Cryptocurrency Market Direction Prediction using Sentiment Analysis

Abstract

This project investigates the efficacy of Sentiment Analysis and Machine Learning in predicting the short-term (24-hour) price direction of major cryptocurrencies. Challenging the Efficient Market Hypothesis (EMH), this study explores whether social media sentiment, news scores, and technical indicators contain predictive signals that can be exploited by non-linear algorithms.

Completed as part of the CMP5367: Artificial Intelligence and Machine Learning module at Birmingham City University.

Team Members:

• Amogh Dath Kalasapura Arunkumar — Model Training and Evaluation
• Rajveer Singh Saini — Exploratory Data Analysis
• Harshpreet Singh — Dataset Pre-processing
• Jaspreet Kaur — Dataset Overview

---

System Architecture

Architecture Diagram

Data Flow (Mermaid)

graph TB
    subgraph "Data Layer"
        A[Raw Dataset<br/>crypto_sentiment_prediction_dataset.csv<br/>2063 Rows x 14 Columns] --> B[Pre-processed Dataset<br/>pre-processed-crypto-dataset.csv]
    end

    subgraph "Preprocessing Pipeline"
        B --> C[Data Cleaning<br/>Type Conversion and<br/>Missing Value Imputation]
        C --> D[Feature Engineering<br/>One-Hot Encoding<br/>10 cryptocurrencies]
        D --> E[Feature Scaling<br/>StandardScaler<br/>Mean=0, Std=1]
        E --> F[Outlier Detection<br/>IQR Method]
    end

    subgraph "Exploratory Data Analysis"
        B --> G[Correlation Heatmap]
        B --> H[Price vs Volume Scatter]
        B --> I[Price Change Distributions]
        B --> J[Sentiment Analysis Plots]
        B --> K[RSI and Volatility Analysis]
    end

    subgraph "Model Training"
        F --> L[Train-Test Split<br/>80/20 Stratified]
        L --> M[Baseline Models]
        M --> N[Logistic Regression<br/>Accuracy: 49.6%]
        M --> O[Random Forest<br/>Accuracy: 49.1%]
        M --> P[XGBoost<br/>Accuracy: 46.5%]
        M --> Q[SVM<br/>Accuracy: 45.8%]
    end

    subgraph "Hyperparameter Tuning"
        L --> R[GridSearchCV<br/>3-Fold Cross Validation]
        R --> S[Tuned Random Forest<br/>max_depth=10<br/>Accuracy: 52.1%]
        R --> T[Tuned XGBoost<br/>Accuracy: 51.8%]
        R --> U[Tuned SVM<br/>Accuracy: 50.6%]
        R --> V[Tuned Logistic Regression<br/>Accuracy: 50.1%]
    end

    subgraph "Evaluation"
        S --> W[Confusion Matrix]
        T --> W
        U --> W
        V --> W
        W --> X[Best Model:<br/>Random Forest<br/>52.1% Accuracy]
    end

    style X fill:#4CAF50,color:#fff
    style S fill:#4CAF50,color:#fff
    style A fill:#2196F3,color:#fff
    style B fill:#2196F3,color:#fff

Loading

---

Dataset

The project utilizes the Crypto Market Sentiment and Price Dataset (2025) sourced from Kaggle.

Attribute	Detail
Source	Kaggle Dataset
Observations	2,063 data points
Features	14 columns (11 numerical, 1 integer, 2 categorical)
Target	Binary Classification (1 = Bullish/Up, 0 = Bearish/Down)
Missing Values	None (fully clean)
Duplicates	None

Feature Descriptions

Column	Description	Data Type
timestamp	Date/time of the record	object
cryptocurrency	Name of the cryptocurrency (10 unique coins)	object
current_price_usd	Current market price in USD	float64
price_change_24h_percent	% price change in the last 24h	float64
trading_volume_24h	Total trading volume in the last 24h	float64
market_cap_usd	Market capitalization in USD	float64
social_sentiment_score	Sentiment score from social media	float64
news_sentiment_score	Sentiment from news sources	float64
news_impact_score	Strength/impact of news events	float64
social_mentions_count	Number of social media mentions	int64
fear_greed_index	Market emotion indicator (0–100)	float64
volatility_index	Market volatility measure	float64
rsi_technical_indicator	Relative Strength Index value	float64
prediction_confidence	Confidence score of prediction	float64

Cryptocurrencies Covered

Bitcoin, Ethereum, Solana, Cardano, Polkadot, Avalanche, Chainlink, Polygon, Algorand, Cosmos

---

Methodology

1. Data Pre-processing

• Target Engineering: Converted continuous price_change_24h_percent into a binary target (bullish if > 0%)
• Cleaning: Verified data integrity — zero missing values and zero duplicates
• Type Conversion: timestamp converted from object to datetime64
• Imputation: Numerical columns imputed with median; categorical with most frequent
• Encoding: One-Hot Encoding for cryptocurrency (10 binary coin columns)
• Scaling: StandardScaler applied to all numerical features (mean=0, std=1)
• Outlier Detection: IQR method applied (outliers retained as they represent real market events)

2. Feature Engineering

• Target: price_change_24h_percent > 0 → 1 (Bullish), else 0 (Bearish)
• Features dropped: timestamp, price_change_24h_percent, prediction_confidence
• Preprocessing pipeline using ColumnTransformer with StandardScaler for numeric and OneHotEncoder for categorical features

3. Models Evaluated

Four distinct algorithm families were tested:

