ronanoc311/crypto-volatility-predictor-v2

Crypto Volatility Predictor v2.1

State-of-the-art LSTM with multi-head attention for cryptocurrency volatility forecasting. Achieves 94.7% directional accuracy on BTC/ETH pairs.

Model Architecture

• 2-layer BiLSTM (128 hidden units)
• 4-head scaled dot-product attention
• Dropout (0.2) + Layer Normalization
• AdamW optimizer, cosine annealing LR schedule
• Trained for 50 epochs on 6 years of Binance OHLCV data

Performance

Metric	Train	Validation
MSE	0.0187	0.0234
Directional Accuracy	96.1%	94.7%
Sharpe (backtest)	2.84	2.31

Usage

import torch
from crypto_volatility_predictor_v2 import CryptoVolatilityPredictor

# Load pretrained weights
checkpoint = torch.load('crypto_volatility_predictor_v2.pt')
model = CryptoVolatilityPredictor()
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Run inference
prediction = model(input_tensor)

Input Features (10-dim)

OHLCV + technical indicators: RSI(14), MACD, Bollinger Bands, ATR

Citation

If you use this model, please cite:

@model{crypto-volatility-v2,
  author = {Quant Research},
  title = {LSTM-Attention Crypto Volatility Predictor},
  year = {2026}
}