A Systemic-Risk Monitoring Framework for High-Valuation Markets

Working paper · May 12, 2026

Abstract

This paper presents a systemic-risk monitoring framework for global financial markets as of May 12, 2026. It distinguishes (i) discrete trigger events, (ii) slow-moving state amplifiers, (iii) downstream outcomes, and (iv) feedback-loop transmission pathways, and translates this taxonomy into an operational monitoring dashboard.

The framework's strongest empirical screen is regime-conditional history. Using Robert Shiller's monthly CAPE series through September 2023, extended through May 2026 with Multpl/GuruFocus monthly readings, only two distinct episodes have reached CAPE ≥40 in 145 years: January 1999 - September 2000 (peak 44.19 in December 1999) and the current March 2026 onward episode. The 1999-2000 episode contributes 21 monthly starting points; 13 of those 21 monthly windows experienced a peak-to-trough ≥25% drawdown within the subsequent 24 months — a rolling-window coverage rate of 61.9%. Because these 21 monthly observations are highly overlapping within a single completed historical episode (n = 1 at the episode level), the 61.9% figure is a descriptive historical stress signal, not a calibrated independent-event probability or confidence-bounded forecast. The paper does not report a confidence interval for this regime: as documented in Appendix K, a stationary block-bootstrap on this single-cluster sample produces a degenerate distribution (the 5th percentile for the broader CAPE ≥35 regime sits above the point estimate) and does not have a frequentist confidence-interval interpretation.

Several state amplifiers are elevated simultaneously: Shiller CAPE in a provider/date range of approximately 40.3–42.2 in early May 2026 (GuruFocus May 1; Multpl May 11); Buffett Indicator 231.7% on May 6, 2026 (Longtermtrends/MacroMicro); official SSGA SPY top-10 weight 39.28% and Magnificent 7 share-class aggregate 34.69% as of May 8, 2026; AI capex shifting from operating cash flow to debt financing (BIS Bulletin 120, published January 7, 2026; BIS Quarterly Review follow-on, March 2026); approximately $14 trillion in 2026 US investment-grade bond supply per Apollo's Torsten Slok (March 24, 2026); private-credit redemption pressure across non-traded BDCs (BCRED, BlackRock HPS, Morgan Stanley North Haven, Cliffwater funds in Q1 2026; Moody's BDC sector outlook cut to negative on April 7, 2026); and an oil shock characterized by the IEA as "the largest disruption in history" — a 10.1 mb/d global supply fall in March 2026 (IEA Oil Market Report, April 14, 2026).

The paper retains two transparent scenario-generation layers, regenerated from a single frozen parameter file as of May 12, 2026 (seed 42, 200,000 trials). Layer A is a factor-copula trigger co-occurrence simulation across eight discrete trigger pathways. Layer B is a structured second-stage severity simulation. Base-case scenario-generator outputs:

  • Layer A: P(≥1 modeled trigger fires within 24 months) = 76.7%; P(≥2) = 54.6%; P(≥3) = 38.3%; expected trigger count = 2.24.
  • Layer B: P(trigger-mediated drawdown ≥25% within 24 months) = 49.9% at slope = 6 base calibration; slope-sensitivity range 16.7% (slope=4) to 67.9% (slope=8); P(≥40%) = 10.4% at base.

These are conditional scenario outputs given a frozen parameter set, not empirically calibrated forecasts. Trigger marginals, factor loadings, severity weights, policy-effectiveness distributions, and the slope parameter mapping severity to drawdown percentages are all author priors anchored to external evidence but not estimated from history. The model layers are useful for channel identification (which transmission pathways are most worth monitoring), trigger-conditional uplift (e.g., iran_hormuz firing raises basis_trade probability 1.53×; ai_capex firing raises taiwan probability 1.80×), and slope sensitivity. They are not estimates of true market odds.

The dashboard layer (Section 12) translates the framework into operational monitoring signals. A pre-specified rule of VIX ≥25 sustained for 5 trading days was backtested against eight historical drawdown episodes 1995-2024 using FRED VIXCLS data. The rule detected 3 of 8 episodes pre-peak — Dot-com 2000 (157 days), GFC 2008 (55 days), 2022 cycle (28 days) — with 38 distinct false-positive events outside episodes over 30 years (≈1.27/year). The rule misses sentiment-driven (Q4 2018), exogenous (COVID 2020), fast-acting (August 2024), and pre-1995-style (LTCM 1998, Europe 2011) drawdowns. Threshold sensitivity (22.5/25/27.5/30) and duration sensitivity (3/5/7/10 days) are reported in §12. These are preliminary backtest results on a pre-specified sample, not validation across the universe of all S&P 500 drawdowns above a fixed threshold.

The framework's appropriate use is stress-testing portfolio concentration, liquidity, private-credit exposure, and oil/Fed-correlation risk under adverse drawdown scenarios; not directional positioning advice and not a calibrated crash forecast. The author is not a registered investment adviser; readers should consult appropriate fiduciaries for portfolio decisions. The framework is symmetric: §13 specifies both the conditions under which tail-risk concern should be reduced and the conditions under which the bull case should be upgraded.

Reading Guide

Section What it answers
§1 The motivating configuration: records and low implied volatility against multi-dimensional stress
§2 Verified market, valuation, concentration, and macro anchors as of May 8-12, 2026
§3 Taxonomy: triggers vs. amplifiers vs. outcomes — definitions and the framework diagram
§4 Five amplifiers (AI capex debt shift, IG supply, concentration, valuation, CRE wall)
§5 Eight discrete triggers with current-state evidence and marginal probabilities
§6 Six feedback-loop transmission pathways
§7 Probability framework: Layer C (history), Layer A (scenario-generator co-occurrence), Layer B (scenario-generator severity)
§8 Policy reaction functions and trigger-specific effectiveness
§9 Historical analog review
§10 September 2026 - February 2027: a catalyst-dense monitoring window (heuristic)
§11 Three conditional severity paths (qualitative scenario weights)
§12 Leading-indicators dashboard (selected-sample backtest results)
§13 Falsification, bull-case upgrade criteria, Brier scoring, auditable forecast log
§14 Conclusion
Appendices A-K Capex reconciliation, evidence anchors, code, source register, reproducibility manifest, statistical caveats

1. Introduction

The starting point matters. As of Friday May 8, 2026, the S&P 500 closed at 7,398.93 — a record close, up 0.84% on the day and the sixth consecutive weekly gain. The VIX closed at 17.19, below its long-term average. WTI crude settled at $95.42; Brent at $101.29. Hyperscaler Q1 2026 earnings had beaten consensus across the Big Four (Microsoft, Alphabet, Amazon, Meta), and Wall Street strategists raised year-end targets through May 11: JPMorgan (Dubravko Lakos-Bujas) to 7,600 on April 21; RBC Capital Markets (Lori Calvasina) to 7,900 on May 8; HSBC (Nicole Inui) to 7,650 on May 11; Yardeni Research (Ed Yardeni) to 8,250 on May 11. Bitcoin traded near $80,000 and gold near $4,700/oz.

This is true while:

  • The IEA reports a 10.1 mb/d global oil supply fall in March 2026 — characterized in its April Oil Market Report as "the largest disruption in history."
  • The Strait of Hormuz has been substantially closed since late February 2026; pre-war crude+condensate transit of ~20 mb/d fell to ~2-4 mb/d in March-April 2026.
  • IEA Director Fatih Birol stated on Norges Bank's In Good Company podcast (reported by CNBC April 1, 2026) that 12 mb/d of oil supply is lost — "more than two of these oil crises [1973 and 1979] put together"; on April 13, 2026 at the Atlantic Council he revised this figure to about 13 mb/d.
  • The Bank of Japan held its policy rate at 0.75% on April 28, 2026 in a 6-3 vote, with three dissenters (Hajime Takata, Naoki Tamura, Junko Nakagawa) proposing a rise to 1.0%.
  • Powell's term as Fed Chair ends May 15, 2026; as of publication time, Kevin Warsh's Senate confirmation vote was scheduled for 11:30 a.m. ET on May 12, 2026, with cloture invoked 49-44 the prior day.
  • Trump rejected Iran's latest peace counter-proposal as "TOTALLY UNACCEPTABLE" on May 10, 2026; WTI rose ~2% on May 11.

This is the central configuration the paper examines. The market prices relief and low conditional risk against a backdrop of unresolved structural stress on multiple dimensions. Either (a) something has materially changed about productivity, asset demand, or policy capacity such that elevated state variables are less threatening than history would suggest, or (b) the conditional asymmetry of outcomes is being mispriced and a catalyst that engages two or more feedback loops simultaneously could produce a sharp move down.

The honest counterweights are real. Q1 2026 earnings came in strongly: FactSet's "Earnings Insight" (Butters, May 8, 2026) reports 84% of S&P 500 companies have beaten EPS estimates with an average surprise of 18.2% — both above 5- and 10-year averages and the highest since Q1 2021. Forward 12-month P/E is 21.0 (FactSet, May 8), above the 5-year average 19.9 and the 10-year average 18.9. Bridgewater's Greg Jensen (February 23, 2026) estimates Big Tech 2026 AI capex at approximately $650B (~$410B in 2025). A reasonable bull thesis exists and is reflected in market pricing.

This paper should be read as a fragility framework. It does not claim that valuation, oil, AI capex, or leverage independently cause a crash. It asks whether the current configuration increases the market's sensitivity to adverse catalysts, identifies the most likely transmission channels, and translates this into operational monitoring signals. The framework's central empirical claim — that the only previous time markets reached the CAPE ≥40 regime, the subsequent path was a substantial drawdown across nearly every monthly entry point in that episode — is single-episode evidence (n = 1 at the episode level) with all the limitations that entails.

2. Current Market State

All values are timestamped to specific providers; intraday quotes are flagged. Full source-register entries appear in Appendix I; the reproducibility manifest appears in Appendix J.

2.1 Index levels (May 8-11, 2026)

Indicator Value Source Date/Time
S&P 500 7,398.93 (close) FRED SP500; CNBC live blog May 8, 2026 (Fri)
S&P 500 (intraday) ~7,412 record high CNBC May 11, 2026
Nasdaq Composite 26,247.08 (close) TheStreet/Nasdaq May 8, 2026
Russell 2000 2,861.21 (close) TheStreet/AP May 8, 2026
Dow Jones Industrial 49,609.16 (close) TheStreet May 8, 2026
VIX 17.19 (close) FRED VIXCLS; Cboe May 8, 2026
VIX (intraday) ~18.41 Yahoo Finance May 11, 2026
WTI front-month $95.42 (settle) CNBC; CME May 8, 2026
WTI (intraday) ~$97-98 TradingEconomics May 11, 2026
Brent front-month $101.29 (settle) CNBC; ICE May 8, 2026
Brent (intraday) $100-104 range TradingEconomics May 11, 2026
Bitcoin (spot) ~$80,800-82,200 Fortune; Yahoo Finance May 11, 2026, 9:15 a.m. ET
Gold (spot) ~$4,700-4,747 Fortune; USAGOLD May 11, 2026
10-yr Treasury 4.38% (close) Fed H.15; FRED DGS10 May 8, 2026
MOVE Index 67.25 (intraday) Yahoo Finance/ICE BofAML May 11, 2026
ICE BofA HY OAS 2.81% (close, May 8); 2.79% (close, May 7) FRED BAMLH0A0HYM2 May 2026

This is a market that has fully recovered from the March-April 2026 Iran-war low and is making new highs, while the underlying oil shock remains the largest in IEA records.

