btc-quant-prob: Probabilistic Bitcoin Price Forecasting

A robust Python repository for probabilistic Bitcoin (BTC-USD) price forecasting, focused on uncertainty quantification, rigorous backtesting, and automated optimization. The system is engineered for data science research, educational exploration, and developing advanced trading strategies — not for financial advice or live trading.

Disclaimer: This project is for research and educational purposes only. It is not investment advice. Financial markets are uncertain; past performance does not guarantee future results.

---

🏛️ Architecture Overview

The system is built modularly, with clear divisions between key processes:

1. Data Pipeline (/data)
   Ingests, validates, and preprocesses raw OHLCV (Open/High/Low/Close/Volume) data.

2. Feature Engineering (/features)
   Constructs features including technical indicators, statistical metrics, volatility ratios, "halving" event counters, and custom “alpha” features with a scikit-learn pipeline.

3. Modeling (/models)
   Implements several probabilistic models:

   • LightGBM Quantile Regression for fast, direct quantile forecasts.
   • Bayesian LSTM (with MC Dropout) for deep temporal uncertainty modeling (PyTorch).
   • Gaussian Process for Bayesian baselines.
   • Ensemble models combining various approaches by performance-weighted stacking.

4. Training & Hyperparameter Tuning (/train, /autotune)
   Unified scripts for model training, leveraging Optuna for automatic hyperparameter search.

5. Backtesting (/backtest)
   Rigorous walk-forward validation with trading simulation, including risk management logic.

6. Evaluation (/eval)
   Tools for forecast quality measurement, such as CRPS, PICP, Winkler Score, calibration plots, and post-hoc recalibration.

7. Deployment & Monitoring (/deploy)
   FastAPI serving for predictions and a performance drift monitor.

---

🚀 Getting Started

Prerequisites

• Python 3.9+
• Git

Installation

1. Clone the repository:

   git clone <repository_url>
   cd btc-quant-prob

2. (Recommended) Create and activate a virtual environment, then install dependencies:

   python -m venv venv
   source venv/bin/activate  # Windows: venv\Scripts\activate
   pip install -r requirements.txt

3. Download Data:
   Get Bitcoin Historical Data from Kaggle and place bitcoin_historical_data.csv inside btc-quant-prob/raw_data/.

   Tip: For quick tests, a mock data generator is also included in the codebase.

---

⚙️ Running the System

Basic workflow:

1. Activate your virtual environment:

   source venv/bin/activate

   You should see (venv) in your terminal prompt.

2. Train a Model:

   python -m train.train --horizon 180 --model_name lgbm_quantile

   Trains a LightGBM quantile regression model for 180-day predictions, saving the trained model in artifacts/.

3. Generate Walk-Forward Backtest Predictions:

   python -m backtest.walkforward --horizon 180 --model_name lgbm_quantile

   Stores predictions in artifacts/backtest/.

4. Evaluate and Visualize:
   Open and execute the notebook notebooks/02_traineval.ipynb for analysis and visualization.

Summary Workflow:

1. Training (train.train) → produces model
2. Backtesting (backtest.walkforward) → produces historical results
3. Notebook evaluation (notebooks/02_traineval.ipynb) → visualization and analysis

---

🧑‍🔬 Advanced Usage & Suggested Improvements

Phase 1: System Optimization

• Tune hyperparameters for each model (see autotune/iterate.py, using Optuna).
• Engineer new, original features ("alpha"). Ideas:
  • Short/long-term volatility ratios
  • Feature interactions (e.g., RSI × volatility)
  • Halving effects (add_halving_features)
• Try more models or ensembles (models/gp_baseline.py, models/ensemble.py).
• Refine backtester: Add stop-loss, volatility-based position sizing, and "high-conviction" trading signals in backtest/simulate.py.

Phase 2: Expand Knowledge Base

• Deepen knowledge in probability/statistics (CRPS, Bayesian methods), time-series analysis (ARIMA, GARCH), linear algebra/calculus.
• Understand financial market concepts: EMH, market structure, options/futures.

Phase 3: Build Portfolio & Community

• Document experiments methodically (“50 features tested, feature X improved Sharpe by 0.2…”).
• Join Kaggle time series competitions.
• Read quant research papers and replicate simple ideas.

---

📝 Example Past Experiment Log

• Applied Optuna hyperparameter tuning (50+ trials). Achieved low CRPS, but trading results poor with default median-based strategy (Sharpe < 0, max drawdown -100%).
• Added "halving" and new interaction features: model uncertainty prediction improved (CRPS = 0.11).
• Tried high-conviction, tail-based, and contrarian trading strategies. Only contrarian logic yielded sustainable returns (Sharpe ≈ 2), but drawdown remains problematic during regime shifts.
• Discovered that feature-rich models can forecast uncertainty, not direction — highlighting the challenge of trading with predictive signals.
• Plan: add volatility-based position sizing and trend regime filters (e.g., SMA-200) for risk management.

---

📚 References

• mczielinski/bitcoin-historical-data (Kaggle)
• CRPS: Continuous Ranked Probability Score
• Quant blog reading: AQR, Renaissance, Two Sigma, etc.

---

📢 Contributing

Research ideas, feature pull requests, and reproducible experiment logs are welcome. Open issues, document your findings, and help make this project a clearer resource for quant-oriented probabilistic research!

---

License

MIT (c) 2024-present
Original author: jwjooth