LSTM Forecasting of Bitcoin Implied Volatility (DVOL)

🟦 Writing Phase
Master's Thesis – Economics (STEM Designated)

Overview

This thesis develops and validates a Long Short-Term Memory (LSTM) neural network model for forecasting Bitcoin implied volatility (DVOL), the Deribit 30-day volatility index. Using a unified framework of 17 models (13 linear/tree baselines + 4 LSTM variants), we demonstrate that LSTM models achieve competitive performance (R² = 0.9287) when properly evaluated.

Research Objective

Develop an LSTM neural network model to forecast Bitcoin implied volatility (DVOL) using on-chain metrics and historical volatility patterns, validated through statistical analysis and economic significance.

Core Predictors

Lagged DVOL

1-day, 7-day, 30-day lags

Justification:

Transaction Volume (USD)

Daily on-chain transaction volume

Justification:

Active Addresses Count

Daily count of active Bitcoin addresses

Justification:

Network Value to Realized Value (NVRV)

Market value/realized value ratio

Justification:

DVOL-RV Spread

Volatility risk premium (DVOL - 30-day realized volatility)

Justification:

Technologies

Python, PyTorch, TensorFlow/Keras, Pandas, NumPy, Scikit-learn, SHAP

Abstract

This thesis develops and validates a Long Short-Term Memory (LSTM) neural network model for forecasting Bitcoin implied volatility (DVOL), the Deribit 30-day volatility index. Using a unified framework of 17 models (13 linear/tree baselines + 4 LSTM variants), we demonstrate that LSTM models achieve competitive performance (R² = 0.9287) when properly evaluated, narrowing the gap to linear/tree models to only 2%.

Key Contributions

  • Multi-window normalization analysis: 72-hour (3-day) window is optimal for R² level prediction
  • Standard Lee-Mykland (2008) implementation: Academically rigorous jump detection with 236 jumps (0.57%)
  • DataParallel evaluation artifact: Identified 13.3% R² degradation in multi-GPU LSTM evaluation
  • LSTM competitiveness: With fixed evaluation, LSTM (R²=0.9287) within 2% of best linear/tree models

Model Comparison Results

All 17 models use identical preprocessing (60/20/20 split, 720h rolling normalization):

ModelTypeFeaturesR²RMSE
RF (Lags + Jumps)Tree110.94921.65
OLS (Lags + Jumps)Linear110.94901.65
HAR-RVLinear30.94541.71
LSTM market_lagsLSTM70.92871.71

Practical Implications

  • Suitable for: Risk management, option pricing, volatility level estimation, VaR calculations
  • NOT suitable for: Directional trading, market timing strategies
  • Directional accuracy ~50% across all models indicates hourly direction is fundamentally unpredictable

Code & Data

The complete codebase, dataset documentation, and reproducibility instructions are available on GitHub.