tri-domain-moe

Model Card: TriDomainMoE (BTCUSD v2.0 Production Checkpoint)

Model Summary

TriDomainMoE is an institutional multi-domain Mixture of Experts (MoE) model engineered for continuous 24/7 financial market prediction on BTCUSD (Bitcoin / US Dollar). It coordinates three specialized domain experts through a continuous Softmax Correlation-Aware Weighting (CAW) router to deliver robust directional drift and meta-calibrated conviction sizing while strictly adhering to a $< 2.50%$ trailing drawdown ceiling.

Architecture Specifications

• Tech Expert Input: 6 Microstructural features (Dilated Causal Convolutions with $d \in {1, 2}$, Parkinson High-Low Volatility Ratio, Bar OFI, Fractional Differencing $d^*=0.45$). Lookback: 32 M5 bars.
• Macro Expert Input: 6 Macro cross-asset features (HiPPO Linear Recurrence SSM, H4/D1 secular trends, 24h/168h volatility term slope).
• Fundamental Expert Input: 8 Narrative sentiment features (Gated Residual Highway, 24h net CVD volume delta, decay kernel $z_t$).
• Router: 8-dimensional regime state vector with continuous Softmax CAW routing and cosine repulsion orthogonal separation.
• Meta-Sizer: Calibrated sigmoid bet sizer outputting continuous trade conviction $s_t \in [0, 2.0]$.

Benchmark Performance (1-Year Real Tick Evaluation)

How to Load and Use

import torch
from src.models.institutional_moe import TriDomainMoE

# Load checkpoint
checkpoint_path = "weights/btcusd_tri_domain_v2.pt"
checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=False)

# Instantiate model from checkpoint configuration
model = TriDomainMoE(
    tech_dim=checkpoint["tech_dim"],      # 6
    macro_dim=checkpoint["macro_dim"],    # 6
    fund_dim=checkpoint["fund_dim"],      # 8
    regime_dim=checkpoint["regime_dim"],  # 8
    hidden_dim=checkpoint["hidden_dim"],  # 48
)
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()

# Dummy input tensors matching operational dimensions
batch_size = 1
x_tech = torch.randn(batch_size, 32, 6)   # 32 M5 bars of 6 tech features
x_macro = torch.randn(batch_size, 32, 6)  # 32 bars of 6 macro features
x_fund = torch.randn(batch_size, 8)       # 8 fundamental features
z_regime = torch.randn(batch_size, 8)     # 8D regime vector

with torch.no_grad():
    outputs = model(x_tech, x_macro, x_fund, z_regime)
    y_pred = outputs["y_pred"].item()       # Directional drift forecast
    conviction = outputs["size"].item()     # Calibrated conviction sizing [0, 2]
    weights = outputs["weights"].squeeze()  # [Tech, Macro, Fund] expert allocation

print(f"Drift Forecast: {y_pred:+.4f} | Conviction: {conviction:.2f}")
print(f"Expert Allocation -> Tech: {weights[0]*100:.1f}% | Macro: {weights[1]*100:.1f}% | Fund: {weights[2]*100:.1f}%")

Intended Use & Limitations

• Intended Use: Algorithmic quantitative research, signal generation, and hedge fund portfolio risk modeling.
• Limitations: Trained on institutional broker floating spreads ($\approx $65$ on BTC). Execution models must account for broker slippage, weekend swap fees, and liquidity conditions.

Links & Ecosystem

• GitHub Repository: https://github.com/ElMoorish/TriDomainMoE
• Official Web Portal: https://primeclub-quant.vercel.app/
• Hugging Face Hub: https://huggingface.co/ElMoorish/tri-domain-moe

💖 Support & Research Grants (Donations)

Developing and live-forward testing institutional algorithmic intelligence requires 24/7 GPU compute, high-frequency tick data streams, and execution infrastructure for TriDomainMoE and FinRL-X-MT5.

If this research provides value to your operations, cryptocurrency grants directly accelerate continuous live testing and open-source model releases:

TRC20 USDT Address:
TC8TFkemSFGEeBPF5ZQKbmjK97FVEGwrwc

License

Apache License 2.0.