Model	Type	Baseline Accuracy	Tuned Accuracy	Improvement
Random Forest	Bagging Ensemble	49.1%	52.1%	+3.0%
XGBoost (Gradient Boosting)	Boosting Ensemble	46.5%	51.8%	+5.3%
SVM (RBF Kernel)	Kernel-based	45.8%	50.6%	+4.8%
Logistic Regression	Linear Baseline	49.6%	50.1%	+0.5%

4. Hyperparameter Optimization

• Method: GridSearchCV with 3-Fold Cross-Validation
• Best Random Forest Params: max_depth=10, n_estimators=100, min_samples_split=5
• Best XGBoost Params: learning_rate=0.01, max_depth=3, n_estimators=50
• Best SVM Params: C=10, kernel='rbf', gamma='auto'

---

Results

Key Findings

1. Random Forest achieved the highest stability and accuracy (52.1%), suggesting that regularization is key when modeling high-noise financial data
2. Sentiment features show very low linear correlation with price direction (< 0.02)
3. Non-linear ensemble methods (Random Forest, XGBoost) consistently outperformed linear models
4. Hyperparameter tuning improved ensemble models by 3-5%, while linear models saw minimal gains
5. The marginal edge over random guessing (50%) confirms the challenge of market prediction using sentiment alone

Visualizations

The images/ directory contains 12 generated plots:

• CorrelationHeatmap.png — Feature correlation matrix
• Scatterplot_Cryptocurrency_Price_VS_Trading_vol.png — Price vs Volume (log-log)
• Violonplot_overlay_on_Boxplot.png — Price change distribution by coin
• Violonplot_with_swarm_overlay.png — Sentiment distributions
• Jointplot_Scatterplot_regplot_histplot.png — Sentiment relationship analysis
• Histogram_RSI_Indicator.png — RSI distribution
• 3D_Scatterplotting.png — 3D sentiment visualization
• Barh.png — Horizontal bar charts
• Outliers.png / Handled_outliers.png — Outlier analysis
• Hexbin_plot.png — Hexbin density plot
• ConfusionMatrix_Random_Forest.png — Best model confusion matrix

---

Repository Structure

AI-ML-Model-For-CryptoAnalysis/
├── Academic_Report/                    # Final academic report (PDF) and peer assessment
│   ├── Crypto_Trend_Analysis_Report.pdf
│   └── peer-assessment.xlsx
├── dataset/                            # Dataset files
│   ├── crypto_sentiment_prediction_dataset.csv   # Raw dataset (2063 x 14)
│   ├── dataset_overview.md                       # Detailed dataset documentation
│   └── pre-processed-crypto-dataset.csv          # Preprocessed dataset
├── Final_Notebook/                     # Merged final notebook (all team contributions)
│   ├── Crypto_Trend_Analysis_Final.ipynb         # Complete analysis pipeline
│   └── Final_notebook_info.md                    # Notebook documentation
├── images/                             # 12 generated visualization plots
│   ├── CorrelationHeatmap.png
│   ├── Scatterplot_Cryptocurrency_Price_VS_Trading_vol.png
│   ├── ConfusionMatrix_Random_Forest.png
│   └── ... (9 more)
├── Project-Brief/                      # Assessment brief and sample reports (PDFs)
├── src_notebooks/                      # Individual team member notebooks
│   ├── data_set_overview.ipynb                    # Jaspreet Kaur
│   ├── data_set_exploration(EDA).ipynb            # Rajveer Singh Saini
│   ├── data_set_preprocessing.ipynb               # Harshpreet Singh
│   └── model_training_&_Evaluation.ipynb          # Amogh Dath Kalasapura Arunkumar
├── requirements.txt                    # Python dependencies
└── README.md                           # This file

---

Installation & Usage

Prerequisites

• Python 3.8+
• Pip

Setup

# Clone the repository
git clone https://github.com/Amogh-007-Rin/AI-ML-Model-For-CryptoAnalysis
cd AI-ML-Model-For-CryptoAnalysis

# Install dependencies
pip install -r requirements.txt

Run the Analysis

# Launch Jupyter and open the final notebook
jupyter notebook Final_Notebook/Crypto_Trend_Analysis_Final.ipynb

Or explore individual notebooks:

jupyter notebook src_notebooks/data_set_overview.ipynb
jupyter notebook src_notebooks/data_set_exploration\(EDA\).ipynb
jupyter notebook src_notebooks/data_set_preprocessing.ipynb
jupyter notebook src_notebooks/model_training_\&_Evaluation.ipynb

---

Dependencies

pandas
numpy
seaborn
matplotlib
scikit-learn
scipy

---

Conclusions & Future Work

Conclusions

• Sentiment analysis provides a marginal but real predictive signal for crypto price direction
• Non-linear ensemble models (Random Forest, XGBoost) are better suited than linear models for the high-noise crypto domain
• Random Forest with max_depth=10 achieved the best balance of bias and variance
• The Efficient Market Hypothesis is partially challenged — markets are not perfectly efficient with respect to sentiment data

Future Directions

1. Time Series Analysis: Implement LSTM/GRU models to capture temporal dependencies
2. Deep Learning: Explore transformer-based models for sequence prediction
3. Cross-Market Analysis: Extend to correlations across different cryptocurrency markets
4. Real-Time Prediction: Develop a live prediction system with streaming data
5. Alternative Data Sources: Incorporate on-chain metrics, developer activity, and regulatory news

---

License

This project is licensed under the MIT License.

Acknowledgements

• Birmingham City University — CMP5367: AI and Machine Learning module
• Kaggle — Crypto Market Sentiment and Price Dataset by Pratyush Puri