Market Risk Scan
MarketRiskScan · Public Research Evidence

What the research supports—and what it does not.

This page summarizes MarketRiskScan’s historical validation work through Phase 18.3. It is intentionally conservative: retrospective reconstruction, point-in-time evidence, failed hypotheses and publication limits are kept separate rather than blended into one performance claim.

Current research position

Evidence update
September 6, 2026
Covers the frozen v1.2 historical work through Phase 18.3.
Public interpretation
Relative vulnerability, not probability
The live model may describe risk state and qualitative downside vulnerability without claiming an exact event probability.

Executive conclusions

Supported

The Market Risk Score is a relative risk-state score, not a probability.

A score of 58 means the model is in a higher relative vulnerability state than lower-scoring conditions. It does not mean a 58% chance of a correction, recession or bear market.
Retrospective only

Higher full-model risk states show meaningful retrospective downside association.

In the frozen 2000+ reconstruction, High+ states had higher 6-month rates of 10%, 15% and 20% downside than Below-High states. The association strengthened with downside severity, but the numerical rates remain retrospective and PIT-blocked.
Not validated

Calibrated crash or correction probabilities are not supported for publication.

Walk-forward probability calibration did not provide sufficiently compelling, stable improvement to justify publishing absolute event probabilities such as “31% chance of a 15% decline.”
Robust

The available strict-PIT score is a reduced diagnostic subset, not the full frozen model.

The strict-PIT inventory represented only 6 of the 16 frozen-core indicators. Architecture-completeness and score-equivalence gates failed, so that reduced subset cannot be treated as a strict-PIT validation of the complete v1.2 model.
Research diagnostic

Missing architecture, rather than shared-indicator vintage alone, explains most strict-PIT High-state misses.

The crossed architecture × vintage audit found that restoring the missing architecture was the dominant explanation for the original High-state false negatives. Vintage effects were real but secondary.
Not validated

A complete historical 16/16 publication-safe reconstruction remains blocked by CAPE vintage lineage.

Most missing indicators have plausible direct-vintage or official archive paths, but a complete authoritative month-by-month as-published Shiller CAPE lineage was not established. MarketRiskScan therefore does not claim full historical exact-vintage validation.
Supported

Qualitative Downside Risk Outlook labels are the appropriate public interpretation.

MarketRiskScan may describe 10%, 15% and 20% downside vulnerability using ordinal labels such as Very Low, Low, Moderate, Elevated, High and Very High, provided they are clearly described as relative vulnerability—not calibrated probability.
Supported

Data Quality describes input freshness and completeness, not forecast certainty.

The operational quality percentage reflects the condition of data feeding the model. It must not be interpreted as the model’s percentage confidence that the market will rise or fall.

Retrospective 6-month downside association

The frozen 2000+ full-model reconstruction showed a clear directional relationship between High+ states (score ≥51) and subsequent downside. These figures are research evidence only: they are not point-in-time calibrated probabilities for today’s market.
Downside event Below High High+ Retrospective rate ratio Episode support
10% correction 18.0% 43.8% 2.43× 7 distinct / 4.09 effective
15% decline 9.5% 31.3% 3.29× 7 distinct / 4.33 effective
20% bear market 4.0% 18.8% 4.69× 6 distinct / 3.83 effective
Evidence-class warning: these numerical rates come from retrospective reconstruction. They remain PIT-blocked and should not be presented on the Dashboard as absolute probabilities. Their appropriate use is to support the qualitative statement that higher model states have historically corresponded to greater downside vulnerability.

Why the Dashboard uses qualitative Downside Risk Outlook labels

The product deliberately stops short of false precision. Members can view 10%, 15% and 20% downside vulnerability over 1-, 3-, 6- and 12-month model horizons using a versioned ordinal scale.
A label such as High means MarketRiskScan detects a high relative vulnerability state for that horizon and downside severity. It does not mean a specified percentage chance that the decline will occur.