2.2 Valuation

Metric Value Historical reference Source
Shiller CAPE provider/date range: 40.33 (GuruFocus, May 1, 2026) to 42.15 (Multpl, May 11, 2026) All-time high 44.19 (Dec 1999); historical mean ~17.8 (1881-2026 monthly) Multpl; GuruFocus; YCharts; MacroMicro
Buffett Indicator 231.7% (May 8, 2026) All-time high ~230.3% (Jan 2026); ~2.4 σ above long-run trend GuruFocus; MacroMicro; Current Market Valuation
Forward 12-month P/E 21.0 (May 8, 2026) 5-yr avg 19.9; 10-yr avg 18.9 FactSet Earnings Insight, May 8, 2026
Excess CAPE Yield (10-yr real expected return proxy) ~1.3-2.0% range — Author calculation from CAPE

The 231.7% Buffett Indicator print is effectively a new all-time high relative to the prior January 2026 peak of ~230.3%. CAPE ≥40 has occurred in only two distinct episodes in 145 years of Shiller's monthly data: January 1999 - September 2000 and March 2026 - present. The decisive empirical link between CAPE level and drawdown frequency comes from §7.3.

2.3 Concentration

Concentration is measured from official SSGA SPY holdings rather than aggregator dashboards. As of May 8, 2026:

Rank Holding Weight in SPY
1 NVIDIA 8.24%
2 Apple 6.78%
3 Microsoft 4.86%
4 Amazon 4.20%
5 Alphabet Class A 3.68%
6 Broadcom 3.21%
7 Alphabet Class C 2.93%
8 Meta 2.10%
9 Tesla 1.90%
10 Berkshire Hathaway B 1.38%
Top 10 total 39.28%

The Magnificent 7 share-class aggregate (NVIDIA + Apple + Microsoft + Amazon + Alphabet A/C + Meta + Tesla) is 34.69% of SPY as of May 8, 2026. This is the figure used throughout the paper, replacing the earlier Motley Fool 33.7% reference (April 14, 2026, retained only as dated context).

Top 10 in Russell 1000 Growth: approximately 56-60% in early 2026 (FTSE Russell factsheet).

The arithmetic amplification is real but bounded: a 10% drawdown limited to the Magnificent 7 share-class aggregate alone would mechanically subtract approximately 3.47 percentage points from SPY before second-round effects. Cap-weighted index funds do not mechanically force-sell on price declines; weights simply adjust, and selling occurs only with investor redemptions, vol-targeted deleveraging, leveraged-position margin calls, and active rebalancing. The dashboard (§12) distinguishes mechanical index arithmetic from flow-driven second-round effects.

2.4 Macro indicators

  • April 2026 nonfarm payrolls (BLS, May 8, 2026): +115,000; unemployment unchanged at 4.3%; AHE +0.2% m/m, +3.6% y/y. February revised to -156k; March revised up to +185k. Healthcare +37k, transportation/warehousing +30k, retail +22k, information -13k.
  • Q4 2025 real GDP, third estimate (BEA, April 9, 2026): +0.5% annualized, revised down from 0.7% second estimate; full-year 2025 +2.1%. BEA notes the October-November 2025 federal shutdown subtracted ~1.0 pp.
  • Inflation: above the Fed's 2% target with oil-driven upside pressure resumed since the Iran conflict began. BOJ's April 28 Outlook Report raised FY2026 core CPI to 2.8% from 1.9%.
  • Fed Chair transition: Powell's term ends May 15, 2026. Kevin Warsh cleared Senate Banking Committee 13-11 on April 29 (CNBC, Al Jazeera); cloture invoked 49-44 on May 11 (Roll Call), with Senators Fetterman and Coons crossing party lines; as of publication time, the final confirmation vote was scheduled 11:30 a.m. ET, May 12, 2026. Powell will remain on the Board of Governors (term through 2028).
  • Bank of Japan, April 28, 2026: held uncollateralized overnight call rate at 0.75% in 6-3 vote. Dissenters Takata, Tamura, and Nakagawa proposed raising to 1.0%. FY2026 GDP cut to 0.5% from 1.0%; FY2026 core CPI raised to 2.8% from 1.9%. Source: BOJ, "Statement on Monetary Policy," April 28, 2026.
  • US federal debt: $38.91 trillion as of May 5, 2026 (Treasury Fiscal Data, "Debt to the Penny"; JEC Debt Dashboard, updated May 8, 2026). Projected to cross $39T around May 18, 2026. Average interest rate on marketable debt 3.373% (April 2026). YoY change +$2.70T. Debt held by the public lower (~$30T); gross debt/GDP ~130%, debt-held-by-public/GDP ~100%.

2.5 The bull case

  • Q1 2026 earnings (FactSet, May 8, 2026): 89% reported; 84% positive EPS surprise rate (highest since Q2 2021); 18.2% aggregate surprise magnitude (highest since Q1 2021's 22.2%); blended revenue growth 11.3%.
  • Forward growth expectations: Q2 EPS +19.9%, Q3 +23.2%, Q4 +20.7%, full-year 2026 +21.0%.
  • Strategist target revisions (April 21 - May 11, 2026): JPM 7,600; RBC 7,900; HSBC 7,650; Yardeni 8,250.
  • AI productivity narrative remains intact in earnings reports; Bridgewater's Greg Jensen (Feb 23, 2026) estimates 2026 Big Tech AI capex contributes ~140 bps to US GDP growth.

The bear case (this paper's principal subject) coexists with these. The framework does not assert the bull case is wrong; it asks whether the conditional asymmetry of outcomes makes additional stress-test attention warranted. The bull case has its own upgrade criteria in §13.

3. Taxonomy: Triggers, Amplifiers, Outcomes

The framework's organizing principle: distinguish discrete events that could fire (triggers), state variables that determine response magnitude (amplifiers), and downstream consequences (outcomes). This distinction matters because the categories are often conflated — for example, treating "valuation extremes" or "regional bank failure" as triggers. Both are wrong: valuation is a state variable that amplifies the response to any trigger; regional bank failure is an outcome that emerges from CRE state plus a credit shock.

Compact diagram:

[ Trigger ] → engages [ Amplifier(s) ] → activates [ Feedback Loop(s) ] → produces [ Outcome ]

Example: AI capex disappointment → engages valuation/concentration/debt-financing →
  activates Loop A (capex-IG supply-yields) → produces sector re-rating + IG spread widening

3.1 Eight triggers

A trigger is a discrete event that could occur within the analysis horizon and that would, by itself, materially reprice at least one market segment.

  1. AI capex / revenue disappointment that materially reprices AI infrastructure equity and credit.
  2. Iran / Hormuz escalation beyond the current state — a kinetic step beyond what is currently priced.
  3. Treasury basis trade unwind — forced deleveraging at one or more large funds with Treasury-market dislocation.
  4. Yen carry unwind — material liquidation (≥15% of relevant positions, approximate threshold) with cross-asset spillover.
  5. Major AI-enabled cyber incident on a globally systemic financial institution or shared infrastructure.
  6. Private-credit semi-liquid fund forced gating, wind-down, or sponsor injection event (BCRED-style but more severe).
  7. Stablecoin de-peg with measurable contagion into money markets or Treasury bills.
  8. Taiwan kinetic escalation — a discrete military step beyond current posture.

3.2 Five amplifiers

Amplifiers are slow-moving state variables that determine how large the response to any trigger will be.

  1. Elevated valuation — CAPE ≈40, Buffett Indicator ≈232%. Low expected forward returns; high downside skew on disappointment.
  2. Index concentration — Magnificent 7 share-class aggregate 34.69%, SPY top 10 39.28% (official SSGA, May 8, 2026).
  3. Treasury / IG supply pressure — approximately $14T 2026 IG supply per Apollo/Slok.
  4. AI capex debt-financing shift — BIS Bulletin 120 (published January 7, 2026) documents leading tech firms increasingly using external financing; couples AI narrative risk directly to corporate-credit markets.
  5. CRE refinancing wall — $875B (17% of $5.0T outstanding) maturing in 2026 per MBA (February 9, 2026).

3.3 Three downstream outcomes (not triggers)

  1. Recession — outcome variable, often caused by triggers 2 (oil-Fed channel), 3 (Treasury market dislocation), or 5 (cyber operational impact).
  2. Regional bank failure — downstream of CRE state variable plus credit shock.
  3. Bond-equity correlation breakdown — outcome of the oil-inflation-Fed loop (Loop C).

4. Five Amplifiers (Detail)

4.1 AI capex debt-financing shift

BIS Bulletin 120, "Financing the AI Boom: From Cash Flows to Debt" (Aldasoro, Doerr, Rees), published January 7, 2026 (bis.org/publ/bisbull120.htm), documents that leading technology firms — Amazon, Alphabet, Microsoft, Meta, and Oracle — are increasingly using external financing for AI investment, with free cash flows now lagging capital expenditures in absolute amounts. The BIS Quarterly Review of March 2026 published a follow-on piece, "Financing the AI infrastructure boom: on- and off-balance sheet borrowing" (Aldasoro/Doerr/Rees), at bis.org/publ/qtrpdf/r_qt2603u.htm.

Capex estimates carry a range. Full source-by-source reconciliation appears in Appendix A. Headline anchors with evidentiary grading:

Estimate Status
~$650B (Bridgewater, Greg Jensen letter Feb 23, 2026, four Big Tech firms) Directly sourced analyst estimate
~$690-725B (Big Four hyperscaler post-Q1 2026 consensus) Constructed/consensus from Q1 earnings-call guides
~$750B (CreditSights, five hyperscalers including Oracle) Constructed/analyst (April 2026)
~$800B (global, including Chinese hyperscalers and CoreWeave) Order-of-magnitude estimate

The evidence supports a material shift toward external financing for AI capex; the exact dollar scale remains estimate-dependent. Bridgewater estimates AI capex contributed ~50 bps to 2025 US GDP growth and projects ~140 bps in 2026. This is amplifier coupling: AI narrative risk feeds directly into IG credit spreads via hyperscaler debt issuance.

4.2 Treasury and IG supply pressure

Apollo Academy / Torsten Slok (March 24, 2026): "Ten trillion dollars in existing US government debt will need to be refinanced over the coming 12 months… The budget deficit this year is about $2 trillion. Total gross corporate bond issuance in 2026 is likely to be around $2 trillion because of increased supply from hyperscalers. Adding it all up, the total amount of investment grade supply coming to the market this year is around $14 trillion." Source: apolloacademy.com/14-trillion-in-ig-supply-coming-to-the-market.

The OECD Global Debt Report 2026 (March 2026, DOI 10.1787/e9d80efd-en) provides corroborating context:

  • Total government + corporate bond market borrowing was $27T in 2025; projected $29T in 2026 (17% above 2024, double the level a decade ago).
  • OECD sovereign gross borrowing reached a record $17T in 2025, projected $18T in 2026.
  • 2025 refinancing requirements were $13.5T — approximately 80% of gross borrowing, a new high.
  • AI hyperscalers: nine firms raised $122B in bond markets in 2025 (nearly half of all tech-firm issuance).

This is an analyst estimate plus an official OECD compendium, not an official Treasury forecast.

4.3 Index concentration

Documented in §2.3. The amplification mechanism is conditional, not automatic. The dashboard (§12) tracks both mechanical contribution (Magnificent 7 weight × Magnificent 7 return) and flow-driven indicators (ETF redemption volumes, volatility-targeted fund deleveraging).

4.4 Valuation extremes

Documented in §2.2. The decisive empirical link between CAPE level and drawdown frequency comes from regime-conditional analysis in §7.3 — this is the framework's strongest single empirical claim, anchored on one completed historical episode (1999-2000).

