A comprehensive project focusing on Bitcoin price prediction using neural network models. This repository demonstrates a structured approach to time series analysis and forecasting, comparing the performance of Multi-Layer Perceptron (MLP) and Convolutional Neural Network (CNN) architectures on Bitcoin's closing price data.
Bitcoin's price is notoriously volatile, making it a challenging yet valuable target for time series forecasting. This project involves:
-
Data Collection and Preprocessing:
- Data sourced from Yahoo Finance for the period 2018-01-01 to 2023-01-01.
- Preprocessing techniques include log transformations, stationary differencing, and fractional differencing.
-
Feature Engineering:
- Lagged features are created to capture temporal dependencies in the dataset.
-
Model Architectures:
- MLP: A neural network with fully connected layers designed for regression tasks.
- CNN: A convolutional architecture utilizing Gramian Angular Field (GAF) transformations to analyze time series as images.
-
Performance Comparison:
- Models are evaluated on datasets with varying levels of preprocessing.
- Key metrics include Root Mean Square Error (RMSE).
- MLP: Best suited for datasets with global dependencies or long-term memory, such as fractionally differenced series.
- CNN: Excels in capturing localized patterns in stationary data but struggles with raw, non-stationary inputs.
This project is licensed under the MIT License. See the LICENSE file for details.