Comparative Analysis of Bitcoin Price Prediction: LSTM, BiLSTM, ARIMA & Transformers

Overview

This project focuses on Bitcoin price prediction using various deep learning and statistical models, including Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Autoregressive Integrated Moving Average (ARIMA), and Auto-regressive Encoder-Decoder Transformers. By leveraging these techniques, the project aims to analyze historical Bitcoin price data and forecast future price trends with improved accuracy.

The proposed approach is tested on a dataset from Yahoo Finance, spanning 10 years of Bitcoin price data. The study demonstrates that the Transformer-based model outperforms traditional approaches, achieving the highest prediction accuracy with an R² score of 0.9941. Below is an overview of the analysis, along with sample outputs and results. This project was done in May' 2024.

---

Publication

• This work was presented in the "2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT)".
• Link to IEEE Publication: https://ieeexplore.ieee.org/abstract/document/10724247/

---

Features

• LSTM & BiLSTM-Based Sequential Prediction: Captures long-term dependencies in Bitcoin price trends for more reliable forecasting.

• Auto-regressive Transformer Model: Employs self-attention mechanisms to enhance temporal relationships and improve price prediction accuracy.

• ARIMA for Traditional Time-Series Forecasting: Utilizes statistical modeling to identify trends and seasonal patterns in Bitcoin prices.

• Fourier Transform Feature Engineering: Decomposes price data into frequency components to enhance prediction performance.

• High Accuracy and Low Error: The Transformer model achieves the highest accuracy, outperforming LSTM, BiLSTM, and ARIMA.

---

Tech Stack

• Python – Core language for model implementation.
• TensorFlow & Keras – Used for deep learning models like LSTM, BiLSTM, and Transformers.
• Statsmodels – Implements ARIMA for time-series forecasting.
• Matplotlib & Seaborn – Used for visualizing Bitcoin price trends and model performance.
• Scikit-Learn – Preprocessing, scaling, and evaluation metrics.

---

Installation

1. Download the Dataset:

   • The Bitcoin historical price dataset can be obtained from Yahoo Finance.

2. Preprocess the Data:

   • Run Comparitive Analysis for Bitcoin.py to clean and normalize the data.

3. Train the Models:

   • Execute Comparitive Analysis for Bitcoin.py for LSTM, BiLSTM, ARIMA and Transformers. The project is divided into cells where there is a section for each the mentioned models.

4. Evaluate the Results:

   • Run Comparitive Analysis for Bitcoin.py to compare the models and visualize results.

---

Running Tests

The models can be tested using historical Bitcoin price data to verify prediction accuracy.

---