⚡
Hybrid Mamba-Transformer • Direction Prediction
CrossMamba™
Stream 1: Mamba (long-range) + Stream 2: Transformer (short-range) + Learned Gating Fusion
19:06:020 DEPLOYED • 0 XGBoost RETAINED
Models Trained
0
CM Wins
0/0
Avg Accuracy
0.0%
Best Model
0.0%
Best Asset
GPU
RunPod A40
CrossMamba™ Architecture
Stream 1: Mamba Expert
• State Space Model (SSM)
• Processes 720 candles (long-range)
• Linear time complexity O(n)
• Captures macro trends, seasonality
• 4 Mamba layers, d_model=128
Stream 2: Transformer Expert
• Multi-Head Self-Attention
• Processes 60-120 candles (short-range)
• Captures precise momentum shifts
• 2 Transformer layers, 8 heads
• Cross-market reaction detection
Fusion: Learned Gating
• Dynamic weight: Mamba vs Transformer
• Trending → more Mamba weight
• Choppy → more Transformer weight
• Output: UP/DOWN + 0-100 strength
• ~3M parameters, CPU inference <50ms
CrossMamba™ vs XGBoost — Per Asset Results
GBPUSD
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:68.9%-68.9%
8H🌳 XGB
CM:0.0%XGB:59.2%-59.2%
24H🌳 XGB
CM:0.0%XGB:64.5%-64.5%
EURUSD
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:65.2%-65.2%
8H🌳 XGB
CM:0.0%XGB:58.6%-58.6%
24H🌳 XGB
CM:0.0%XGB:51.1%-51.1%
USDJPY
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:65.8%-65.8%
8H🌳 XGB
CM:0.0%XGB:58.6%-58.6%
24H🌳 XGB
CM:0.0%XGB:61.9%-61.9%
USDCHF
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:65.8%-65.8%
8H🌳 XGB
CM:0.0%XGB:55.8%-55.8%
24H🌳 XGB
CM:0.0%XGB:43.7%-43.7%
AUDUSD
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:61.1%-61.1%
8H🌳 XGB
CM:0.0%XGB:60.2%-60.2%
24H🌳 XGB
CM:0.0%XGB:52.5%-52.5%
USDCAD
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:58.2%-58.2%
8H🌳 XGB
CM:0.0%XGB:54.3%-54.3%
24H🌳 XGB
CM:0.0%XGB:37.7%-37.7%
NZDUSD
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:62.2%-62.2%
8H🌳 XGB
CM:0.0%XGB:64.1%-64.1%
24H🌳 XGB
CM:0.0%XGB:75.1%-75.1%
EURGBP
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:59.7%-59.7%
8H🌳 XGB
CM:0.0%XGB:54.7%-54.7%
24H🌳 XGB
CM:0.0%XGB:48.6%-48.6%
EURJPY
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:60.8%-60.8%
8H🌳 XGB
CM:0.0%XGB:54.1%-54.1%
24H🌳 XGB
CM:0.0%XGB:68.0%-68.0%
XAUUSD
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:49.3%-49.3%
8H🌳 XGB
CM:0.0%XGB:53.0%-53.0%
24H🌳 XGB
CM:0.0%XGB:28.7%-28.7%
SPX500
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:47.9%-47.9%
8H🌳 XGB
CM:0.0%XGB:54.7%-54.7%
24H🌳 XGB
CM:0.0%XGB:50.7%-50.7%
NAS100
CM wins 0/30.0%
4H🌳 XGB
CM:0.0%XGB:54.0%-54.0%
8H🌳 XGB
CM:0.0%XGB:57.5%-57.5%
24H🌳 XGB
CM:0.0%XGB:49.1%-49.1%
Training Configuration
Architecture
• 4 Mamba SSM layers
• 2 Transformer layers
• 8 attention heads
• d_model = 128
• Dropout = 0.2
Training
• 100 epochs max
• Early stopping (15 patience)
• AdamW optimizer
• LR = 0.0003 + scheduler
• Walk-forward split 70/15/15
Data
• D1 candles (10 years)
• 153-173 features per asset
• Cross-market correlations
• Sequence length: 120
• Batch size: 64
Infrastructure
• RunPod A40 GPU (48GB)
• PyTorch 2.4 + CUDA 12.8
• ~25 min for 12 assets
• Cost: ~$0.20 per run
• Weekly auto-retrain Saturday
CrossMamba™ — Hybrid Mamba-Transformer for Financial Time Series • © 2026 Flexi Analysis • CONFIDENTIAL