets.

Dataset Description

The dataset for this investigation is sourced from publicly accessible databases and Application Programming Interfaces (APIs) that allow for seamless integration with Python programming.

a. Bitcoin Price Data

The Bitcoin market data was obtained through the Historic-Crypto library, which is an open-source tool facilitating interaction with the CoinBase Pro API. This dataset encompasses hourly data points for Bitcoin, including the Open, High, Low, Close, and Volume metrics, spanning from February 5th, 2021, to June 20th, 2021.

b. Twitter Data

The Twitter dataset was compiled from Kaggle.com, utilizing the opendatasets library for initial access, and subsequently employing the Tweepy Python library to gather tweets tagged with #bitcoin and #BTC. The dataset features include user-related information (username, location, description, account creation date, follower and friend counts, favorites) and tweet-specific details (timestamp, hashtags, platform used, retweet flag). For the purpose of this study, only the date and textual content of tweets within the specified period (February 5th, 2021, to June 20th, 2021) were analyzed and modeled. The tweets were recorded at 1-minute intervals.

c. Google Search Trends

Data on Google search trends related to Bitcoin and BTC were extracted using the pytrends API, a Python interface. This involved querying the Historical Hourly Interest feature for the aforementioned keywords, covering the same period as the Bitcoin and Twitter datasets (February 5th, 2021, to June 20th, 2021).

Data preprocessing and transformation

The framework operates through an iterative process comprising eight key phases: Problem Definition, Data Acquisition, Preprocessing and Transformation of Data, Exploratory Data Analysis, Correlation Analysis, Feature Selection, Model Development, and Model Optimization (Fig. 1). In the data preprocessing and transformation phase, the initial step involved cleansing the tweet text data to ensure a pristine dataset. This fundamental cleansing removed irrelevant characters such as URLs, hashtags (e.g., #BTC, #bitcoin), user mentions, and punctuation, employing regular expressions for efficient automation [8]. Subsequently, morphological analysis was conducted using the nltk library, which facilitated the removal of stopwords—common but insignificant words like 'the', 'is', and 'a'—that could potentially obscure the analytical clarity of the text data [8]. Additionally, this phase included word tokenization, breaking down sentences into individual words or tokens, and lemmatization, a process of reducing words to their base or dictionary form to ensure consistency in the text data [9].

Feature engineering played a pivotal role in enhancing the dataset by extracting predictive features from the raw data [10]. This included the computation of tweet volume, which aggregated the number of tweets per hour, providing a quantitative measure of Twitter activity related to Bitcoin. Sentiment analysis of the tweets was performed to gauge public opinion towards Bitcoin, categorizing sentiments into positive, negative, and neutral. This was achieved using the Textblob library for basic polarity and subjectivity analysis, and the Vader scoring system for a more nuanced sentiment analysis that could interpret slang and emoticons commonly found in social media texts[11, 12]

To align the various data sources for analysis, data merging was undertaken, standardizing the sentiment data to hourly intervals to match the granularity of the Bitcoin price and Google Trends data. This consolidation used the date as a primary key for merging datasets.

Correlation analysis was conducted to understand the relationships between variables, using the Pearson correlation coefficient formula to quantify the strength and direction of these relationships [7]. Feature selection was then applied to refine the model by identifying and retaining only those features with significant predictive power, thereby enhancing computational efficiency and reducing the risk of overfitting[5]. Various feature selection techniques, including filter and wrapper methods, were explored to determine the most effective features for predicting Bitcoin prices. The F-test, a statistical filtering approach, was particularly noteworthy for its ability to rank features based on their individual contributions to the model's performance, using F-statistics and p-values to highlight significant attributes [5].The SelectKBest function ranks the F-score from highest to lowest. It returns the most significant features, ranked from the highest correlated to the least correlated features to the target variable.

SHAP (Shapley Additive exPlanations) values emerge from game theory and offer a method to attribute the contribution of each feature in a predictive model to the overall prediction outcome [6]. Originating from the work of Lloyd Shapley, these values address the fundamental question: "How to fairly distribute the 'payout' (or contribution to the outcome) among the 'players' (or features) in a coalition (or model)?" [6]. The fairness criterion is assessed by considering all possible combinations or coalitions of features both with and without a particular feature, to determine its marginal contribution to the prediction. In practice, this involves systematically omitting each feature from the model, measuring the change in prediction accuracy, and then averaging this impact over all possible feature combinations. This average marginal contribution of a feature to the prediction accuracy of the model is its Shapley value [6].