4.5 CRE refinancing wall

MBA's 2025 Commercial Real Estate Survey of Loan Maturity Volumes (released February 9, 2026 at the CREF Convention): 17% of $5.0 trillion of outstanding commercial and multifamily mortgages — approximately $875 billion — is scheduled to mature in 2026, a 9% decline from $957B scheduled to mature in 2025. Source: mba.org/news-and-research/newsroom/news/2026/02/09/17-percent-of-commercial-and-multifamily-mortgage-balances-to-mature-in-2026.

Property-type breakdown of 2026 maturities: hotel/motel 30%, industrial 23%, office 17%, healthcare 15%, multifamily 13%. Investor-type breakdown: depositories $396B, CMBS/CLO/ABS $200B, credit companies/warehouse/other $163B, life insurance $76B, GSEs $39B. 2027 maturities are $652B.

MBA's CREF Loan Performance Survey Q1 2026 (April 27, 2026): overall commercial mortgage delinquency rate 4.02%, up from 3.86% in Q4 2025. CMBS 5.21% (up from 4.97%); life insurance 1.47%; GSE 0.97%; FHA 0.96%. CRED iQ parallel data: March 2026 CMBS distress rate (including specially serviced loans) ~12%.

Bank exposure: Florida Atlantic University's screener using Q3 2025 Call Report data finds 59 of the 155 largest US banks have CRE exposures exceeding 300% of equity capital (down from 62 banks the previous quarter). Aggregate CRE/equity ratio fell from 138% to 132%. This is industry-derived academic screening based on official Call Report data; it is not an FDIC-published quarterly tally. The 300% threshold derives from interagency supervisory guidance (FDIC FIL-23-2023; reaffirming 2006 OCC/Fed/FDIC joint guidance). Identified high-concentration banks (Q4 2024 vintage screener): Flagstar, Zions Bancorp, Valley National, Synovus, Umpqua, Old National — each >$50B in total assets.

This is a slow amplifier — distress is realized at refinancing dates over years — but it shapes the latent loss curve under any rate-rise or credit-shock scenario.

5. Eight Triggers (Detail)

For each trigger: current-state description, empirical anchor for the marginal probability, and low/base/high probability range used in the Monte Carlo. The base probability is in each case an author prior informed by the listed evidence; it is not mechanically derived from the anchors. Reproducibility flags appear alongside each anchor.

5.1 AI capex / revenue disappointment

Current state. Hyperscaler capex is increasingly debt-funded. Q3 and Q4 2026 earnings will be the first concrete tests of whether AI revenue is converting at the pace required. Specific stress points:

  • Microsoft tracking $120B+ for FY26 per company guidance; analyst extensions to ~$190B for calendar 2026 (CreditSights; Yahoo Finance/om.co).
  • Alphabet raised 2026 capex guide to $180-190B from $175-185B (CFO Anat Ashkenazi, Q1 earnings call).
  • Amazon ~$200B for 2026 (Feb 2026 guidance); Morgan Stanley sees FY26 free cash flow -$17B; BofA -$28B.
  • Meta raised 2026 capex range to $125-145B from $115-135B; stock -6% after-hours on the print.
  • Oracle: $50B+ 2026 capex (Futurum/Introl); capex-to-revenue 86%; RPO $523B. Moody's downgraded Oracle's outlook to Negative from Stable in September 2025; rating maintained at Baa2 (lowered-end investment grade), reaffirmed in Moody's December 15, 2025 Credit Outlook. S&P loosened its downgrade trigger to 4× debt/EBITDA from 3.5×. Barclays Fixed Income Research (analyst Andrew Keches) downgraded Oracle's debt to Underweight on November 11, 2025, citing capex-cash gap and counterparty risk from the ~$300B OpenAI agreement; flagged that cash reserves could be depleted by November 2026 and debt could fall to BBB-. Note: this is a debt/credit downgrade; Barclays' equity research (Raimo Lenschow) maintained Overweight through the period.
  • Michael Burry (Scion Asset Management): On May 10, 2026 via his Cassandra Unchained Substack, Burry disclosed that approximately 80% of Scion's portfolio is in NVDA + PLTR puts (notional ~$1.1B): 1,000,000 NVDA puts strike $110 expiring December 2027; PLTR puts strike $100 (December 2026) and $50 (June 2027) plus a direct short. Scion was deregistered after the Q3 2025 13F, so disclosure is via Substack/X, not a 13F filing. (TradingKey; Foreign Policy Journal, May 7, 2026.)

Empirical anchors. Polymarket "AI bubble burst in 2026?" was trading near 16-28% YES as of May 11-12, 2026 (multi-outcome market; specific outcome and snapshot time matter for the precise number). Historical base rate for sector re-ratings ≥30% from peak valuation within 24 months: approximately 35-45%. Bloomberg NVDA option surfaces show 12-month left-tail skew implying P(-20%+ move) ≈ 18-22% (proprietary surface; illustrative).

Marginal probability (12-month): low 15%, base 25%, high 40%.

5.2 Iran / Hormuz escalation

Current state. A ten-week-plus closure of the Strait of Hormuz against a fragile, repeatedly violated ceasefire. Iran's response to Trump's latest peace proposal was rejected as "TOTALLY UNACCEPTABLE" on May 10-11. WTI rose 2%+ on May 11.

Disaggregation:

  • IEA-reported supply fall (realized, March 2026): 10.1 mb/d ("the largest disruption in history" — IEA OMR, April 14, 2026).
  • Pre-war Hormuz crude+condensate transit: ~20 mb/d, falling to ~2-4 mb/d in March-April 2026 (the flow disruption, not the realized supply loss, because of bypass via Saudi Red Sea ports, UAE Fujairah, and Iraq-Türkiye pipeline).
  • Saudi Aramco CEO Amin Nasser: ~100 million barrels per week potentially at risk if Hormuz remains closed (TradingEconomics; Reuters; May 10-11, 2026).
  • IEA Director Fatih Birol (Norges Bank In Good Company podcast; CNBC April 1, 2026): "Today, we lost 12 million barrels per day — more than two of these oil crises [1973 and 1979] put together." On April 13, 2026 at the Atlantic Council in Washington, Birol revised this to ~13 mb/d shut-in with more than 80 energy facilities damaged.
  • IEA emergency release: March 11, 2026 — IEA members agreed to release 400 million barrels from emergency reserves.

Empirical anchors (Polymarket, May 11-12, 2026 snapshots). "US-Iran peace deal by May 31, 2026?" 13-28% YES across market configurations (paper uses 17.5%); "Strait of Hormuz traffic returns to normal by May 15?" ~0.7%; "Iran agrees to end uranium enrichment by June 30?" ~25-28%. Polymarket prices are sensitive to liquidity, settlement-text mapping, manipulation risk, and participant composition; all values are time-of-day specific and used as triangulation, not calibration.

Marginal probability of meaningful further escalation (12-month): low 20%, base 30%, high 45%.

5.3 Treasury basis trade unwind

Current state. OFR Brief 26-01 (March 3, 2026), "Hedge Fund Participation in Cleared Repo," states: "Recent estimates suggest that about 75% of all hedge fund Treasury repo activity is not centrally cleared." (financialresearch.gov/briefs/2026/03/03/hedge-fund-participation-cleared-repo). A companion OFR blog (January 29, 2026) estimates that under the SEC central-clearing rule, 77% would have been cleared versus the actually-cleared 45% of daily Treasury repo outstanding in the first 8 months of 2025.

FSB published "Vulnerabilities in Government Bond-backed Repo Markets" on February 4, 2026 (fsb.org/2026/02/vulnerabilities-in-government-bond-backed-repo-markets/; PDF at fsb.org/uploads/P040226.pdf). Key statistics:

  • ~$16 trillion in repo trades backed by government bonds outstanding at end-2024 (~80% of total repo stock); up ~20% since 2022.
  • US accounts for ~60% of activity; ~40% of repo outstanding involves counterparties in different jurisdictions.
  • ~70% of the non-centrally cleared segment operates with zero haircuts.

The SEC's Treasury central-clearing rule (compliance dates extended February 25, 2025) is effective December 31, 2026 for eligible cash market transactions and June 30, 2027 for eligible repo. The SEC approved CME Securities Clearing as a Treasury Covered Clearing Agency (December 1, 2025) and ICE Clear Credit (January 30, 2026), expanding CCP options beyond FICC.

Earlier industry summaries referencing >$6T in repo financing, 18:1 leverage, and 73.8% of repo at zero/negative haircuts are directionally consistent with the OFR/FSB picture but are treated here as directional rather than specifically sourced. The substantive claim — that the safest market is now the most leveraged, with policy ammunition more constrained than in 2020 — holds.

Empirical anchors. Historical base rate for Treasury-market dislocations requiring Fed intervention since the SRF was established in 2021: zero realized (the August 2024 yen carry tremor produced a near-miss). Prior to 2021: roughly one episode per 5-10 years (1998 LTCM, 2008 GFC funding stress, March 2020). MOVE index at 67-72 in early May 2026 is well below the 130+ threshold historically associated with stress periods. No direct prediction-market anchor exists.

Marginal probability (12-month): low 8%, base 12%, high 20%.

5.4 Yen carry unwind

Current state. BOJ held at 0.75% on April 28, 2026 in a 6-3 vote, with three dissenters (Takata, Tamura, Nakagawa) voting to raise to 1.0%. The August 5, 2024 yen tremor (-12.4% Nikkei on a single day; second-largest percentage drop in Nikkei history after October 1987) is the relevant precedent. Carry-trade size estimates vary by source; Appendix C provides a structured exposure matrix rather than a single estimate.

Empirical anchors. Polymarket BOJ-rate markets that were active in 2024-2025 are not currently active for the 2026 horizon. USD/JPY 1-year implied volatility at approximately 9-10% (CME/Bloomberg, May 2026) implies P(10%+ JPY appreciation within 12 months) ≈ 12-18% under a lognormal flat-vol approximation (this is not a risk-neutral or model-free estimate; treated as illustrative). The August 2024 historical precedent — ~5% of relevant positions liquidating produced a -12.4% Nikkei single-day move and ~3% S&P move — is the relevant base for what a "material" unwind transmits.

Marginal probability of >15% material unwind (12-month): low 10%, base 18%, high 30%.

5.5 Major AI-enabled cyber incident

Current state. Anthropic announced Project Glasswing on April 7, 2026 (anthropic.com/glasswing; anthropic.com/project/glasswing): an initiative to secure critical software powered by Claude Mythos Preview, which Anthropic reports identified "thousands of zero-day vulnerabilities" across every major OS and browser. Launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic committed $100M in Mythos Preview usage credits + $4M in direct donations to open-source security organizations. Anthropic stated it does not plan to release Mythos Preview generally. (Note: the "thousands of zero-days" is a self-reported Anthropic claim, not an independently audited finding.)

The IMF published "Financial Stability Risks Mount as Artificial Intelligence Fuels Cyberattacks" on May 7, 2026 (imf.org/en/blogs/articles/2026/05/07/...) authored by Tobias Adrian, Tamás Gaidosch, and Rangachary Ravikumar. Verbatim: "IMF analysis suggests that extreme cyber-incident losses could trigger funding strains, raise solvency concerns, and disrupt broader markets." The blog explicitly references Claude Mythos Preview as evidence of how quickly risks are increasing.

