CRYPTO CURRENCY PRICE PREDICTION

USING BITCOIN LSTM MODEL ALONG

WITH SENTIMENT ANALYSIS
Harsh Kumar | Kevin Benjamin | Rayala Viswanath | Santosh Kumar
REALLGREATSITE

Introduction

Crypto – currency is a decentralized currency which can be something that can be traded
between individuals or groups. Investment is possible through different commercial centers
known as Bitcoin trades. These enable individuals to trade on Bitcoins utilizing various currencies.
Mt. Gox is the biggest Bitcoin exchange, where Bitcoin is stored as a virtual digital bank. The
record of the considerable number of exchanges, the timestamp information handled in this
market is called Block chain. Each record of block chain information is encrypted. Trades done by
the client's name are made private only wallet ID is made open. The only need to predict the
Crypto prices is because crypto has no governing body so any rumors or news about crypto
results in the prices variation rapidly. So, there is a need to predict the prices of the crypto based
on previous trends so that one can reduce the loss.
 Forecasting Directional Bitcoin Price Returns Using Aspect-Based
LITERATURE SURVEY - 2021

Sentiment Analysis on Online Text Data

Objective Methodology Results Gaps

The study aims to predict Bitcoin's The study analyzes Bitcoin prices Sentiment adjustments The study's reliance on Twitter
weekly price direction, using data from Investing.com increased accuracy and F1 data limits its robustness; future
determining if prices will rise or and Twitter sentiment from Score, with a 1-day tweet work could incorporate sources
fall. Kaggle, processed with regular window achieving the highest like Reddit or news articles for
expressions and classified with accuracy of 84.01% and an F1 enhanced prediction accuracy.
It compares the effectiveness of Flair. Score of 80.03%, a +2.7%
machine learning models against It employs an LSTM-based improvement over the baseline. Exploring longer investment
the traditional 'buy and hold' model incorporating Bitcoin horizons and including other
strategy. price features and profit Longer tweet windows (3-day cryptocurrencies could improve
calculations, adjusting actions and 7-day) showed smaller the model's generalizability,
The goal is to see if ML-based based on Twitter sentiment improvements with accuracy while a real-time adaptive
predictions provide better when prediction confidence is increases of +2.0% and +0.6%, threshold and larger rolling
accuracy and financial returns low. respectively, due to added noise. windows could enhance
than conventional methods. Performance is evaluated flexibility and trend capture.
through accuracy, F1 Score, Both the F-test and T-test (p <
confusion matrix, and statistical 0.05) confirmed that the
tests, with a 1-day sentiment adjusted models significantly
window providing the best improved predictions over the
results. baseline.
 Bitcoin, Sentiment Analysis and the Efficient Market Hypothesis
LITERATURE SURVEY - 2023

Objective Methodology Results Gaps

This study evaluates the The study examined Bitcoin data Best Performing model was the Logistic regression and simple
predictability of Bitcoin's weekly from September 2017 to August SVM model which gave the decision trees were ineffective
price movements to test the 2021, creating four datasets: accuracy of 67.74% during the and failed to meet
Efficient Market Hypothesis (EMH) Bitcoin's past prices, prices with testing phase. requirements.
in the cryptocurrency market. macroeconomic indicators, and This model also made no false Some models achieved high
two others with Google Trends predictions of the price accuracy but struggled with
It incorporates technical, asset- and momentum features. increases which is a very predicting price increases and
based, and sentiment-based Machine learning models like important aspect. false predictions, both crucial in
indicators, including Google Logistic Regression, SVM, While other models such as real-world applications.
Trends data, into machine learning Decision Trees, and Random Random Forest gave an accuracy Models only performed well with
models. Forests were used, with 2021 of 61.29% and the Logistic high-quality data, limiting their
data set aside for testing model regression performed poor with usability due to data variability
The research aims to determine if generalization. accuracy rate of 38.7% only and potential complexities.
sentiment analysis can enhance Model performance was Returns after investing using
trading strategies over traditional measured using sensitivity, these models were 56.05%
buy-and-hold methods. specificity, accuracy, and the whereas the buy and hold
ROC curve, comparing results strategy gave only 18.84% of
against a benchmark of returns.
unchanged Bitcoin price
direction.
 Cryptocurrency Price Prediction Using Lstm
LITERATURE SURVEY - 2023

