metadata
tags:
- forex
- trading
- transformer
- time-series
license: mit
OMEGA Forex v3
Dedicated forex trading signal model — CNN + RegimeGate + Transformer + Standard Candle.
Architecture
- Conv1d(k=3) → Conv1d(k=5) → RegimeGate → 3-layer Transformer (pre-norm) → Classifier
- d_model=64, 4 attention heads, FFN=256
- ~184,354 parameters
- Input: seq_len=24 × 46 features → binary UP/DOWN
Features (46)
22 technical + 11 session + 5 Keplerian lag + 5 regime + 3 USD/JPY Standard Candle anchor
Pairs
EUR/USD, GBP/USD, USD/JPY, USD/CHF, AUD/USD, GBP/JPY
Inference
# run.py reads config.json fresh each hour — anchor features require USD/JPY to be
# fetched and processed first, then injected into each other pair's feature window.
Not financial advice. Research and educational use only.