BNB-NOVADAX-TRADE — Phase 1 Supervised Pre-training

Resultados

• Steps: 204,000
• Épocas: 2
• Tempo de treino: 11.00h
• Val Loss: 2.4436
• Acc 30s: 0.178
• Acc 60s: 0.246
• Acc 120s: 0.317

Arquitetura

• Transformer Encoder | d_model=256 | 8 heads | 6 layers | FFN=1024
• ~10.7M parâmetros
• Contexto: 120 segundos
• Features: 12 features por timestep
• Multi-task: direção 30s/60s/120s + spread futuro + NDX follow

Dataset

• Binance BTCUSDT 1s + NovaDax BTC_USDT | 2020-2026
• Split: 2020-2023 treino | 2024 val | 2025-2026 test (blind)

Como usar (Phase 2 — RL Fine-tuning)

ckpt = torch.load('checkpoint_best.pt', map_location='cpu')
model = TradingTransformer(ckpt['cfg'])
model.load_state_dict(ckpt['model_state'])
emb = model.get_embedding(x)  # usar como encoder no PPO