For machine learning models, SHAP values offer a granular explanation of how each feature contributes to individual predictions, thereby enhancing model transparency and interpretability. They quantify the additive impact of including a specific feature (Xi) in the prediction by comparing the model's performance with and without (Xi), across all possible feature subsets. The SHAP value of (Xi) is thus the weighted average of the differences in prediction outcomes with and without (Xi), reflecting its contribution to the model's predictive power[6]. This methodological approach not only enhances our understanding of model behavior but also aids in identifying features that significantly influence model predictions, contributing to more informed decision-making and model improvement.

In Python, the SHAP library implements Tree SHAP, specifically designed for decision tree-based models, including ensemble techniques like Random Forest and Gradient Boosted Trees. Tree SHAP optimizes the computation of Shapley values by focusing on a subset of features and fitting the data to produce approximations of these values, thus offering a practical solution to the otherwise computationally demanding task of calculating exact Shapley values across all feature permutations [13].

Feature Selection Algorithm

The feature selection process utilizes tree-based algorithms, notably CART (Classification and Regression Trees). In this context, the Random Forest model, an ensemble approach, aggregates predictions from multiple decision trees to enhance prediction accuracy and reliability [12, 14]

Modeling with LSTM

LSTM (Long Short-Term Memory) networks, an advanced variant of Recurrent Neural Networks (RNNs), are adept at capturing long-term dependencies in sequence data[15]. Unlike traditional ANNs and RNNs, which may struggle with long sequence data due to short-term memory limitations, LSTMs incorporate a system of gates that manage information flow, thereby mitigating these issues.

LSTMs are characterized by three types of gates:

Forget Gate: Utilizes a sigmoid function to decide which information from the cell state should be retained or discarded, with values close to 0 indicating "forget" and values near 1 signifying "retain" [16].
Input Gate: Determines updates to the cell state by combining the sigmoid function's output (which filters information) with the output of a tanh function (which scales data between − 1 and 1), effectively deciding new information to be added to the cell state [16].
Output Gate: Dictates the next hidden state by filtering the cell state through a tanh function, scaled by the sigmoid function's output, thus determining the portion of the cell state to contribute to the output [16].

These gates interact within the LSTM architecture to maintain a cell state that carries relevant information through the sequence, enabling the network to make informed predictions based on both recent and long-term past information. The LSTM model processes input (X_t), utilizes memories from the previous cell state (C_t−1), and the output from the last LSTM unit (H_t−1) to update its gates and cell state, ultimately producing the new cell state (C_t) and the output or next hidden state (H_t).

Activation functions within the LSTM gates, specifically the sigmoid (sig) and tanh functions, play crucial roles in regulating the information flow, ensuring that the LSTM can learn and remember over long sequences effectively (Fig. 2) [17].

The feature set prepared for analysis encompasses a diverse array of data points, integrating sentiment metrics derived from Twitter data—such as subjectivity, polarity, and Vader scores—with tweet volume, key Bitcoin trading figures (including low, high, open, and close prices), trading volume, and hourly Google search trends related to Bitcoin. A comprehensive descriptive statistical analysis was conducted to summarize the characteristics of these features, offering insights into their distribution and central tendencies. The dataset earmarked for modeling comprises 134 observations, meticulously chosen based on the criteria of data completeness and availability. This subset corresponds to a specific timeframe, spanning from February 5th to February 10th, 2021. Table 1 presents a detailed statistical breakdown of the dataset, with a particular focus on the target variable, 'close'—the closing price of Bitcoin during this period. The descriptive analysis reveals a maximum closing price of $48,192, a minimum of $37,400, and an average value of $41,996, providing a snapshot of Bitcoin's price dynamics within the selected timeframe.

Table 1

Summary Statistics
	Compound	Neg	Neu	pos	Subjectivity	Polarity	Tweet_vol	Low	High	Open	Close	Volume	Google_trends_btc
Mean	0.16	0.03	0.88	0.08	0.28	0.10	160.63	41591.81	42315.98	41939.68	41996	1259.99	46.64
SD	0.04	0.009	0.01	0.01	0.06	0.03	96.67	3488.22	3657.76	3586.34	3572.07	1116.64	23.19
Min	0.01	0.00	0.80	0.04	0.09	0.03	11.00	37239.42	37676.38	37278.08	37400.68	299.67	24.00
25%	0.12	0.02	0.87	0.07	0.25	0.08	109.25	38477.49	39013.38	38758.07	38768.67	641.44	29.00
50%	0.15	0.03	0.88	0.08	0.27	0.09	144.00	39989.52	40499.99	40155.23	40198.09	982.93	36.00
75%	0.18	0.03	0.88	0.09	0.29	0.29	180.75	45528.17	46553.85	46002.85	46002.84	1444.08	64.00
max	0.31	0.089	0.93	0.12	0.53	0.27	920.00	47487.35	48200.00	48192.15	48192.15	9348.67	124.00