Objective Methodology Results Gaps

This review aims to evaluate the The reviewed studies utilized ML Latent source models showed Many models lacked proper
potential of machine learning algorithms like SVM, Random strong profitability, achieving up validation, raising concerns
techniques for forecasting Bitcoin Forest, GLM, ANN, and RNN to 89% returns with a Sharpe about their reliability in various
price movements. It focuses on (including LSTM) with data from ratio of 4.1. scenarios.
analyzing different ML models, social sentiment, search trends, SVMs and ANNs, when relying
comparing data sources such as and blockchain-specific features solely on blockchain data, had Sentiment predictions were
sentiment indicators and . moderate directional accuracy affected by misinformation and
blockchain metrics, and examining Optimization methods such as around 55%. sample bias, and some models
the role of optimization strategies genetic algorithms and GPU- RNNs and LSTMs were more failed to predict upward price
and computational advancements based training were explored to effective in modeling sequential movements accurately.
in enhancing predictive accuracy. enhance model performance or time-based data, making
and efficiency. them suitable for financial The success of these models
forecasting. relied on clean, consistent data,
GPU acceleration significantly which isn't always feasible in
improved model training and real-world applications.
testing speeds, with reports of
performance being up to 80
times faster than CPU-based
training.
 Enhancing Bitcoin Price Direction Prediction Using Deep Learning
LITERATURE SURVEY - 2024

and On-Chain Data

Objective Methodology Results Gaps

To evaluate the impact of feature The dataset from Glassnode The CNN–LSTM model, Research on TCN and LSTNet in
selection methods (Boruta, GA, (2013–2023) included 87 Bitcoin enhanced with Boruta-selected cryptocurrency is limited,
LightGBM) on deep learning model metrics, with missing values features, achieved a test especially in integrating on-chain
performance. handled through listwise accuracy of 82.44%, effectively data and using advanced feature
To compare CNN–LSTM, LSTNet, deletion and regression-based capturing spatial and temporal selection methods like Boruta
and TCN models—underexplored techniques, and the task was to Bitcoin price dependencies. and Genetic Algorithms for
in this context—for Bitcoin price forecast price changes via binary Statistical tests confirmed its better model accuracy.
direction prediction. classification. superiority over other models, The relationship between model
To backtest and analyze the Feature selection was enhanced boasting an F1 Score of 81.92% performance and trading
profitability of model-based using Boruta, Genetic Algorithm and MCC of 0.6489. profitability needs more
trading strategies compared to (GA), and LightGBM, while deep The Long-and-Short Buy-and-Sell exploration, highlighting a gap in
traditional MACD-based strategies learning models like CNN–LSTM, strategy delivered exceptional evaluating real-world
LSTNet, and TCN captured data results with an Annual Return of applicability.
complexities. 6653%, a Sharpe Ratio of 1.86, There's a need to incorporate
Models were evaluated against and a Maximum Drawdown of sentiment analysis, technical
ARIMA using Accuracy, Precision, -7.04%, demonstrating effective indicators, and on-chain metrics
and AUC-ROC metrics, with real-world trading and risk for improved predictive
trading strategies simulated for management. performance, with a focus on
real-world applicability and long-term predictions in the
optimized through randomized cryptocurrency market.
search and cross-validation.
 Proposed
Work

01 Focus exclusively on 02 Utilize a database of 03 Employ the BitcoinLSTM

predicting Bitcoin price cryptocurrency influencer model with training and
movements using previous tweets (2021-2023) for trend testing sets, evaluating
day's tweets. analysis and merge with performance with MSE
historical Bitcoin price data. and R2 metrics.
System Architecture

This workflow represents our Bitcoin price prediction model. Sentiment data from social platforms is analyzed and
merged with historical market data.

The combined dataset is then fed into an LSTM model, which generates future price predictions with improved
accuracy.
REALLGREATSITE

Results
The model was able to work efficiently on the test set
providing satisfactory results. We recorded the
following metrics:

MSE R2

AI algorithms monitor systems and networks AI-powered systems can automatically

for suspicious activity, triggering alerts and isolate affected systems, prevent the spread
initiating automated responses. of malware, and mitigate the impact of
attacks.
 Future Scope
Additionally, we plan to incorporate a real-time prediction
01 Our goal is to develop this into a tool that includes a graphical
user interface (GUI), making it user-friendly and accessible for
02 feature that analyzes tweet sentiments as they are posted. This
users who prefer visual interactions over text-based capability will enable users to receive immediate insights into
commands. This interface will allow users to navigate and the mood and opinions expressed on Twitter, allowing for
utilize the tool's features effortlessly. timely and informed decision-making.
Thank You