İçeriğe geç
MVIE Pro
Back to Research
Quantitative Research

MVIE Validation Study 01: What Has Been Measured So Far

MVIE Pro ResearchObservational — Not Investment Advice11 min readShare

MVIE Validation Study 01: What Has Been Measured So Far

Publication Date: 18 November 2026 Classification: Observational Research — Not Investment Advice

This post reports measurements. It does not say that the score describes market stress. Under the study's pre-registration, only the forward test (Section 6) can support such a sentence, and it has not run yet.

1. Why this post replaces an earlier one

A post published here earlier in 2026 was withdrawn in October. Its sample table, alert thresholds and rates were illustrative or synthetic, and no code or data traced them. This post reports measured numbers instead. Every number below comes from the study's pre-registration and its run records.

2. What is tested, and how

What is not tested. The score at a session says how unusual the 60-session correlation structure of the macro universe is, compared with the same measure over the previous 252 sessions. It does this through three components, weighted 0.40 / 0.35 / 0.25:

  • concentration (AR);
  • change against the previous window (IS);
  • change in the centrality ranking (PD).

That is the formula itself, and code checks it.

What is tested. The name "instability" and the GREEN / AMBER / RED labels invite a reading: that the score's highest values fall in periods the market called stressful. The study tests only that reading. It tests no forecast and no direction.

The method was fixed before any data was loaded. The pre-registration was frozen on 2026-11-14.

  • Events are dated by an official primary document: a central-bank decision, a government declaration or a regulator's announcement. Each event's window is the 60 sessions starting on its date.
  • Ordinary windows are the 60-session windows of the same period that stay clear of every listed event, from 20 sessions before it to 79 sessions after it.
  • M: each event window's peak score is ranked against the ordinary windows' peaks, and M is the mean of these percentiles. M = 0.5 means the event windows peaked like ordinary ones; M = 1.0 means every event window peaked above every ordinary window.
  • p comes from a permutation test: 10,000 draws of non-overlapping ordinary windows. The smallest possible p is 0.0001.
  • The 95% interval is a bootstrap over the events, and it is coarse when there are few of them.
  • Two simple measures get the same statistic, to show what the score adds: VIX-z, the z-score of the VIX level against its own past year, and mean |ρ|-z, the z-score of the mean absolute Spearman correlation over all asset pairs.

3. Blind historical evidence, low power (2008–2016)

Before the study, the score had never been computed for a session before May 2016, so this period was blind to the design.

  • It covers 2008-08-20 to 2016-04-29 with six assets: the macro universe without Bitcoin, which has no history that far back.
  • It was run once, on 2026-11-16, and is never run again.

Five events qualified:

DateEventSource
2008-09-15Lehman Brothers files for Chapter 11secondary
2010-05-03IMF and euro-area programme for Greeceprimary
2011-08-08S&P downgrades the United Statessecondary
2013-05-22Bernanke's testimony to the Joint Economic Committeeprimary
2015-08-11The People's Bank of China changes the renminbi fixingsecondary

Secondary: the primary document could not be read, so the date comes from two independent secondary sources.

MeasureMp95% interval
MVIE score0.9340.00010.929 – 0.940
VIX-z0.9530.00010.881 – 1.000
Mean ∣ρ∣-z0.5400.3650.254 – 0.794
MVIE score, the two primary-sourced events only0.9350.00540.928 – 0.941

The ordinary windows:

  • there are 1,145 of them;
  • the five events' percentiles are 0.941, 0.941, 0.928, 0.928 and 0.932;
  • the ordinary windows overlap, so their peaks repeat. That is why the percentiles sit close together and the interval is narrow.

The RED threshold:

  • no event window's peak reached 80, while 4.9% of the ordinary windows' peaks did;
  • inside the event windows, 60–77% of sessions scored 60 or more, against 10.3% in ordinary windows;
  • no session in an event window scored 80 or more (ordinary windows: 0.08%).

Lehman: the peak stayed below 80. The pre-registration named this case in advance: a turbulent previous year (BNP Paribas, Bear Stearns) makes the peak smaller, because the score measures against its own past year.

Against a financial stress index:

  • the index is the weekly St. Louis Fed Financial Stress Index (STLFSI4), with VIX-z held fixed;
  • partial Spearman correlation 0.437 (0.291 – 0.567), over 402 weeks;
  • the index is revised after the fact and contains the VIX, so this is descriptive only.

A prototype's rows for April 2016: a database from an earlier prototype (a 15-asset universe) held rows for that month. With the ordinary windows touching it removed, M is 0.933 and p is unchanged.

What these numbers are not. Five events are few. Under the pre-registration, a large M and a small p at five events count only as suggestive. They get no label and change nothing in the product.

4. In-sample development test (2016–2023)

This period was seen before the study, and the design was tuned on it: early in 2026, settings were tried until COVID-19 read RED and 2017–2018 read calm. Its numbers check the analysis code end to end; they are not evidence.

Bitcoin's history on the study's data source starts on 2017-08-28, so seven-asset scores with a full year of history start on 2019-02-21. The period was therefore also run with six assets.

7 assets6 assets, same span6 assets, whole period
Span2019-02-21 – 2023-08-312019-02-21 – 2023-08-312016-05-02 – 2023-08-31
Events / ordinary windows5 / 3675 / 3676 / 955
MVIE score: M (p)0.926 (0.0001)0.817 (0.0001)0.794 (0.0008)
VIX-z: M (p)0.779 (0.0001)0.779 (0.0001)0.671 (0.031)
Mean ∣ρ∣-z: M (p)0.772 (0.0001)0.667 (0.0001)0.586 (0.147)
Event windows peaking at 80 or more2 of 51 of 51 of 6
Ordinary windows peaking at 80 or more0%18.0%9.1%
STLFSI4 partial correlation0.2720.2410.211

The events:

  • 2019-08-05: the US Treasury designates China a currency manipulator;
  • 2020-01-30: the WHO declares COVID-19 a public health emergency of international concern;
  • 2022-02-24: Russia's invasion of Ukraine;
  • 2022-09-23: the UK "Growth Plan";
  • 2023-03-10: Silicon Valley Bank is closed;
  • the whole-period column adds 2016-06-24, the Brexit referendum result (secondary source).