Data Visualization

Utilizing Data subsets for visualization by grouping Bitcoin prices and volume, google search and tweet volume, and subjectivity and polarity for plotting. The Bitcoin prices (open, close, high, and low) follow a linear relationship with a multimodal distribution due to the presence of more than a singular peak, as shown in Fig. 2a. In addition, the Bitcoin volume distribution shows a right skew with an outlier point at 9,348. Subjectivity and Polarity data trend along the same line. Polarity data is more left-skewed than subjectivity data. Polarity data shows mild positivity with values greater than 0, and with most subjectivity, data less than 0.5 indicates overall opinions and attitudes are more subjective than objective information. Google trends data and tweet volume is left-skewed and show a nonlinear relationship between variables., as seen in Fig. 2b and 2c.

Trend Analysis

The trend analysis reveals across google searches, tweet volume, and Bitcoin volume an upward spike on February 8th, 2021, which trends upwards with bitcoin prices. With online news search, this spike reflects the news release of Telsa's purchase of $1.5 billion in bitcoin and its acceptance as a form of payment for its offerings[18].

Correlation Analysis and FRegression

The Pearson correlation coefficient of that data shows a strong correlation among the Bitcoin price attributes between 0.99 and 1, and google trends data show a positive correlation of 0.76. In contrast, tweet volume shows a low correlation of 0.3 concerning the close price. Overall sentiment data, as seen in Fig. 2.2 attributes, reveal a weak correlation concerning close price, which is on par with Ji[7] results, as shown in Fig. 3a. As seen in Fig. 3b, the f regression computation output reveals high, low, open, close, and google trends and tweet volume as the top five predictors.

Shapley values

As seen in Fig. 4a coupled with the Random Forest, the output of the Shapley values algorithm reveals the top predictors as follows high, google trend, open, low. Examining the prediction of the first observation based on Shapley values and Random Forest operation. The output value f(x) represents the prediction for the close price at 38,805.90. The base value, which is the mean value across the target values y, the features depicted with blue, push the value lower from the base value. Overall high, google trends search, open and close impact the outcome prediction, As seen in Fig. 4b. Given the base value of $41,990, the individual features are lower than their respective mean values, as seen in Table 3, thus pushing the prediction to the left from the base value.

Table 3

Average values of variables in the test set
Variable	Compound	neg	neu	pos	subjectivity	polarity	tweet_vol	low	high	open	volume	Google_trends_btc
Mean	0.16	0.03	0.88	0.08	0.28	0.10	167.08	41677.38	42428.98	42030.15	1346.79	47.97

The LSTM model was trained using an 80% split on train data and a 20% split on test data. In addition, hyperparameters such as units, batch size, and epochs were set for the model (For Base model: Units (5), Batch Size (6), and Epochs (37); whereas for base model w/ google trends: Units (5), Batch Size (6), and Epochs (29)). The LSTM model was used to test the impact of the addition of the google trends features alongside base features: close, high, low, and open in model performance. Additionally, the time series dataset for the features was framed to a pair of input and target sequences with five-time steps across all scaled features for modeling. The modeling results for the base model and base model features, including google trends, yield 539.7 and 421.7 for the Root mean square error and R² squared values of 0.604 and 0.758, respectively.

The base model and updated model produce a good fit with a slight gap between the training and test loss curves. The training and test loss decreased and stabilized to the same level, indicating a good fit for the data model. The actual and predicted time series for the base and base models with google trends data are depicted in Figs. 5a and 5b.