Empirical anchors. SEC public-company cyber incident disclosures under Item 1.05 of Form 8-K (effective December 18, 2023): per Debevoise & Plimpton's Cybersecurity Incident Disclosure Tracker (April 10, 2026 update), there have been approximately 44 cumulative Item 1.05 filings through April 10, 2026 — Dec 2023: 2; 2024: 25; 2025: 13; Q1 2026: 3 plus 2 additional in early April. Six (~14%) were filed by financial-sector (SIC 6000-series) issuers including First American Financial, South State, Prudential, FHLBNY, B. Riley, and BayFirst Financial. Item 1.05 filing pace slowed in 2025 vs 2024 after SEC Corp Fin Director Erik Gerding's May 2024 Statement redirecting non-material incidents to Item 8.01.

The IMF's May 7, 2026 framing argues that the AI-enabled rate is likely above this historical baseline.

Marginal probability of a financially material AI-enabled cyber incident (12-month): low 5%, base 10%, high 20%.

5.6 Private credit fund forced gating

Current state. Fitch reports US private credit default rates of 9.2% in 2025 (Reuters, March 6, 2026) — a record in Fitch's Private Markets Research series (begun 2018); up from 8.1% in 2024. Smaller issuers (EBITDA ≤$25M) had a 15.8% default rate vs 4% for larger borrowers (EBITDA >$100M). Six of eight cases with ultimate recoveries paid in full. Fitch monitored 302 companies and recorded 38 defaults among 28 borrowers.

The FSB published "Report on Vulnerabilities in Private Credit" on May 6, 2026 (fsb.org/2026/05/report-on-vulnerabilities-in-private-credit/; PDF P060526.pdf). Key data:

  • Total private-credit market: $1.5-2.0T at end-2024 (FSB official scope); commercial estimates of "broad private credit" reach $3.5T (ACC/AIMA Financing the Economy 2025 with Houlihan Lokey).
  • Geographic concentration: US ~$1T; euro area + UK; smaller markets in Canada, Hong Kong, Japan, Switzerland, South Africa.
  • Bank direct exposure: FSB official estimate ~$220B drawn + undrawn credit lines to private-credit funds; commercial estimates $270-500B. Bank private-credit exposure <0.5% of total bank assets.
  • Retail investor share rose from ~0% to ~13% over the past 10 years.
  • Insurer exposure: ~10% of life insurer portfolios may be private credit.

BCRED Q1 2026 redemption event (verified across CNBC March 5, 2026; Bloomberg March 2, 2026; AltsWire): Investors requested redemptions of 7.9% of NAV (approximately $3.7-3.8 billion) against a standard 5% cap. Blackstone raised the cap from 5% to 7% and injected $400 million of internal capital ($250M from the firm + $150M from >25 senior executives), honoring 100% of requests. BCRED Q1 NAV was approximately $45B (down from $47.6B at year-end 2025); non-accruals 2.4% at cost / 1.4% at fair value; Medallia (mark 60.3) and Affordable Care (69.8) were cited as top contributors to mark declines.

Industry-wide stress (Bloomberg March 11, 2026; Reuters; Alternative Credit Investor April 7, 2026):

  • Cliffwater Corporate Lending Fund (~$33B AUM): redemptions ~14% of NAV; capped withdrawals.
  • BlackRock HPS Corporate Lending Fund (~$26B AUM): $1.2B redemption requests (9.3% of NAV); paid out $620M under the 5% cap.
  • Morgan Stanley North Haven Private Income Fund (~$7.6B AUM): redemptions 10.9% of shares; returned ~$169M (45.8% of requests).
  • Blue Owl Credit Income Corp: 21.9% of shares requested redemption in Q1 2026; paid ~14% under cap.
  • Blue Owl Technology Income Corp: 40.7% of shares requested redemption; paid ~14% under cap.
  • Moody's BDC sector outlook cut from Stable to Negative on April 7, 2026 (Reuters), citing "rising redemption pressures, higher leverage, and weakening access to funding markets." Total BDC sector assets ~$400B; non-traded BDCs >60% of sector AUM.

Empirical anchors. Historical base rate for semi-liquid credit fund gating events: approximately one major event per 3-5 years in stress periods. Current Fitch default rate (9.2%) is the record in Fitch's PMR series and is consistent with elevated forward gating risk.

Marginal probability of additional forced gating event (12-month): low 8%, base 15%, high 25%.

5.7 Stablecoin de-peg with contagion

Current state. Total stablecoin market capitalization $322.7B as of May 10, 2026 per DefiLlama (cryptonews.net); USDT $189.6B (58.76% dominance); USDC $78.96B. The GENIUS Act was signed July 18, 2025 (Senate 68-30 on June 17; House 308-122 on July 17), establishing the US framework for payment stablecoins: 100% reserve backing with USD or short-term Treasuries; monthly public reserve disclosures; Bank Secrecy Act applicability; insolvency priority for stablecoin holders; classification as non-securities/non-commodities. Eligible issuers limited to insured depository institution subsidiaries, OCC-approved federal nonbanks, and state-qualified issuers. Effective date: earlier of 18 months after enactment or 120 days after final rules.

Tether Q1 2026 attestation (BDO Italia ISAE 3000R, published late April/early May 2026; tether.io/news/...): As of March 31, 2026 — total assets $191.77B; total liabilities $183.54B; direct US Treasury bill holdings $117.0 billion (line item: $117,035,732,050); direct + indirect T-bill exposure (including reverse repos and money-market funds) ~$141 billion (Tether's headline number); excess reserves $8.23B (record); Q1 net profit $1.04B; gold $19.84B (~132 metric tons at XAU $4,668.06/oz).

The systemic transmission mechanism — stablecoin redemptions forcing Treasury-bill sales — operates on the direct T-bill subset ($117B). The headline $141B includes indirect exposure that would also be liquidated in stress but with longer paths.

Empirical anchors. Historical base rate: UST/Luna (May 2022, ~$60B destruction, contagion contained to crypto), USDC brief de-peg (March 2023 SVB exposure, recovered within 48 hours). Approximately one event per 1-2 years in crypto markets, with only one cross-market-contagion event historically.

Marginal probability of de-peg with measurable TradFi contagion (12-month): low 2%, base 5%, high 12%.

5.8 Taiwan kinetic escalation

Current state. Bloomberg Economics' February 10, 2026 update, "The $10 Trillion Fight: Modeling a US-China War Over Taiwan" (bloomberg.com/news/articles/2026-02-10/...) estimates that in the most extreme case, a US-China conflict over Taiwan would cost the global economy about $10.6 trillion, roughly 9.6% of global GDP, in the first year alone — "eclipsing the impact of the Covid-19 pandemic and the 2007-09 global financial crisis." First-year impacts by economy: China -11%, US -6.6%, Taiwan -40%, S. Korea -23.3%, Japan -14.7%, EU -10.9%, India -8%, UK -6.1%. The probability is low; the magnitude is extreme.

PLA Q1 2026 activity around Taiwan: large-scale amphibious exercises opposite Taiwan in late April; routine Liaoning carrier transits April 20; East China Sea patrols April 18. ODNI's mid-March 2026 Annual Threat Assessment concluded PLA "likely lacks plans for a 2027 offensive or sooner." Beijing announced 10 new economic incentives for Taiwan on April 12, prioritizing gray-zone coercion over kinetic conflict.

Empirical anchors (Polymarket, May 11-12, 2026 snapshots).

Polymarket market Implied probability
Will China invade Taiwan by June 30, 2026? ~0% (paper uses 1.8%)
Will China invade Taiwan by end of 2026? ~7% (paper uses 7.4%)
China x Taiwan military clash before 2027? ~9-12% (paper uses 12.0%)
Will China invade Taiwan by Dec 31, 2027? ~17-18%

The framework's "Taiwan kinetic escalation" trigger maps most closely to "military clash before 2027" at approximately 9-12%, not the narrower "full invasion" framings.

Marginal probability of discrete kinetic step (12-month): low 4%, base 9%, high 15%.

6. Feedback Loops

The feedback-loop topology is the framework's most durable analytical contribution. Each loop describes a self-reinforcing transmission pathway: a trigger creates pressure that engages a state variable that creates further pressure on the same or adjacent triggers.

Loop A: AI capex ↔ IG supply ↔ yields ↔ AI valuations. Hyperscaler debt issuance lifts IG supply → marginal-buyer pressure raises yields → AI cash flows discounted lower → if AI revenue misses, Oracle credit (and credit of similarly-positioned issuers) widens → loop tightens. Already partially activated (Moody's Oracle outlook negative September 2025; Barclays fixed-income downgrade November 2025).

Loop B: AI capex ↔ private credit ↔ BDC retail flows ↔ public credit. Private credit funds AI infrastructure via asset-based finance and JV debt. A 20-30% capex compression → NAV write-downs → BCRED-style redemptions → fire sales into public credit. The Q1 2026 BCRED/HPS/Cliffwater/Morgan Stanley/Blue Owl events were a low-magnitude test; the system bent but did not break.

Loop C: Iran ↔ oil ↔ inflation ↔ Fed ↔ bond-equity correlation. $100+ Brent keeps inflation sticky → Fed cannot cut → bonds and equities fall together → 60/40 fails → risk-parity forced selling. The IMF's April 2026 GFSR identified this directly: more frequent supply shocks have eroded the equity-bond hedging relationship, raising risks of simultaneous deleveraging in both asset classes.

Loop D: Yen carry ↔ Mag 7 ↔ passive flows ↔ index concentration. BOJ hikes → carry unwinds → tech selling concentrated in Mag 7 → because the Mag 7 share-class aggregate is ~34.7% of SPY (and the S&P 500 broadly), index moves mechanically. Cap-weighted funds do not themselves force-sell on price declines; index weights simply adjust. Second-round selling pressure comes from passive redemptions, volatility-targeted deleveraging (risk-parity and vol-control funds), margin calls on leveraged positions, and active/quant rebalancing. 0DTE gamma hedging on Mag 7 single-stock options amplifies intraday velocity. August 5, 2024 (Nikkei -12.4%) is the partial precedent.

Loop E: Basis trade ↔ Treasury liquidity ↔ everything. Leveraged hedge funds + repo financing (75% non-centrally cleared per OFR Brief 26-01). Vol spike → margin calls → Treasury dislocation → because Treasuries are the global risk-free curve, all assets reprice. March 2020 precedent. SEC's central-clearing rule (effective December 31, 2026 for cash, June 30, 2027 for repo) is the right structural fix, but the transition itself creates risk.

Loop F: Private credit / NBFI ↔ banks ↔ refinancing risk. EU banks' 9.2% NBFI exposure (FSB, EBA reporting) is NOT the same as direct private-credit exposure. Three sub-channels:

  • Direct on-balance-sheet bank exposure to private-credit funds: materially smaller than 9.2% (FSB May 2026: ~$220B drawn + undrawn, <0.5% of bank assets); not itemized in standard bank capital reporting.
  • Broad NBFI exposure (the 9.2% figure): includes hedge funds, family offices, sovereign wealth, asset managers, and private credit. Most is non-correlated with private-credit stress.
  • Indirect linkage: SRT (significant risk transfer) trades, warehouse facilities, repo against private-credit collateral, and shared counterparty risk. This is the real systemic concern. If private-credit funds retrench, the SRT investor base shrinks, raising bank refinancing risk.

Loop evidence ratings:

Loop Conceptual strength Evidence strength Quantified?
A — AI/IG/yields High Medium-high Partly
B — AI/private credit/BDC High Medium-high Partly
C — Oil/Fed/correlation High High Partly (IMF GFSR)
D — Yen/Mag 7/passive Medium Medium Weakly
E — Basis/Treasury liquidity High High Partly (OFR/FSB)
F — Private credit/NBFI/banks High Medium Weakly

The factor-copula structure in §7.4 captures cross-loop correlation: triggers that share factor loadings cluster, so loop activations are correlated through latent risk-off, inflation/Fed, liquidity/funding, and tech/AI factors.

