Menu
Menu
Close
Close
CryptoForecast: Bitcoin Price Prediction
An end-to-end machine learning pipeline for Bitcoin price prediction, covering the full cycle from data collection and preprocessing to feature engineering, model training, evaluation, and visualization.
Overview
CryptoForecast is an end-to-end machine learning pipeline for Bitcoin price prediction. It covers the full cycle from data collection and preprocessing to feature engineering, model training, evaluation, and visualization. The project leverages multiple data sources (e.g., Binance, CryptoCompare, Yahoo Finance) and extensive feature engineering (100+ technical indicators, patterns, and market regime features) to provide accurate price forecasts and market insights.
From
raw
market
data
to
actionable
predictions
Key Features
Modeling Approaches
The project explores multiple modeling techniques, implemented in a modular way (organized under `src/models/` for traditional, deep learning, ensemble, and RL models), allowing comprehensive comparison of their performance:
Traditional ML Models
Random Forest, Gradient Boosting (XGBoost, LightGBM, CatBoost) - powerful tree-based algorithms that excel at capturing non-linear relationships in financial time series data.
Deep Learning Models
LSTM and GRU recurrent neural networks, including bidirectional LSTM - designed to capture long-term dependencies and temporal patterns in price sequences.
Ensemble Methods
Voting, Stacking, and Blending ensembles that combine predictions from multiple models to improve robustness and generalization.
Reinforcement Learning
Agents using PPO (Proximal Policy Optimization) and A2C (Advantage Actor-Critic) algorithms to make trading decisions based on price movements and market conditions.
Diverse
models,
unified
framework
Results and Insights
Extensive experimentation revealed notable findings:
Best Overall Performance
The best overall performance was achieved by a Stacking Ensemble model combining multiple gradient boosting algorithms, demonstrating the power of ensemble methods in financial forecasting.
Highest Standalone Accuracy
CatBoost (a gradient boosting model) with custom engineered features had the highest standalone accuracy, showcasing the importance of thoughtful feature engineering.
Most Influential Features
Indicators such as RSI (Relative Strength Index), Bollinger Bands, and volume-based metrics were among the top predictors of price movement, highlighting the significance of technical analysis in cryptocurrency markets.
Directional Accuracy
The models could predict directional movement with up to ~75% accuracy, indicating significant skill in forecasting price direction, which is often more valuable than exact price prediction for trading strategies.
Market Regime Impact
Model performance varied across market regimes (bull vs. bear markets), suggesting that different market conditions impact prediction accuracy. This insight emphasizes the importance of adaptive modeling strategies.
A
powerful
framework
for
crypto
prediction
Overall, CryptoForecast demonstrates a powerful framework for crypto price prediction by integrating diverse data sources, sophisticated features, and multiple modeling techniques in a unified pipeline. The modular architecture allows for easy extension and experimentation, making it a valuable tool for both research and practical applications in cryptocurrency trading and analysis.
