Skip to content
MVIE Pro
Methodology

A white box, on purpose

Every number the engine reports can be traced back to a published method. This page covers the mathematics inside, why the model refuses to be a black box, and how it differs from VaR.

Inside the Brain: how MVIE works

We don't hide the math. Every layer is documented and reproducible.

1. Rolling-Window Instability Engine

The MVIE composite blends three structural metrics — Absorption Ratio (eigenvalue concentration of the correlation matrix), Instability Score (mean change in pairwise correlations) and Priority Drift (MST centrality rank drift). Each is measured on the latest 60 trading sessions, standardised against its own previous 252 sessions and mapped to 0–100:

Composite formula · pseudocode

# Pseudocode — MVIE composite (one rolling 60-session window)# C_t = Spearman correlation of the latest 60 sessions, C_b = the 60 beforeAR = sum(top_k_eigenvalues(C_t)) / trace(C_t)     # k = max(1, ceil(0.20 * n))IS = mean(abs(C_t - C_b) over i != j)PD = footrule(rank(eigencentrality(MST(C_t))),              rank(eigencentrality(MST(C_b))))    # normalised to [0, 1]# Each component against its own previous 252 sessions (min 63), never itselfz = (x_t - mean(x[t-252 .. t-1])) / std(x[t-252 .. t-1])s = clip(50 + z * 50/3, 0, 100)MVIE = (0.40*s_AR + 0.35*s_IS + 0.25*s_PD) / (sum of available weights)

Regime thresholds:

  • 0–60 GREEN — Stable absorption / correlation structure; centrality ranks steady
  • 60–80 AMBER — PCA concentration or correlation drift rising; MST hub order shifting
  • 80–100 RED — Systemic stress band; correlation breakdown and centrality drift elevated

2. Network Topology Analysis

Spearman correlation builds a Minimum Spanning Tree (d_ij = 1 − |ρ_ij|, so strong negative correlation also counts as close). Eigenvector, betweenness and degree centrality on the MST quantify systemic hubs; Priority Drift tracks how the eigenvector-centrality ranking moves from the previous 60 sessions to the latest 60:

  • Eigenvector centrality — Hub assets whose stress propagates through the MST (e.g., SPY)
  • Betweenness centrality — Bridge nodes linking otherwise disconnected clusters (contagion paths)
  • Degree centrality — Direct connectivity count on the MST

3. Signal Points and Scenario Ranges

Each asset gets a points-based label from three technical indicators and a range calibrated on past 21-session moves. The composite score enters only through the veto:

Signal points · engine rules

Signal points (per asset):  RSI:              3  MACD:             2  Bollinger Bands:  2  Maximum:          7Label:     BULLISH or BEARISH when that side has more points and at least 3,           otherwise NEUTRALStrength:  points / 7 (a count, not a probability)Range:     68% / 90% band over the next 21 sessions, multipliers measured           per asset class on past 21-session movesVeto:      RED state (MVIE ≥ 80) turns BULLISH into NEUTRAL

4. Veto Protocol

The Veto Protocol is a model override rule. When the MVIE state is RED (score ≥ 80), every BULLISH label becomes NEUTRAL, regardless of indicator strength. The state is the only trigger. In crisis regimes the engine treats single-asset indicator readings as high-noise.

5. LLM Analyst Layer

A large language model layer turns the quantitative output into a structured written summary. 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.

Transparent Quant · No Black Boxes

Every number traces back to a published formula. Not a black box.

MVIE Pro reads markets with transparent mathematics — eigenvalue decomposition, changes in pairwise correlation, and Minimum Spanning Tree centrality. The LLM layer turns the quantitative output into a plain-language written summary. You can inspect every Z-score, every weight, every threshold.

MVIEComposite MVIE Score

MVIE = 0.40·s_AR + 0.35·s_IS + 0.25·s_PD

Weighted mean of three 0–100 component scores. Each component is first standardised against its own previous 252 sessions (z), then mapped to 0–100: s = clip(50 + 50·z/3, 0, 100).

ARAbsorption Ratio

AR = Σλ(top 20%) / Σλ(all)

Share of total variance in the top 20% of eigenvalues (at least one) of the Spearman correlation matrix. Correlation, not covariance, so one volatile asset cannot dominate it. High AR → correlated, fragile regime.

ISInstability Score

IS = mean_{i≠j} |C_t[i,j] − C_b[i,j]|

Mean absolute change in the pairwise Spearman correlations between the latest 60 sessions and the 60 before them. Averaging over pairs keeps it comparable across portfolio sizes.

PDPriority Drift

PD = Σ |rank_t(i) − rank_b(i)| / ⌊n²/2⌋

Spearman footrule between the eigenvector-centrality rankings on the Minimum Spanning Tree of the latest 60 sessions and of the 60 before them, scaled to 0–1. Measures leadership rotation in the asset network.

λEigenvalue Decomposition→ZZ-Score Normalization→ΣWeighted Composite

⚠ All outputs are educational scenario descriptions, not investment advice. The LLM layer is compliance-locked — no price targets, no recommendations.

Traditional approach vs. MVIE Pro

Traditional Risk ToolsMVIE Pro
One correlation snapshotRolling 60-session window, compared with the 60 sessions before it
Static risk metricsDynamic MVIE score (0-100) with regime classification
Ignores asset network effectsMinimum Spanning Tree topology + centrality metrics
Buy/sell labels with no range or contextPoints-based labels with a 68% / 90% range calibrated on past 21-session moves
No macro override mechanismVeto Protocol — macro risk overrides micro signals
Black-box ML modelsTransparent mathematics — every Z-score is auditable