We don't hide the math. Every layer is documented and reproducible.
1. Multi-Window Instability Engine
The MVIE score is a composite of three Z-score components, computed across 30/60/90-day rolling windows:
# Pseudocode — actual implementation in core/mvie_engine.py
IS = z_score(eigenvalue_dispersion(correlation_matrix))
AR = z_score(autocorrelation(returns))
PD = z_score(singular_value_decay(returns_matrix))
MVIE = sigmoid(0.4 * IS + 0.3 * AR + 0.3 * PD) * 100
Regime thresholds:
- 0–50 GREEN — Stable correlation structure, normal regime
- 50–75 AMBER — Eigenvalue dispersion rising, regime transition warning
- 75–100 RED — Crisis regime band; correlation structure stress elevated or active in the model
2. Network Topology Analysis
A Minimum Spanning Tree is constructed from the asset correlation matrix using the distance metric d_ij = sqrt(2(1 - ρ_ij)). From the MST, three centrality metrics surface systemic importance:
- Eigenvector centrality — Identifies hubs whose stress propagates through the network (e.g., SPY in equities, DXY in FX)
- Betweenness centrality — Surfaces bridge assets that connect otherwise disconnected clusters (contagion paths)
- Degree centrality — Raw connection count, useful as a sanity check
3. Probabilistic Strategy Engine
Unlike binary directional labels, MVIE Pro outputs probability-weighted illustrative scenarios:
Signal weights:
RSI: 30%
MACD: 30%
Bollinger: 20%
ATR-vol: 20%
Output:
Direction: BULLISH | BEARISH | NEUTRAL
Probability: P(direction) ∈ [0, 1]
Envelope: ±2σ ATR-based confidence band
Horizon: 5 trading days, with best/base/worst projections
4. Veto Protocol
The Veto Protocol is a model override rule. When MVIE > 75 (RED regime), all BULLISH scenario labels are reclassified to NEUTRAL or defensive overlays in the published set, regardless of micro-indicator strength. In crisis regimes, the engine treats micro readings as high-noise; the defensive scenario narrative takes precedence in the illustrative output.
5. LLM Analyst Layer
A large language model layer synthesizes the quantitative output into a structured institutional report. The prompt is compliance-locked — no investment advice, only analytical scenarios. Output is markdown-formatted with explicit risk framework sections. The underlying model can be swapped as newer APIs become available.