7. Probability Framework

The framework uses a multi-layer stack. The dominant layer is empirical-historical, not modeled. Historical regime-conditional data answers a sharper question than any model output: given that markets have historically started from CAPE regimes like the current one, what is the realized frequency of forward drawdowns?

The layers, in order of priority for decision-making:

  • Layer C-conditional (§7.3): Historical rolling-window drawdown coverage rate, conditioned on starting CAPE regime. The strongest empirical anchor — but at the episode level, n = 1 for CAPE ≥40.
  • Layer C-unconditional (§7.2): All-history rolling-window drawdown coverage rate. Context.
  • Layer A (§7.4): Scenario-generator trigger co-occurrence via factor-copula Monte Carlo. Not a calibrated probability forecast.
  • Layer B (§7.5): Scenario-generator trigger-mediated severity via structured second-stage Monte Carlo. Not a calibrated probability forecast.
  • Layer D (§7.6): Dashboard-conditioned posterior updating as monitoring indicators fire.

§7.7 brings the layers together. They should not be combined into a single number.

7.1 Approach: empirical anchor over modeled forecast

A model is only as good as its calibration. The eight-trigger Monte Carlo carries parameter uncertainty in three places: marginal probabilities (author priors anchored to external evidence but not directly extracted from market prices), correlation structure (factor loadings calibrated to economic logic, not estimated), and severity transmission (slope parameter calibrated to historical analogs ranging from 3 to 14, with median 5-7). Every output therefore depends on the frozen parameter set in Appendix J.2; nothing in Layers A and B is an empirical probability of a future event.

Conversely, the historical 24-month rolling-window drawdown coverage rate starting from CAPE ≥40 — computed below — is descriptive empirical history for the available sample. The sample is dominated by one completed episode (January 1999 – September 2000); this is one historical episode, not many independent observations. The computation is fully reproducible (Appendix F).

The framework therefore leads with §7.3 (regime-conditional history), uses §7.2 (unconditional history) for context, and treats the scenario-generator layers §7.4 (trigger co-occurrence) and §7.5 (trigger-mediated severity) as channel-identification augmentation rather than probability forecasts.

7.2 Layer C unconditional: all-history base rates

Methodology.

  • S&P 500 monthly close, 1881-2026: Robert Shiller monthly series through September 2023, extended October 2023-May 2026 as described below.
  • Price-only index (excludes dividends).
  • Forward-extension beyond September 2023 (last month of Shiller's published file): monthly S&P 500 close + monthly CAPE from Multpl/GuruFocus, October 2023 - May 2026 (33 additional monthly observations).
  • For each rolling k-month window (k ∈ {12, 18, 24, 36}), find the maximum peak-to-trough drawdown from the rolling all-time high within the window.
  • Score window = 1 if max drawdown within the window ≥ D; average across all rolling windows.
  • This is coverage frequency, not event-count frequency — the answer to "starting from a randomly selected month, what is the probability that the next k months include a ≥D drawdown."

Empirical results (1881-2026, monthly rolling-window coverage frequency):

Drawdown threshold P(within 12 mo) P(within 18 mo) P(within 24 mo) P(within 36 mo)
≥ 10% 37.8% 51.6% 63.0% 81.0%
≥ 15% 21.6% 32.7% 42.7% 60.7%
≥ 20% 13.5% 21.8% 28.9% 42.3%
≥ 25% 7.4% 13.6% 19.9% 30.9%
≥ 30% 4.6% 7.1% 10.1% 17.4%
≥ 40% 2.0% 3.7% 5.2% 8.8%
≥ 50% 0.7% 1.6% 2.6% 4.0%

Reproducible code in Appendix F. This is the unconditional 1881-2026 monthly base rate. It averages across all starting regimes, including the many low-CAPE windows that mechanically have less downside available.

7.3 Layer C-conditional: regime-conditional historical coverage frequency

The unconditional rate is a misleading comparison for the current market. CAPE ~40 is between the 95th and 99th percentile of the 1881-2026 monthly CAPE distribution. Markets starting from this regime have a different historical track record than the average starting month.

Methodology.

  • S&P 500 monthly close 1881-2026 (Shiller through September 2023 plus the documented October 2023-May 2026 extension).
  • Shiller's published monthly CAPE 1881-2023, extended through May 2026 with Multpl monthly readings.
  • For each rolling 24-month window, filter starting months by CAPE regime.
  • Average drawdown coverage frequency within each regime.

CAPE distribution 1881-2026 (monthly values, equal-weighted by month):

Statistic Value
Mean 17.75
Median 16.60
75th percentile 21.45
90th percentile 28.02
95th percentile 32.84
99th percentile 40.99
All-time max 44.20 (December 1999)

Current CAPE (~40.4 as of May 8, 2026) is at the right tail — between the 95th and 99th percentile.

Regime-conditional 24-month forward drawdown coverage frequency (central table):

Starting CAPE regime N monthly windows P(≥10%) P(≥20%) P(≥25%) P(≥30%) P(≥40%)
All starting months 1,722 63.0% 28.9% 19.9% 10.1% 5.2%
Low CAPE (<15) 690 59.7% 23.3% 14.2% 4.6% 1.0%
Medium (15-25) 775 65.0% 32.7% 24.0% 12.9% 7.1%
High (25-35) 208 59.6% 23.1% 19.7% 15.9% 13.0%
Very high (≥35) 49 91.8% 71.4% 36.7% 18.4% 2.0%
Extreme narrow (≥38) 27 100.0% 85.2% 55.6% 22.2% 3.7%
Extreme (≥40) — current 21 100.0% 90.5% 61.9% 19.0% 4.8%
Very extreme (≥42, 1999 only) 13 100.0% 100.0% 69.2% 23.1% 0.0%

12-month forward (for shorter-horizon planning):

Starting CAPE regime P(≥10%) P(≥20%) P(≥25%) P(≥30%)
All starting months 37.8% 13.5% 7.4% 4.6%
Very high (≥35) 50.0% 23.3% 1.7% 0.0%
Extreme (≥40) 47.6% 33.3% 4.8% 0.0%

Headline. In the available historical sample, monthly rolling 24-month forward windows beginning from CAPE ≥40 had a 61.9% coverage frequency of a peak-to-trough ≥25% drawdown (13 of 21 monthly observations). Uplift versus unconditional: 3.11x. The 36-month coverage frequency at this regime reaches 100% at ≥25% and 61.9% at ≥40% — meaning all 21 monthly windows experienced a ≥25% drawdown within 36 months.

This is rolling-window coverage frequency, not an independent-event probability. Because these windows are highly overlapping and concentrated in a small number of distinct historical regimes, the result should be interpreted as a historical stress signal anchored on one episode, not as a calibrated 24-month forecast probability.

Distinct CAPE ≥40 episodes (1881-2026 monthly data):

Episode Start End Months Max CAPE Max CAPE date Forward 24m max DD range 24m P(≥25%)
Dot-com peak 1999-01 2000-09 21 44.19 1999-12 10.4% - 40.9% (mean 26.8%) 13/21 = 61.9%
Current 2026-03 2026-05+ 3 (and continuing) 40.55 2026-04 Incomplete — paper publication May 12, 2026 N/A

At CAPE ≥38, broader sample (5 distinct episodes): 1998-07; 1998-12 to 2000-11 (24 months); 2021-11 to 2021-12 (2 months); 2024-11 to 2024-12 (2 months); 2025-09 to 2026-05 (9 months and ongoing). The 2021 and 2024 peaks (38.58 and 38.78 respectively) did not cross 40. Note on cross-period sourcing: the 2021 reading (38.58) comes from Shiller's published file (period covered by Shiller through September 2023); the 2024 reading (38.78) comes from the Multpl/GuruFocus forward extension documented in Appendix J.4a. Both sources are documented in §7.2 methodology; the regime filter does not change at any plausible tiebreaking choice.

Uncertainty: why this paper does not report a confidence interval. The stationary block-bootstrap output for this regime is shown below as a diagnostic, not as a confidence interval. It is not statistically interpretable for this sample. The bootstrap output itself displays the diagnostic that exposes the problem:

Regime Threshold Point estimate Bootstrap 5th pct Bootstrap 50th pct Bootstrap 95th pct
CAPE ≥ 35 ≥25% 36.7% 60.7% 85.1% 100.0%
CAPE ≥ 38 ≥25% 55.6% 63.0% 91.7% 100.0%
CAPE ≥ 40 ≥25% 61.9% 57.7% 97.4% 100.0%

The CAPE ≥35 row shows the issue starkly: the bootstrap 5th percentile (60.7%) lies above the point estimate (36.7%). A 5th percentile above the point estimate is a pathology, not a feature. It happens here because block bootstrap on a rare regime concentrated in one historical cluster either resamples around that cluster (producing very high coverage) or misses it entirely (producing replications discarded as undefined). The resulting distribution does not have a frequentist confidence-interval interpretation. Appendix K documents this in full. A Wilson interval treating the 21 windows as independent Bernoulli trials (which they are not — they overlap) would be approximately [44%, 77%] at 90% confidence; even that is not appropriate because the observations come from one regime.

The honest statement is: n = 1 at the episode level for CAPE ≥40, and the 61.9% figure is descriptive history within that one completed episode.

Non-overlapping-window sensitivity (sample every 24 months so windows do not overlap): only 1 truly independent observation at CAPE ≥40 (the 1999 episode contributes 1 non-overlapping window with peak-to-trough max drawdown 40.9%, exceeding 25%). At CAPE ≥35 there are 2 non-overlapping observations (1/2 hit 25%). This confirms that the 61.9% figure is fundamentally a single-episode observation rather than a 21-window statistical result.

Caveats.

  1. Sample composition. The CAPE ≥40 sample of 21 monthly observations is dominated by the January 1999 – September 2000 dot-com episode. The current episode (started March 2026) cannot yet contribute completed 24-month forward observations.

  2. Causation vs coincidence. We cannot fully separate "high CAPE causes drawdowns" from "the specific episode that produced a drawdown happened to coincide with high CAPE." The robust interpretation: in the one completed historical regime where CAPE reached current levels, the subsequent path was a substantial drawdown across nearly every monthly entry point in that episode. The proximate cause was the dot-com bust; the present episode's proximate cause has not yet been determined.

  3. Single-episode dominance. The 61.9% point estimate is functionally equivalent to: "in the one completed historical episode where CAPE ≥40, 13 of 21 monthly entry points were followed by a ≥25% peak-to-trough drawdown within 24 months."

  4. Forward extrapolation. The empirical signal points to elevated fragility from current valuation conditions. It does not imply that a ≥25% drawdown is 61.9% likely within the next 24 months as an unconditional forecast. The right framing is: "the only previous time we saw this configuration, the subsequent path was almost always a substantial drawdown — but n = 1 at the episode level."

This is computed historical coverage rate, not model output. The framework's scenario-generator layers (§7.4-7.5) add channel-identification value, but they do not transform this single-episode evidence into a calibrated probability.