Research milestones

Research phase Finding Status What was found Interpretation
Phases 3–14C Historical framework, vulnerability/stress decomposition and PIT-vs-revision audits Mixed / conditional Established the historical reconstruction framework, separated retrospective from stricter point-in-time evidence, and showed that no single component should be marketed as a universal crash predictor. Supports a multi-dimensional risk-state framework and strict evidence governance.
Phase 17A Longer historical sample and multi-horizon downside outcomes Retrospective only Extended the frozen architecture into the 1990s and a lower-confidence 1980s sensitivity period without changing the production model. The full 2000+ core remained the primary research sample. Broadened episode coverage while keeping lower-quality historical reconstruction clearly separated.
Phase 17A.8 High-state warning-spell and control analysis Retrospective only High+ states were associated with increased subsequent downside, but long warning spells and false alarms showed that the score should not be presented as a crash timer. Supports vulnerability/warning-state language rather than precise timing claims.
Phase 17A.9 Prior-3-month score direction candidate Not validated Adding prior-3-month change in the 12-month score worsened walk-forward performance and produced adverse coefficient direction. The candidate was rejected and no direction-window search was continued.
Phase 17B Probability calibration Not validated No primary outcome cell met the pre-specified standard for statistically compelling, decision-relevant calibration improvement. Do not convert the 0–100 score into absolute event probabilities.
Phase 18 Publication-safe claim audit Robust Confirmed that changing a retrospective statistic from probability to rate, multiple or band frequency does not upgrade its evidence class. Retrospective numerical rates remain research context, even when the qualitative direction is useful.
Phase 18.1 Exact-vintage High-state confirmation gate Not validated The stored strict-PIT subset covered only 6/16 frozen-core indicators; High-state agreement and sensitivity gates also failed. The reduced score is not an exact-vintage validation of the full frozen architecture.
Phases 18.1a–18.2a False-negative and crossed architecture × vintage attribution Research diagnostic Missed full-model High states were themselves risky. Crossed testing showed missing architecture was the dominant source of the mismatch and remained dominant after controlling the shared-six vintage treatment. Explains why the reduced strict-PIT score attenuated the full-model retrospective relationship without treating the result as new predictive validation.
Phase 18.3 Publication-safe source feasibility Not validated Direct-vintage or official archive paths were identified for many missing indicators. A complete as-published CAPE lineage was not established, so the 16/16 historical Standard 1 gate remains blocked. End the archival reconstruction effort rather than weaken the evidence standard; build pristine prospective history going forward.

Public claims governance

What MarketRiskScan may safely say

  • MarketRiskScan summarizes multiple dimensions of U.S. equity-market vulnerability into a transparent relative risk framework.
  • The Market Risk Score is a relative 0–100 risk-state score; it is not the percentage probability of a crash, correction, recession or bear market.
  • Higher scores indicate a more adverse or fragile combination of modeled conditions within the MarketRiskScan framework.
  • Retrospective reconstruction found that High+ states were associated with higher subsequent downside-event rates, particularly for more severe 6-month declines.
  • The Downside Risk Outlook may use qualitative labels such as Very Low, Low, Moderate, Elevated, High and Very High to describe relative vulnerability.
  • Those qualitative labels are not calibrated probabilities and do not identify the exact timing of a future decline.
  • Data Quality describes freshness, completeness and coverage of the model inputs.
  • MarketRiskScan distinguishes retrospective reconstruction from stricter point-in-time evidence and preserves negative research results.
  • MarketRiskScan is a research and risk-assessment tool, not individualized investment advice.

What MarketRiskScan should avoid

  • “MarketRiskScan predicts crashes.”
  • “A score of 70 means a 70% chance of a market decline.”
  • “High means a specific percentage probability that the market will fall.”
  • “Backtesting proves the model works.”
  • “The full v1.2 architecture has been completely validated with exact-vintage historical data.”
  • “The model successfully predicted the 2000, 2008 or 2020 crashes.”
  • Any exact correction or bear-market probability unless a separately calibrated, point-in-time model is later validated.
  • Any wording that turns Data Quality into confidence that the market will rise or fall.

Prospective validation going forward

MarketRiskScan now has a cleaner path than continued historical data archaeology: archive every live production snapshot with the exact inputs, observation dates, publication/vintage information, transformed indicator scores, category scores, horizon scores and model version. That prospective record will become the strongest future evidence because it captures what the model actually knew and calculated at the time.
Future changes to weights, risk bands or qualitative downside mappings should be versioned and documented rather than silently rewritten after outcomes are known.
See how the model reads today’s market. Methodology and research are public. Registration is required for the live Dashboard and Daily Report.
Create Account