The detailed analysis conducted in this study, incorporating correlation analysis, F-statistic tests, and Shapley values, has yielded insightful findings regarding the predictive power of various features on Bitcoin prices. Consistently across these analytical methods, tweet sentiment metrics—specifically subjectivity and polarity—were ranked lower in importance compared to Google search trends and tweet volume. This suggests a more pronounced influence of web search behaviors over sentiment indicators in forecasting Bitcoin price movements, a conclusion that resonates with the findings of seminal works by Mittal et al.[4] and Albariqi and Winarko [3]. The robustness of search data as a predictor underscores the significant role of public interest and internet search trends in the cryptocurrency market dynamics. The efficacy of correlation analysis, F-statistic tests, and Shapley values as tools for feature selection was further demonstrated through their consistent ranking of the top predictive features. Notably, Shapley values provided a nuanced perspective by highlighting Google search trends as a particularly influential predictor for Bitcoin's closing price, surpassing even traditional price metrics such as high, low, and open values in predictive importance. This distinction underscores the unique contribution of web search data to understanding market sentiments and its potential impact on price.

Incorporating Google Trends data into the predictive model marked a significant improvement, enhancing the model's fit by 15% compared to the baseline and reducing the RMSE by $118. This improvement illustrates the substantial value of integrating time-series historical price data with correlated external features, thereby enriching the model's predictive capacity. The approach adopted in this study, which extends beyond the reliance on previous day's price data to include additional features, has demonstrated a marked improvement in model performance on test datasets. The strategic application of early stopping techniques further refined the modeling process by identifying an optimal number of epochs, effectively curbing the risk of overfitting and ensuring model robustness.

The overarching goal of this research was to explore the predictive utility of auxiliary features such as sentiment data and Google search trends, in conjunction with traditional Bitcoin price indicators (close, high, low, open), in forecasting the closing price of Bitcoin. The quantitative analysis conducted for feature selection has affirmed the value of a selective approach, leveraging correlation analysis, F-statistic tests, and Shapley values for a more targeted and efficient feature inclusion strategy. This approach not only enhances model performance but also optimizes computational resources. The findings unequivocally demonstrate the significant impact of Google search trends on model accuracy, advocating for their incorporation into predictive models for Bitcoin trading. This suggests a broader applicability of the approach, potentially extending to other cryptocurrencies such as Ethereum, and offering valuable insights for traders and analysts in the cryptocurrency domain. The study thus contributes to the evolving understanding of market dynamics in the cryptocurrency sector, highlighting the interplay between internet search behaviors, public sentiment, and market prices.

The research aimed to assess the impact of additional features such as sentiment data and Google Trends search data, alongside traditional Bitcoin price indicators (close, high, low, open), on predicting the closing price of Bitcoin.

For a more comprehensive understanding of these findings, it would be beneficial to expand the dataset for tweet sentiment analysis to enhance model optimization further. Currently, the Twitter API limits access to only seven days of historical data, requiring a developer license for extended access, thus presenting a challenge in data availability. Alternatively, publicly available datasets (e.g., from Kaggle.com) need to be thoroughly vetted for completeness prior to their use in analysis and modeling. Future research could also explore search data related to Bitcoin and BTC on other search engines like Bing, Baidu, and Duck Duck Go.

Moreover, analyzing top news websites for Bitcoin-related headlines, by scraping comments and assessing the volume of likes, shares, and comments, could provide additional insights for Bitcoin price prediction. Ethical considerations in future studies analyzing tweet data should prioritize the privacy of user IDs during the publication and discussion of sentiment analysis findings. The primary advantage of accessing Twitter data lies in advancing machine learning techniques to enhance the analysis of tweet volume and sentiment data, rather than infringing on users' freedom of speech. A significant limitation of this study is the emphasis on data quality for training machine learning models over hyperparameter optimization. After excluding text input features to improve data quality, only the first 137 observations were deemed suitable for a thorough and effective analysis. Future research should investigate whether different or longer time frames might yield divergent outcomes with respect to tweet sentiment, Google Trends data, and historical price data. Additionally, this study was confined to a brief time window, with the 137 hours under review characterized by a specific event—the announcement by Elon Musk regarding Tesla—which might not be representative. Extending the model to cover longer periods with greater market volatility could offer new perspectives and opportunities for optimizing model parameters.