7.4 Layer A: trigger co-occurrence scenario generator (factor-copula Monte Carlo)

This is a scenario-generation layer. Outputs are conditional on the frozen parameter set; they are not calibrated probability forecasts. For each of the eight triggers, a 12-month marginal probability prior is specified (§5). Co-occurrence is generated via Gaussian copula with a four-factor latent correlation structure:

  • F1 Risk-off (affects all triggers)
  • F2 Inflation/Fed (affects oil-driven and credit triggers)
  • F3 Liquidity/funding (affects Treasury, repo, carry, stablecoin)
  • F4 Tech/AI (affects AI-specific and Taiwan-via-chips)

Factor loadings:

Trigger Risk-off Inflation Liquidity Tech/AI Idiosyncratic
ai_capex 0.40 0.10 0.30 0.65 0.32
iran_hormuz 0.40 0.70 0.20 0.10 0.30
basis_trade 0.45 0.25 0.70 0.05 0.24
yen_carry 0.40 0.30 0.55 0.30 0.36
cyber 0.30 0.05 0.20 0.20 0.83
private_credit 0.40 0.25 0.55 0.20 0.43
stablecoin 0.30 0.10 0.50 0.15 0.63
taiwan 0.50 0.20 0.20 0.45 0.47

12-month marginal probabilities (regenerated frozen parameter file, seed=42, 200,000 trials):

Trigger 12m base 24m implied
ai_capex 0.25 0.4375
iran_hormuz 0.30 0.5100
basis_trade 0.12 0.2256
yen_carry 0.18 0.3276
cyber 0.10 0.1900
private_credit 0.15 0.2775
stablecoin 0.05 0.0975
taiwan 0.09 0.1719

Average pairwise binary correlation: 0.25 (range 0.10 to 0.42).

Base case results (24-month horizon, factor copula, 200,000 trials):

Metric Value
P(≥1 trigger fires) 76.7%
P(≥2 triggers fire) 54.6%
P(≥3 triggers fire) 38.3%
P(≥4 triggers fire) 25.7%
P(≥5 triggers fire) 16.0%
Expected trigger count 2.24

Conditional cluster amplification (factor-driven):

Conditioning Target trigger P(unconditional) P(conditional) Uplift
iran_hormuz fires basis_trade 22.6% 34.6% 1.53x
iran_hormuz fires yen_carry 32.7% 47.7% 1.46x
iran_hormuz fires private_credit 27.9% 40.4% 1.45x
iran_hormuz fires taiwan 17.3% 25.8% 1.49x
ai_capex fires taiwan 17.3% 31.1% 1.80x
ai_capex fires yen_carry 32.7% 51.9% 1.59x

The factor structure correctly captures cluster behavior: shared loadings on the liquidity factor pull basis_trade, yen_carry, and private_credit together; shared loadings on tech/AI pull ai_capex and taiwan (via chip-supply concerns) together. Cyber, with mostly idiosyncratic variance (~83%), barely correlates with anything.

Code in Appendix D.

7.5 Layer B: trigger-mediated severity scenario generator (structured Monte Carlo)

This is a scenario-generation layer. Outputs are conditional on the frozen parameter set; they are not calibrated probability forecasts. A second-stage Monte Carlo takes Layer A trigger vectors and maps them to drawdown distributions via:

  1. Primary severity = max(standalone severity) among fired triggers.
  2. Additional severity = 0.5 × (n_fired − 1) (diminishing returns from additional triggers).
  3. Coupling term = 0.3 × average liquidity coupling of fired triggers.
  4. Loop boost = sum of activated-loop boost values (additive, capped at 1.5).
  5. Policy offset = stochastic, weighted by trigger standalone severity; per-trigger p_eff_i calibrated from §8.7.
  6. Net severity = gross × (1 − policy offset).
  7. Drawdown = N(slope × net_severity, std), clipped to [0, 90%].

The slope parameter is the central calibration choice. Historical analogs imply:

  • LTCM 1998 (net severity ~2, realized 19.3% S&P drawdown): slope ~9-10.
  • GFC 2008 (net severity ~5, realized 57%): slope ~11-14.
  • August 2024 yen tremor (net severity ~2, realized 6-8%): slope ~3-4.
  • Median historical anchor: slope ~5-7.

The base case uses slope = 6.0 (median). Sensitivity range slope ∈ {4, 8} reflects whether you believe current state resembles "fast-policy-response" (slope=4) or "GFC-style policy slow" (slope=8) more closely. Policy-offset parameters are scenario assumptions, not estimated coefficients.

Base case results (factor copula, 24-month horizon, slope = 6.0, 200,000 trials):

Threshold P(trigger-mediated drawdown ≥ threshold over 24 months)
≥10% 74.9%
≥15% 71.6%
≥20% 63.9%
≥25% 49.9%
≥30% 33.8%
≥40% 10.4%
≥50% 1.5%

Slope sensitivity (P(≥25%, 24mo) and P(≥40%, 24mo)):

Slope Interpretation P(≥25%) P(≥40%)
4.0 Fast-policy-response (Aug 2024 anchor) 16.7% 0.4%
5.0 Conservative 34.0% 3.0%
6.0 Median historical anchor (base) 49.9% 10.4%
7.0 Moderate-aggressive (LTCM anchor) 61.4% 21.4%
8.0 Aggressive (GFC anchor) 67.9% 33.7%

Global sensitivity check (base slope = 6.0, 200,000 trials). The slope parameter remains the largest single severity-mapping choice, but other frozen assumptions also move the output. These variants perturb one assumption family at a time and leave all other inputs unchanged; they are robustness checks, not alternative forecasts.

Variant P(≥2 triggers) P(drawdown ≥25%) P(drawdown ≥40%) Interpretation
Base frozen file 54.6% 49.9% 10.4% Published scenario-generator baseline
Trigger marginals −20% 45.4% 42.1% 7.9% Lower author-prior trigger rates
Trigger marginals +20% 62.6% 56.9% 13.0% Higher author-prior trigger rates
Factor loadings −20% 59.9% 51.2% 9.2% Less clustering, more independent trigger incidence
Factor loadings +20% 48.1% 48.1% 11.9% More clustering, fewer multi-trigger trials but fatter tails when clustered
Policy offsets +20% 54.6% 43.0% 6.6% More effective policy response
Policy offsets −20% 54.6% 56.0% 15.2% Less effective policy response
Loop boosts set to zero 54.6% 40.6% 4.5% Transmission loops removed
Loop boosts +25% 54.6% 51.4% 11.2% Stronger loop amplification

Conditional on trigger count (base case, 24-month, slope = 6.0):

Triggers fired P(trial) Mean drawdown P(≥10%) P(≥25%) P(≥40%)
0 23.3% 1.0% 0.0% 0.0% 0.0%
1 22.1% 22.8% 93.3% 39.9% 1.1%
2 16.3% 26.5% 98.4% 56.3% 4.8%
3 12.6% 29.6% 99.7% 70.6% 10.4%
4 9.8% 32.6% 99.9% 82.8% 18.3%
5+ 16.0% 37.8% 100.0% 93.9% 39.4%

Internal consistency check: P(0 triggers) = 23.3% = 1 − P(≥1) = 100% − 76.7%. ✓

The structured MC gives an actual model output (not author overlay) for the scenario-generator's severity distribution. At base calibration, the scenario-generator output is P(trigger-mediated ≥25% drawdown within 24 months) = 49.9%. This is conditional on the frozen parameter set: change the slope from 6 to 4 and the figure becomes 16.7%; change to 8 and it becomes 67.9%. The structured MC's value is not in the headline number but in the slope sensitivity, the trigger-conditional structure (showing how severity scales with number of triggers fired, which channels couple most strongly), and the factor-conditional uplift results in §7.4.

Code in Appendix E.

7.6 Layer D: dashboard-conditioned posterior

The dashboard (§12) is the operational expression of Bayesian updating. As Tier 1 indicators fire, the framework's qualitative concern level should shift toward the upper end of the trigger-marginal ranges (§5) and the upper end of the slope sensitivity (§7.5). Empirical lead time of VIX ≥25 sustained 5 days is 28-157 days in 3 of 8 historical drawdown episodes (§12.1). Dashboard-conditioned updating is operationalized via Brier Skill Score against trigger-specific climatology (§13.3).

7.7 How the layers fit together

The layers answer different questions and should not be combined into a single number. In particular, the Layer C historical anchor (descriptive history) and the Layer B scenario-generator output (modeled conditional on frozen parameters) are not the same kind of object and should not be averaged or aggregated.

Layer Question answered Headline output
Layer C-conditional (CAPE ≥40) Historical rolling-window coverage rate from current regime 61.9%; n = 1 at episode level; no statistically meaningful CI
Layer C-conditional (CAPE ≥35) Broader current-regime coverage 36.7%
Layer C unconditional All-history monthly average 19.9%
Layer A (scenario generator) Trigger co-occurrence given frozen priors P(≥2 in 24mo) = 54.6%
Layer B (slope = 6, scenario generator) Trigger-mediated severity given frozen priors 49.9% P(≥25%, 24mo)
Layer B (slope = 4, scenario generator) Fast-policy-response assumption 16.7%
Layer B (slope = 8, scenario generator) GFC-style policy-slow assumption 67.9%

Interpretation. The framework's empirical-historical anchor (Layer C) and its scenario-generator output (Layer B) are independent constructs. Both point to elevated fragility under the current configuration, but neither produces a calibrated probability forecast. The framework's substantive contribution is channel identification — telling the user which pathways are most worth monitoring — and operational early-warning signals via the dashboard.

Layers should not be added; they overlap and they answer different questions. Specifically:

  • Layer C-unconditional history includes both trigger-driven and non-trigger drawdowns (sentiment, exogenous shocks).
  • Layer C-conditional is the same universe filtered by starting regime.
  • Layer B is trigger-mediated only; it misses sentiment-driven corrections (Q4 2018 analog) and exogenous shocks (March 2020 analog).
  • Layer A is co-occurrence, not severity.

Defensible composite framing:

In the available historical sample, the CAPE ≥40 regime has occurred in one completed episode (January 1999 – September 2000). Across that episode's 21 overlapping monthly starting points, 13 were followed by a peak-to-trough ≥25% drawdown within 24 months — a rolling-window coverage rate of 61.9%. This is descriptive single-episode evidence (n = 1 at the episode level), not a calibrated probability. The scenario-generator layers identify AI capex / oil-Fed / Treasury basis / yen carry as the channels most worth monitoring; Layer B at base slope generates 49.9% P(≥25%, 24mo) conditional on its frozen parameters, with slope-sensitivity range 16.7% to 67.9%. The dashboard (§12) provides selected-sample early-warning signals.

The framework's position is "elevated fragility, channels identified, monitoring discipline warranted" — not "calibrated crash probability."

8. Policy Reaction Functions

Drawdown severity depends materially on policy response. The framework treats policy effectiveness as trigger-specific — Fed liquidity facilities address basis-trade unwinds well but cannot reopen Hormuz, restore a stablecoin issuer's reserves, or repair systemic cyber damage.

8.1 Federal Reserve

  • Standing Repo Facility (SRF): Established July 28, 2021 by FOMC. Permanent overnight repo facility against Treasury, agency MBS, and agency debt collateral. Counterparties initially primary dealers, later expanded to depository institutions. Minimum bid rate set at or above administered rates. Reduces basis-trade unwind severity if conditions stay funding-driven rather than solvency-driven.
  • Discount Window and FIMA repo facility: Available to foreign central banks.
  • Emergency liquidity facilities (Section 13(3)): Available with Treasury approval. Historical precedents: 2008 (Bear Stearns / Maiden Lane I-III; AIG; PDCF; TALF). 2020 (PMCCF, SMCCF, MMLF, TALF 2.0, MLF, Main Street Lending Program). 2010 Dodd-Frank requires "broad-based eligibility" and Treasury Secretary approval.
  • Constraint: If Warsh is confirmed, the Powell-to-Warsh transition may reduce implicit "Fed put" credibility short-term. Warsh is historically viewed as more hawkish; deviation from precedent in a crisis would itself be destabilizing.

