Precision LSTM Trading Model v2
Bidirectional LSTM with Multi-Head Self-Attention for cryptocurrency trading signal prediction.
Model Details
- Architecture: BiLSTM(128) + BiLSTM(64) + Multi-Head Attention(4 heads) + GlobalAveragePooling + Dense layers
- Features: 28 engineered technical indicators across 96 candles (24-hour window)
- Output Classes: LONG, SHORT, HOLD
- Loss Function: Focal Loss with label smoothing
Performance
- Test Loss: 0.0961
- Test Accuracy: 0.4563
- LONG F1 Score: 0.5818
- SHORT F1 Score: 0.2674
Training Configuration
- Epochs: 20
- Batch Size: 128
- Training Samples: 159777
- Test Samples: 39603
- Regularization: L1L2 (L1=1e-05, L2=0.0001)
Features (28 total)
- OHLCV (5): open, high, low, close, volume
- Trend (3): SMA_20, STD_20, returns
- SMC (6): FVG bullish/bearish, liquidity sweeps, break of structure
- Momentum (3): RSI_14, MACD, MACD_signal
- Volatility (1): ATR_14
- Bollinger (2): upper/lower distance
- Volume (1): volume SMA ratio
- Candle (3): body, upper shadow, lower shadow
- Advanced SMC (4): order blocks, change of character
Usage
\python import tensorflow as tf import joblib import numpy as np
Load model and scaler
model = tf.keras.models.load_model('trading_model.keras') scaler = joblib.load('feature_scaler.joblib')
Prepare features (shape: samples, 96, 28)
features_scaled = scaler.transform(features)
Make predictions
predictions = model.predict(features_scaled) \
Training Date
2026-05-11T17:56:52.932268+00:00
License
MIT
- Downloads last month
- 4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support