One period behind the 7-asset column: its ordinary windows come almost entirely from 2020-07 to 2021-10. That column compares the events with that one period, not with a broad range of ordinary times.

5. The bands and the dead band (2023–2026)

60 and 80 are not fitted to data: in the score's formula they correspond to a weighted +0.6σ and +1.8σ above the past year's average. They had been calibrated on Yahoo data.

The study re-measured them on licensed data, on six portfolio shapes over 752 sessions each (2023-10-02 to 2026-09-30). The four conditions were set in advance:

  • RED in 2–5% of sessions;
  • AMBER in 10–25%;
  • a median GREEN run of at least 4 sessions;
  • at most 60 state changes per shape in three years.

This step looks only at how the score is distributed, never at events.

Plain thresholds did not meet the conditions.

  • RED 2.17% and AMBER 19.9% were inside the limits.
  • But the median GREEN run was 3.0 sessions, and two shapes changed state 62 and 67 times.

The options compared, on the same scores:

60/8065/8060/80, dead band 2dead band 3dead band 5
RED / AMBER share2.17% / 19.9%2.17% / 12.3%2.44% / 21.7%2.70% / 22.4%3.01% / 24.5%
Median GREEN run (sessions)3.04.06.511.020.5
State changes, six shapes299238209179129
Delay of a downward change, median / max (sessions)——0 / 30 / 61 / 10
Sessions labelled unlike 60/800%7.6%2.4%3.5%6.3%
Meets the conditionsnoyesyesyesyes

Chosen: 60/80 with a dead band of 2 points (2026-11-15).

  • The state still goes up at 60 and 80. It comes down only below 58 (AMBER to GREEN) and 78 (RED to AMBER).
  • It meets all four conditions with the smallest cost: the thresholds stay where they were, and a downward change comes at most 3 sessions late.
  • The choice was made before either historical test above was run, and no band was set from an event result.

The cost: the state now depends on the sessions before it.

  • A classification started at any session of the window agreed with the full history within 3 sessions.
  • The same input still gives the same score and the same state.
  • The weekly engine self-check now watches this on the sample portfolios.

The limits of this step:

  • the conditions were set knowing the old Yahoo result;
  • the comparison with that result changes four things at once: the data source, the trading calendar, the portfolio shapes, and price instead of total returns.

6. The forward test

The study's verdict comes only from a forward test on events after 2026-11-14, the day the pre-registration was frozen.

Universe: the seven assets of the Global Macro score, with a six-asset run alongside for comparison with Section 3.

Events: only those matching a closed list of types written before the test began:

  • an unscheduled FOMC decision on the federal funds target;
  • a new Federal Reserve emergency facility under section 13(3), or a new swap line;
  • the failure of a US bank with $50 billion or more in assets, or a systemic-risk exception;
  • a downgrade of the US long-term rating;
  • the resolution, or a publicly supported forced merger, of a global systemically important bank;
  • a new, unscheduled programme or facility of another G7 central bank that cites market functioning or financial stability;
  • a WHO pandemic characterisation;
  • a US bankruptcy filing by a crypto platform that holds customer assets, with $1 billion or more in liabilities.

When: once, when six such events have occurred. The 2016–2026 record of these types gives about 0.64 a year, so this is expected to take nine to ten years, give or take about four.

Labels, fixed in advance:

  • DESCRIBES: p < 0.05 and M ≥ 0.75;
  • WEAK: M ≥ 0.60 otherwise;
  • INDISTINGUISHABLE: M < 0.60.

Only DESCRIBES allows a sentence about what the score describes. That sentence gives the score's M together with VIX-z's and mean |ρ|-z's, the number of events, the intervals, the universe and the period. It uses no forecasting language.

The result is published here, whatever the label.

The withdrawn post's claim about Priority Drift is tested only there, as the study's secondary test.

7. Limits

  • The 2008–2023 event list was chosen knowing market history; most of its events are official responses to market moves.
  • The blind period is not the product's universe: it has no Bitcoin.
  • Five events give low power. The ordinary windows overlap, so the percentiles are not independent; the permutation test accounts for this only in part.
  • Three of the five blind-period events were dated from secondary sources.
  • The 60-session window is fixed, so a response that comes later falls outside it.
  • The forward test's event types are fixed, but its record is kept while the live score is visible, so the record is not blind to the score.

Data

  • Prices: Twelve Data, split-adjusted daily closes, kept in a separate research store. Only derived statistics are shown here: no price, return or index value.
  • VIX: Chicago Board Options Exchange, CBOE Volatility Index: VIX [VIXCLS], retrieved from FRED, Federal Reserve Bank of St. Louis.
  • Financial stress: Federal Reserve Bank of St. Louis, St. Louis Fed Financial Stress Index [STLFSI4], retrieved from FRED, Federal Reserve Bank of St. Louis.
  • Event dates: the issuing institutions' own documents (the Federal Reserve, the US Treasury, the FDIC, the IMF, the ECB, the WHO, the UK government, the Bank of England, FINMA), or two secondary sources where marked.

This document is produced by MVIE Pro for educational purposes only. Not investment advice.