8.2 Treasury

  • Buyback program: Active since 2024, designed to support off-the-run Treasury liquidity.
  • Issuance composition management: Can shift duration mix toward T-bills to reduce price-sensitive long-end supply. Constrained by ~$10T 2026 refinancing — limited room for further T-bill shift without distorting front-end markets.

8.3 Bank of Japan

  • Communication channel: Can signal pause or slowdown in normalization without changing rates. Used effectively after August 2024.
  • JGB yield-curve operations: Available if 10-year yields rise too rapidly.
  • Constraint: Three dissenters (Takata, Tamura, Nakagawa) favor faster hikes; hawkish drift constrains BOJ's use of pause-as-signal.

8.4 Strategic petroleum and energy

  • SPR drawdowns can absorb 1-2 mb/d of supply shortfall short-term. With IEA citing 10.1 mb/d global supply fall, SPR is palliative not solution.
  • IEA member coordination already partially executed: March 11, 2026 IEA agreed to release 400 million barrels from emergency reserves; Birol stated IEA is "ready to dig deeper."

8.5 Regulatory forbearance and market structure

  • Capital rule relief for banks under stress (precedent: 2020 SLR adjustment).
  • Money market fund liquidity rule modifications.
  • LULD circuit breakers; market-wide circuit breakers at 7%, 13%, 20% S&P declines.
  • 0DTE option position limits under SEC review.

8.6 Central bank coordination

  • Fed-ECB-BOE-BOJ-SNB swap lines: standing arrangements activatable on hours-not-days timing.
  • G7/G20 statements: signaling channel.

8.7 Trigger-specific policy effectiveness

The single biggest determinant of severity: which trigger fires governs how effective standard policy tools are. Coarse calibration of per-trigger effectiveness:

Trigger Primary policy actor Effectiveness Constraint
Treasury basis unwind Fed (SRF + 13(3)) + Treasury High (if liquidity-only) Inflation/political constraints on Fed expansion
Iran/Hormuz escalation IEA / SPR / G7 / military / diplomacy Low-medium Physical supply constraints; SPR limits
AI capex disappointment Fed / Treasury / regulators Low Earnings/valuation issue, not liquidity
Yen carry unwind BOJ / Fed / MOF Medium BOJ inflation credibility constraint
AI-enabled cyber Fed / Treasury / CISA / private sector Low-medium Operational recovery, not just liquidity
Private credit gating Fed / Treasury / SEC / sponsors Medium Solvency vs liquidity; sponsor variation
Stablecoin de-peg Treasury / Fed / issuers Medium Reserve quality and redemption capacity
Taiwan kinetic escalation Governments / central banks Low Real-economy supply shock

Numerical p_eff values per trigger (base case, used in §7.5 structured MC):

Trigger p_eff (low/base/high)
Treasury basis 0.55 / 0.70 / 0.85
Iran/Hormuz 0.10 / 0.20 / 0.35
AI capex 0.05 / 0.10 / 0.20
Yen carry 0.35 / 0.50 / 0.70
AI cyber 0.15 / 0.25 / 0.40
Private credit 0.35 / 0.50 / 0.65
Stablecoin 0.30 / 0.50 / 0.65
Taiwan 0.05 / 0.10 / 0.20

On the gap between Layer A and Layer B. P(≥2 triggers fire) = 54.6% in Layer A is higher than P(severe drawdown) = 49.9% in Layer B at base slope. Policy offsets are one contributor to that difference. The difference should NOT be interpreted as a direct estimate of "institutional response value," because it also reflects trigger identity (some firings, like minor stablecoin events, transmit little); severity weights (low-severity triggers contribute little to drawdowns); loop activation (only specific combinations engage loops); threshold choice (the ≥25% cutoff is itself a parameter); and stochastic drawdown mapping. Multiple things move at once; the gap is not a clean decomposition.

Policy effectiveness is itself a state variable that can deteriorate. Fed-Treasury coordination strain, BOJ political pressure, fiscal capacity constraints, and transition periods such as Powell-to-Warsh, if confirmed, can all reduce realized policy offset relative to base-case values.

9. Historical Analog Review

This is a qualitative review of how the framework's vocabulary maps to past episodes — not a formal backtest of signal rules. For formal dashboard threshold backtesting, see §12.

Flagging criterion: "Flagged" means that at least one of the framework's named loops or triggers would have been identifiable before the major drawdown using information available at the time. "Flagged: Yes" requires explicit loop activation pre-peak; "Flagged: Partial" requires either loop or trigger identification but with major caveats; "Miss" requires no available framework category.

Episode Framework loop or trigger Flagging
1998 LTCM (S&P -19.3%) Loop E (basis trade) Partial — loop available, but pre-OFR/FSB documentation thin
2000-02 Dot-com (-49%) Valuation amplifier + Loop A analog Partial — valuation extreme identified; specific loop weaker
2008 GFC (-57%) Loop F + CRE amplifier Yes — both pre-identified by 2007
Q4 2018 (-19.8%) No clean trigger; valuation only Miss — framework's main weakness for sentiment-driven
March 2020 COVID (-34%) Exogenous, outside framework scope Miss by design
August 2024 yen carry (-8.5%) Trigger 4 + Loop D Yes — trigger and loop both identifiable
2022 inflation/rate cycle (-25.4%) Loop C analog Partial — oil-inflation channel, but Fed reaction was the proximate cause

Framework strength: structural-vulnerability identification. Framework weakness: sentiment-driven corrections without clean triggers. Out of scope: exogenous shocks.

This is a stress-testing framework. It will produce more risk signals than realized crashes. The honest framing: if a crash occurs, it will most likely involve one or more of the documented loops; the framework does not assert a crash will occur.

10. Monitoring Window: September 2026 – February 2027

September 2026 – February 2027 is a catalyst-dense monitoring window — more high-information events cluster in this period than in adjacent periods:

  • Hyperscaler earnings cycles: Q3 2026 (October-November) and Q4 2026 (January-February). Both critical for AI capex justification.
  • 2027 capex guidance: typically released at Q3 or Q4 earnings calls.
  • BOJ meetings: September 18, October 30, December 18, 2026; further meetings January-March 2027.
  • US midterm elections: November 3, 2026.
  • Fed leadership: assuming confirmation, Warsh would be fully bedded in by Q3 2026.
  • Treasury year-end refunding: completes by December 2026.
  • Iran war anniversaries: 6-month mark August-September 2026; 1-year mark February 2027.

Heuristic catalyst-density comparison — Sep 2026 – Feb 2027 vs. Mar – Aug 2027:

Catalyst type Sep 2026 – Feb 2027 Mar 2027 – Aug 2027
Hyperscaler earnings (Big-Four quarters) 8 4
BOJ meetings 4 3
Fed FOMC 4 4
Major Treasury refunding/auctions 6 4
US elections / fiscal milestones 2 0
Iran war milestones 2 1
Significant 2027 capex disclosures yes typically not
Approximate total ~25-27 ~15-17

This is a heuristic count, not a statistical ranking. Sep-Feb has roughly 1.6x the event density of the adjacent six-month window.

This is not a statistically derived peak in crash probability. The Monte Carlo operates at 6-, 12-, and 24-month horizons without temporal granularity beyond that — the simulation cannot statistically rank the Sep-Feb window against, say, Mar-Aug 2027. The framework asserts no such ranking. The window is valuable because information arrival rate is highest then — conditional probabilities update most through windows like this. The discipline is Bayesian updating of the trigger-probability vector as evidence arrives.

11. Three Conditional Severity Paths (Qualitative Scenario Weights)

Important framing. The scenario weights below are qualitative author-calibrated conditional severity paths given that a material stress episode begins. They are not unconditional 24-month probabilities, are not formally model-derived, and should not be visually mixed with the modeled probabilities in §7. They should not be summed to infer market drawdown odds and are not suitable for portfolio optimization or VaR modeling without external validation.

If a stress episode begins within the 24-month window, the conditional probability that it resolves as a rolling correction vs. a recession-bear vs. a multi-loop systemic event vs. a structural crisis is what §11 describes.

Base case path — Rolling correction (~45-55% conditional path weight, author-calibrated)

S&P 500 retraces 10-20% from current 7,400. Magnificent 7 underperforms broader market by 5-10%. Single-name AI infrastructure casualties (Oracle, possibly a second tier of AI compute companies). Private credit gating spreads to a second tier of BDCs. Volatility elevated to 25-30 range. No multi-loop engagement. Recovery 6-12 months.

Adverse case path — Recession-driven bear (~25-35% conditional path weight, author-calibrated)

Peak-to-trough decline of 25-35%. One major hyperscaler cuts capex by 20%+. Oracle credit downgraded to junk by a major agency. Multiple regional banks need intervention. Industry-wide private credit gating. Magnificent 7 underperforms by 15-20%. Recovery 18-24 months.

Severe case path — Multi-loop systemic event (~10-20% conditional path weight, author-calibrated, conditional also on policy reaction)

Peak-to-trough decline of 40-55%. Triggered by simultaneous activation of 2+ feedback loops — most plausibly Loop C (Iran/oil/Fed) plus Loop A (AI capex/IG supply) or Loop D (yen/Mag 7). Treasury market dislocation requires emergency Fed intervention. Several private credit funds wound down. Recovery 3-5 years.

Tail case path — Structural crisis (~3-7% conditional path weight, author-calibrated)

Decline of 50%+. Loss of Treasury safe-asset status, dollar reserve role, or critical infrastructure. Triggers: Taiwan conflict, Treasury market crisis with fiscal trajectory loss of confidence, systemic AI cyber event. Recovery 5-10 years.

These conditional path weights are author calibration based on historical analog episodes (§9). They are not formally model-derived. They should be read as: given that the framework's identified channels do transmit, this is the rough distribution of severity outcomes.

12. Leading Indicators Dashboard

This is the operational expression of Layer D. The dashboard has been preliminarily backtested against eight selected historical drawdown episodes 1995-2024 using FRED data; it has not been validated across the universe of all S&P 500 drawdowns above pre-specified thresholds. Results below should be read with that caveat.

12.1 Empirical backtest (preliminary)

Tested against: LTCM 1998, Dot-com 2000-02, GFC 2008, Europe 2011, Q4 2018, COVID 2020, 2022 cycle, August 2024.

VIX warning (≥25 sustained 5 trading days) — base case rule:

Episode S&P drawdown Pre-peak sustained crossing? Lead time
LTCM 1998 (S&P -19.3%) -19.3% No (single-day touch April 27, 1998 at 26.09 but not 5d sustained) —
Dot-com 2000-02 (-49%) -49% Yes (1999-10-19) 157 days
GFC 2008 (-57%) -57% Yes (2007-08-15) 55 days
Europe 2011 (-19%) -19.4% No —
Q4 2018 (-19.8%) -19.8% No (sentiment-driven, fast) —
COVID 2020 (-34%) -33.9% No (exogenous) —
2022 cycle (-25.4%) -25.4% Yes (2021-12-06) 28 days
August 2024 (-8.5%) -8.5% No (fast-acting) —

3 of 8 episodes detected pre-peak with 28-157 day lead time. Missed: LTCM 1998 (VIX touched 26 but did not sustain), Europe 2011 (sovereign-debt stress played out below 25-VIX threshold), Q4 2018 (sentiment), COVID 2020 (exogenous), August 2024 (fast-acting).

