02 Feature Engineering Pipeline

📊 02. Feature Engineering Pipeline

The TriDomainMoE feature engineering suite is implemented in src/features/tri_domain_features.py to ensure 100% strict mathematical parity between offline historical training, vectorized backtesting, and sub-20ms live MT5 execution.

---

1. Technical Domain Features (6D)

Index	Feature Symbol	Mathematical Formulation	Economic Rationale
0	log_ret	$\ln(C_t / C_{t-1})$	Instantaneous bar return
1	vol_20	$\text{Std}(\text{log_ret}, N=20)$	Short-term realized volatility baseline
2	fracdiff	$(1 - B)^d C_t, \quad d^*=0.45$	Stationarity preserving continuous long-memory price action
3	vol_mom	$V_t / \text{SMA}(V, 20) - 1.0$	Surge volume above rolling equilibrium
4	parkinson_ratio	$\sigma_{\text{parkinson}} / \sigma_{\text{close}}$	Distinguishes wick fakeouts from authentic directional expansion
5	bar_ofi	$CLV_t \times \text{vol_mom}_t$	Bar-level Order Flow Imbalance proxy ($CLV = \frac{(C - L) - (H - C)}{H - L}$)

1.1 Fractional Differencing ($d^* = 0.45$)

Integer differencing ($d=1.0$) strips all historical memory, creating an unlearnable white-noise series. We apply fractional differencing via the binomial series expansion:

$$(1 - B)^d = \sum_{k=0}^{\infty} (-1)^k \binom{d}{k} B^k = 1 - d B + \frac{d(d-1)}{2!} B^2 - \dots$$

Using the Augmented Dickey-Fuller (ADF) test, $d^* = 0.45$ achieves stationarity ($p < 0.01$) while retaining over $90%$ of price level correlation.

1.2 Parkinson Volatility Ratio

Using intraday High and Low bounds, the Parkinson volatility estimator is:

$$\sigma_{\text{parkinson}} = \sqrt{\frac{1}{4 \ln 2 \cdot N} \sum_{i=1}^N \left(\ln \frac{H_i}{L_i}\right)^2}$$

When $\sigma_{\text{parkinson}} / \sigma_{\text{close}} \gg 1.0$, market activity consists of volatile whipsaws rather than trend momentum, prompting defensive position sizing.

---

2. Macro Term Structure Domain Features (6D)

Index	Feature Symbol	Mathematical Formulation	Economic Rationale
0	h1_log_ret	$\ln(C_{H1, t} / C_{H1, t-1})$	Higher-timeframe macro drift
1	h1_trend_50	$(C_{H1} - \text{EMA}{50}) / \text{ATR}{14}$	Macro trend persistence relative to average true range
2	h1_vol_24h	$\text{Std}(r_{H1}, 24) \times \sqrt{24 \times 365}$	24-hour annualized historical crypto volatility
3	h4_trend_200	$(C_{H1} - \text{EMA}{200}) / \text{ATR}{14}$	Multi-day swing regime anchor
4	d1_trend_600	$(C_{H1} - \text{EMA}{600}) / \text{ATR}{14}$	Secular macro regime anchor (institutional bias)
5	vol_term_slope	$\sigma_{24\text{h}} / \sigma_{168\text{h}}$	Term structure curve (short vs weekly volatility expansion)

---

3. Fundamental Crypto Sentiment Domain Features (8D)

Index	Feature Symbol	Mathematical Formulation	Economic Rationale
0	polarity	$\in [-1.0, +1.0]$	Net directional sentiment across crypto news releases
1	uncertainty	$\in [0.0, 1.0]$	Entropy / ambiguity score in institutional reportage
2	monetary_score	$\in [-1.0, +1.0]$	Central bank liquidity expectations (Fed / DXY correlation)
3	geopolitics	$\in [0.0, 1.0]$	Geopolitical risk and safe-haven rotation intensity
4	corporate_flow	$\in [-1.0, +1.0]$	Institutional crypto treasury and ETF net flows
5	commodities	$\in [-1.0, +1.0]$	Cross-asset commodity / energy inflation pressures
6	cvd_flow_24h	$\sum_{t=1}^{288} \text{OFI}_t$	Continuous 24h rolling Cumulative Volume Delta (CVD)
7	spread_basis_ratio	$\text{Spread}_t / \text{ATR}_t$	Real-time liquidity friction relative to available price volatility

---

4. Microstructure Order Flow Imbalance (OFI) & VPIN

From tick-level trade streams, Volume-Synchronized OFI calculates the net pressure on the bid and ask queues:

$$\text{OFI}_k = \sum_{t \in \text{Bar}_k} \Delta Q_{\text{bid}, t} - \Delta Q_{\text{ask}, t}$$

Volume-Synchronized Probability of Toxicity (VPIN) tracks informed order dumping:

$$\text{VPIN} = \frac{\sum_{\tau=1}^V |V_\tau^B - V_\tau^S|}{V \cdot N}$$

When $\text{VPIN} > 0.65$, adverse selection risk spikes; the system automatically triggers the Pre-Trade Anti-Adverse Book Shield, refusing toxic incoming orders.