01 Architecture and Foundations

🏛 01. Architecture & Foundations

The core design philosophy of TriDomainMoE is the orthogonal decomposition of financial alpha. Rather than forcing a single neural network to predict price movements across conflicting time horizons, market mechanics are separated into three specialized neural domain experts coordinated by a Continuous Softmax Correlation-Aware Router (CAW).

---

1. The Three Orthogonal Neural Experts

1.1 Microstructure Technical Expert

• Neural Backbone: Dilated Causal 1D Convolutions (CausalConv1d) with dilation rates $d \in {1, 2}$, kernel size $k=3$, and LeakyReLU activations.
• Causal Guarantee: Output at time step $t$ depends strictly on features $\le t$. Padding is applied exclusively to the left ($p = (k-1) \cdot d$), guaranteeing zero future lookahead.
• Receptive Field: Expands across 32 continuous M5 bars ($2.67$ hours of intra-bar liquidity flow).
• Features Captured: Fractional differencing ($d^*=0.45$), Parkinson High-Low volatility ratios, Volume-Synchronized Order Flow Imbalance (OFI), and normalized log return momentum.

1.2 Macro Term Structure State-Space Model (SSM)

• Neural Backbone: Diagonal Linear Recurrent State-Space Block initialized with a HiPPO log-spaced timescale spectrum: $$\alpha_i = \text{sigmoid}\left(\text{logit}(0.85) + \frac{i}{D-1}(\text{logit}(0.995) - \text{logit}(0.85))\right)$$ $$h_t^{(i)} = \alpha_i h_{t-1}^{(i)} + (1 - \alpha_i) x_t^{(i)}$$
• Multi-Cycle Retention: Because decay parameters are bounded between $0.85$ and $0.995$, individual state-space channels retain memory across hours, days, and multi-week macro secular cycles without vanishing gradients.
• Features Captured: H1 trend alignment, 24h rolling volatility, H4 swing trend (rolling 200 H1 bars), D1 secular regime (rolling 600 H1 bars), and volatility term structure slope ($\sigma_{24\text{h}} / \sigma_{168\text{h}}$).

1.3 Fundamental Narrative Sentiment Expert

• Neural Backbone: Deep Gated Residual Highway Network with LayerNorm, GELU non-linearities, and direct pre-gated skip connections: $$h_{\text{gated}} = \sigma(W_g x) \odot \tanh(W_v x)$$ $$\text{output} = h_{\text{gated}} + \text{MLP}(h_{\text{gated}})$$
• Non-Linear News Modulation: Eliminates gradient saturation during high-impact macroeconomic releases (FOMC, CPI, geopolitical announcements).
• Features Captured: Monetary policy polarity, institutional crypto ETF flows, regulatory risk indices, and continuous 24-hour Cumulative Volume Delta (CVD).

---

2. Continuous Softmax Correlation-Aware Router (CAW)

The router computes continuous, dynamic weighting coefficients across the three experts conditioned on an 8-dimensional regime representation $z_t$:

$$g(z_t) = \text{Softmax}\left(\frac{W_r z_t + b_r}{\tau}\right) \in [0, 1]^3, \quad \sum_{i=1}^3 g_i(z_t) = 1.0$$

Where $\tau = 1.0$ is the temperature parameter.

2.1 Cosine Repulsion Orthogonality Loss

To prevent the three domain experts from collapsing into identical redundant representations, the composite training loss penalizes feature correlation:

$$\mathcal{L}_{\text{rep}} = \frac{1}{M(M-1)} \sum_{i \neq j} \max\left(0, \cos(h_i, h_j)\right)$$

Where $h_i$ represents the normalized latent embedding of expert $i$. This forces the technical, macro, and fundamental channels into mutually orthogonal mathematical representations.

2.2 Shannon Entropy Collapse Prevention

The gating distribution's Shannon entropy is continuously tracked:

$$H(g) = -\sum_{i=1}^3 g_i \log g_i$$

• Normal Adaptive Regime: $0.35 \le H(g) < 1.00$.
• Conflicted Market Chop: When $H(g) \ge 1.00$, the experts are in conflict (e.g. Microstructure says BUY, Macro SSM says SELL). The system activates the Vector 3 False Alarm Filter, elevating the required drift threshold from $0.0300$ to $0.0380$.
• Collapse Detection: If $H(g) < 0.35$, the router has collapsed into a single dominant expert; a diversity penalty immediately restores balance.

---

3. Meta-Sizer (Continuous Conviction)

The directional forecast $\hat{y}_t$ is combined with the regime vector to determine capital allocation:

$$s_t = 2.0 \cdot \sigma(W_s [\hat{y}_t; z_t] + b_s) \in [0, 2.0]$$

• Sizing $s_t < 0.20$: Trade is pruned (insufficient edge).
• Sizing $0.20 \le s_t < 1.00$: Scaled defensive lot sizing.
• Sizing $s_t \ge 1.00$: Full conviction trade execution.