Author Contribution

All authors reviewed the manuscript

Limba, T., Stankevičius, A., Andrulevičius, A.: Cryptocurrency as disruptive technology: theoretical insights. Entrepreneurship and Sustainability Issues. 6, 2068–2080 (2019). https://doi.org/10.9770/jesi.2019.6.4(36).
Bitcoin Market Cap (I:BMC).
Albariqi, R., Winarko, E.: Prediction of Bitcoin Price Change using Neural Networks. In: 2020 International Conference on Smart Technology and Applications (ICoSTA). pp. 1–4 (2020). https://doi.org/10.1109/ICoSTA48221.2020.1570610936.
Mittal, A., Dhiman, V., Singh, A., Prakash, C.: Short-Term Bitcoin Price Fluctuation Prediction Using Social Media and Web Search Data. In: 2019 Twelfth International Conference on Contemporary Computing (IC3). pp. 1–6 (2019). https://doi.org/10.1109/IC3.2019.8844899.
Pirbazari, A.M., Chakravorty, A., Rong, C.: Evaluating Feature Selection Methods for Short-Term Load Forecasting. In: 2019 IEEE International Conference on Big Data and Smart Computing (BigComp). pp. 1–8 (2019). https://doi.org/10.1109/BIGCOMP.2019.8679188.
Lundberg, S.M., Lee, S.-I.: A unified approach to interpreting model predictions. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. pp. 4768–4777. Curran Associates Inc., Red Hook, NY, USA (2017).
Ji, S.-H., Baek, U.-J., Shin, M.-G., Goo, Y.-H., Park, J.-S., Kim, M.-S.: Best Feature Selection using Correlation Analysis for Prediction of Bitcoin Transaction Count. In: 2019 20th Asia-Pacific Network Operations and Management Symposium (APNOMS). pp. 1–6 (2019). https://doi.org/10.23919/APNOMS.2019.8892896.
’Angiani, G., ’Ferrari, L., ’Fontanini, T., ’Fornacciari, P., ’Iotti, E., ’Magliani, F., ’Manicardi, S.’: A Comparison between Preprocessing Techniques for Sentiment Analysis in Twitter. In: Kdweb 2016. , Cagliari (2016).
Nguyen, G., Dlugolinsky, S., Bobák, M., Tran, V., López García, Á., Heredia, I., Malík, P., Hluchý, L.: Machine Learning and Deep Learning frameworks and libraries for large-scale data mining: a survey. Artif Intell Rev. 52, 77–124 (2019). https://doi.org/10.1007/s10462-018-09679-z.
Jiang, W.: Applications of deep learning in stock market prediction: Recent progress. Expert Syst Appl. 184, 115537 (2021). https://doi.org/https://doi.org/10.1016/j.eswa.2021.115537.
Mohapatra, S., Ahmed, N., Alencar, P.: KryptoOracle: A Real-Time Cryptocurrency Price Prediction Platform Using Twitter Sentiments. In: 2019 IEEE International Conference on Big Data (Big Data). pp. 5544–5551. IEEE (2019). https://doi.org/10.1109/BigData47090.2019.9006554.
Ibrahim, A.: Forecasting the Early Market Movement in Bitcoin Using Twitter’s Sentiment Analysis: An Ensemble-based Prediction Model. In: 2021 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS). pp. 1–5 (2021). https://doi.org/10.1109/IEMTRONICS52119.2021.9422647.
’Tan, T.: Back to Basics: Assumptions of Common Machine Learning Models.
Ray, S.: A Quick Review of Machine Learning Algorithms. In: 2019 International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COMITCon). pp. 35–39 (2019). https://doi.org/10.1109/COMITCon.2019.8862451.
Siami-Namini, S., Tavakoli, N., Siami Namin, A.: A Comparison of ARIMA and LSTM in Forecasting Time Series. In: 2018 17th IEEE International Conference on Machine Learning and Applications (ICMLA). pp. 1394–1401 (2018). https://doi.org/10.1109/ICMLA.2018.00227.
’Phi, M.’: Illustrated Guide to LSTM’s and GRU’s: A step by step explanation.
Feng, W., Guan, N., Li, Y., Zhang, X., Luo, Z.: Audio visual speech recognition with multimodal recurrent neural networks. In: 2017 International Joint Conference on Neural Networks (IJCNN). pp. 681–688. IEEE (2017). https://doi.org/10.1109/IJCNN.2017.7965918.
“Stevekovach”: Tesla buys $1.5 billion in bitcoin, plans to accept it as payment.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Sentiment Analyis and Bitcoin Price Prediction

Status:

Version 1

Abstract

Figures

Introduction

Research problem

Methods

Dataset Description

a. Bitcoin Price Data

b. Twitter Data

c. Google Search Trends

Data preprocessing and transformation

Feature Selection Algorithm

Modeling with LSTM

Results

Data Visualization

Trend Analysis

Correlation Analysis and FRegression

Shapley values

Discussion

Conclusion

Declarations

Author Contribution

References

Additional Declarations

Status:

Version 1