False positive analysis 1995-2024: 80 sustained-crossing events 1995-2024 (separated by ≥30 days for distinct-event counting); 42 inside episode windows; 38 false positives. Over 30 calendar years, that is ≈1.27 false positives per year.

Threshold and duration sensitivity:

Threshold Duration Episodes detected pre-peak Median lead (d) FP/year
22.5 3 5/8 57 1.87
22.5 5 4/8 43 1.73
22.5 7 3/8 53 1.57
25.0 3 3/8 57 1.43
25.0 5 3/8 55 1.27
25.0 7 1/8 53 0.90
27.5 3 2/8 42 0.93
27.5 5 0/8 — 0.50
30.0 5 0/8 — 0.27

The sensitivity table makes clear that lowering the threshold (22.5 instead of 25) detects more episodes but at materially higher FP rate. Raising it (27.5 or 30) sacrifices nearly all pre-peak signal. 25.0 sustained 5 days is a defensible compromise but is not magic.

10y Treasury warning (level ≥ 4.75%): detected in 3 of 8 episodes (LTCM, dot-com, GFC) with 107-270 day lead time; 5 missed. The level threshold is too restrictive — it fires only when nominal rates are in a high regime. A 30-day rate-of-change variant would generalize better; flagged for future revision.

WTI ≥ $105 (Brent ~$110 proxy): detected in 1 of 8 episodes (Europe 2011 at 28 days lead); generally a structural threshold not a leading indicator.

HY OAS ≥ 4.5%: the ICE BofA US HY OAS series (FRED BAMLH0A0HYM2) extends back to December 31, 1996, so this can in principle be backtested. The threshold of 4.5% was crossed in 1998 (LTCM), 2000-02 (dot-com), 2008-09 (GFC), 2011 (Europe), 2015-16 (energy), 2020 (COVID), 2022 (cycle). Lead-time sustained-crossing tests are flagged for full validation; preliminary inspection suggests 5 of 8 episodes detected but with longer lead times (90-300 days) and higher FP rate than VIX.

Dashboard validation summary (preliminary):

Indicator True positives False positive rate Recommendation
VIX ≥25 sustained 5d 3 of 8 ~1.27/year KEEP — best-performing selected-sample indicator
10y Treasury ≥4.75% 3 of 8 ~0.3/year REVISE — restrictive level; use 30d rate change
WTI ≥$105 (Brent proxy) 1 of 8 ~0.1/year KEEP as structural signature
HY OAS ≥4.5% 5 of 8 (preliminary) TBD INCLUDE; full validation pending

This is "preliminarily backtested on selected historical episodes," not "empirically validated." A full backtest across all S&P 500 drawdowns ≥10%, ≥15%, ≥20%, ≥25%, and ≥30% in the FRED VIX era (1990-2026) with precision/recall and full lead-time distribution is the next iteration.

12.2 Tier 1 — Daily / weekly indicators

Indicator Source Current (May 8-11, 2026) Warning Action Cadence
MOVE Index Cboe / ICE BofAML ~67 >130 sustained 5d >150 Daily
VIX Cboe / FRED 17.19 (May 8) ≥25 sustained 5d ≥35 Daily
VIX9D / VIX30 Cboe normal inverted >3d inverted >7d Daily
Brent crude ICE $101.29 (May 8) >$110 >$125 Daily
HY OAS FRED BAMLH0A0HYM2 2.81% (May 8, 2026) ≥4.5% ≥5.5% Daily
JPY/USD 3m basis Bloomberg / OFR tight widening >20bp >40bp Weekly
10y Treasury FRED DGS10 4.38% (May 8, 2026) >4.75% or +50bp/30d >5.0% Daily

12.3 Tier 2 — Weekly / monthly positioning and structural

Indicator Source Action threshold Cadence
Hyperscaler FCF/capex Company filings 2 quarters of widening gap Quarterly
BDC NAV discount NAV reports >10% sustained Weekly
JGB 10y Bloomberg >2.5% Daily
Treasury auction tail Treasury Direct 2 consecutive sub-2.4x bid-cover Per auction
Oracle CDS Markit >150 bp Weekly
Brent curve backwardation ICE Steepening trend Weekly

12.4 Tier 3 — Quarterly structural

Indicator Source Action threshold
Hyperscaler revenue/capex Company filings Below 1.0 sustained
Bank stress test Fed Regional CRE failures
Moody's BDC outlook Moody's Continued Negative
Private credit defaults Fitch / Moody's >12%
Coordinated central bank pivot Statements Joint dovish action

12.5 Dashboard interpretation — threshold cluster to scenario mapping

Cluster pattern Scenario interpretation Indicated drawdown range
0 tier-1 warnings, 0 tier-2 warnings Base case stable ≤10% rolling
1 tier-1 warning Single-vector elevated risk 10-15% near-term plausible
2+ tier-1 warnings OR 1 tier-1 action Multi-vector stress emerging 15-25%; tactical hedging
2+ tier-1 actions OR 2+ tier-2 actions Active stress, multi-loop possible 25-35% Adverse range; defensive review
3+ tier-1 actions + tier-2 confirmations Severe-case territory 35-55% Severe range; full defensive review
Multi-tier across all categories + CB intervention Tail-case territory 50%+; capital preservation review

12.6 Update protocol

  • Daily review: Tier 1 indicators against thresholds.
  • Weekly review: All Tier 1 + Tier 2 plus narrative news.
  • Monthly review: Full dashboard + framework parameter reassessment.
  • Quarterly review: Backtest of dashboard signals against realized market behavior; threshold recalibration if signal-noise has shifted.

The dashboard is a live monitoring tool, not a one-time snapshot. Its utility comes from disciplined application over time.

13. Falsification and Bull-Case Upgrade Criteria

Reframed as Bayesian posterior updating: framework parameters revise as evidence accumulates, with explicit decision thresholds. Brier Skill Score against trigger-specific climatology serves as the calibration discipline.

13.1 Trigger marginal probability updating

Each trigger marginal probability is reviewed quarterly. If observed conditions move materially away from the anchor (§5), the marginal is revised.

AI capex risk downgraded if:

  • Q4 2026 hyperscaler 2027 capex guidance materially above $725B with documented revenue backing; AND
  • OpenAI revenue growth at or above 50% annualized through 2026 with verified enterprise conversion; AND
  • Oracle credit rating maintained at Baa2 or above through 2027; AND
  • Capex-to-FCF ratios stabilizing at Big-Four hyperscalers.

Treasury / repo basis trade risk downgraded if:

  • SEC central clearing transition (Dec 31, 2026 for cash; Jun 30, 2027 for repo) completes without disruption; AND
  • 2-year period without Treasury auction tail or repo funding spike.

Yen carry risk downgraded if:

  • BOJ pause or reversal of normalization through 2026-2027; AND
  • USD/JPY stable in 145-160 range without 5%+ moves; AND
  • Japan-US rate differential stable.

Iran/oil amplifier downgraded if:

  • Hormuz reopening with traffic returning to >70% of pre-conflict levels; AND
  • Brent sustained below $75 for >3 months; AND
  • Stated Iran-US ceasefire holding for >6 months.

13.2 Framework-wide downgrade conditions

Observed (24-month rolling) Framework adjustment
0 triggers fire, all amplifiers stable Marginals reduced 20-30% across the board
0 triggers fire, amplifiers materially soften (CAPE <30, Buffett <180, VIX <14 sustained) Marginals reduced 40-50%; correlation reduced
1-2 triggers fire but resolve without contagion Modal case; no substantial framework change
2+ triggers fire with multi-loop engagement Framework supported for stress-period channel identification; calibration sharpens

The Layer C regime-conditional finding (§7.3) is itself empirical and does not "falsify" in the model-output sense — but the realization in this cycle either confirms the regime base rate or supplies new evidence to revise it.

13.3 Calibration scoring: Brier Skill Score

Brier Skill Score against trigger-specific climatology:

BSS_trigger = 1 − BS_model / BS_climatology
  • BSS > 0: framework improves on the base-rate forecast for that trigger.
  • BSS = 0: framework is equivalent to always predicting the base rate.
  • BSS < 0: framework is worse than the base rate and should be recalibrated.

Computed per trigger, not as composite — trigger base rates differ by orders of magnitude.

Forecast-horizon scoring rule. The framework forecasts 12-month event probabilities. Quarterly score evaluations convert to quarterly hazards before scoring: p_quarter = 1 − (1 − p_12m)^(3/12).

Operational scoring table. Because each trigger has a different natural base rate and no common event universe, scoring is trigger-specific. The initial climatology proxy below is the frozen 12-month base marginal from §5; quarterly scoring converts it to a quarterly hazard using the formula above. Dashboard updates should move the live posterior only within each trigger's low/base/high range unless the forecast log records a structural recalibration.

Trigger 12m climatology proxy / initial prior Low-base-high range Dashboard update rule for live posterior Scoring horizon
AI capex disappointment 25% 15 / 25 / 40% Move toward high if hyperscaler FCF/capex gap widens for 2 quarters or Oracle CDS/rating stress worsens; move toward low if §13.1 downgrade criteria hold Quarterly hazard from 12m prior
Iran/Hormuz escalation 30% 20 / 30 / 45% Move toward high if Brent >$110 or further kinetic escalation occurs; move toward low if Hormuz reopening and Brent/ceasefire criteria hold Quarterly hazard from 12m prior
Treasury basis unwind 12% 8 / 12 / 20% Move toward high if MOVE >130 sustained, auction tails cluster, or repo funding spikes; move toward low after clearing transition proceeds without funding stress Quarterly hazard from 12m prior
Yen carry unwind 18% 10 / 18 / 30% Move toward high if JPY basis widens >20 bp, BOJ hawkishness rises, or USD/JPY moves abruptly; move toward low if BOJ pause/stability criteria hold Quarterly hazard from 12m prior
AI-enabled cyber incident 10% 5 / 10 / 20% Move toward high after material financial-sector cyber disclosure or shared-infrastructure incident; move toward low after four clean quarters without financial-sector severity Quarterly hazard from 12m prior
Private-credit gating 15% 8 / 15 / 25% Move toward high if BDC discounts widen, redemption queues rise, or defaults >12%; move toward low if redemptions normalize and defaults decline Quarterly hazard from 12m prior
Stablecoin de-peg with contagion 5% 2 / 5 / 12% Move toward high if reserve quality, redemption, or bill-market stress appears; move toward low after stable reserve disclosures and no de-peg stress Quarterly hazard from 12m prior
Taiwan kinetic escalation 9% 4 / 9 / 15% Move toward high after direct military incident or materially higher clash-market pricing; move toward low if gray-zone activity de-escalates and ODNI-style assessment remains benign Quarterly hazard from 12m prior

Rare-event guardrails. For triggers with annual base rates below ~10% (Taiwan, stablecoin, AI cyber), four-quarter samples have very low statistical power. BSS is read as a diagnostic rather than an automatic recalibration trigger for these. Rolling 3-year and 5-year calibration windows are weighted more heavily.

Decision rules:

  • For common-base-rate triggers (annual base rate >15%): if BSS < 0 over 4 consecutive quarters, revise marginal toward base rate.
  • For rare-event triggers: BSS is diagnostic only; do not over-react to no-event quarters.
  • If BSS < 0 sustained over 4 quarters for ≥3 common-base-rate triggers, framework requires structural recalibration.

Log score as complementary diagnostic, computed as −Σ y_i log(p_i) − (1 − y_i) log(1 − p_i). Log scores penalize highly confident wrong predictions more than Brier scores; informative when the framework moves marginal probabilities far from base rate.

13.4 Auditable forecast log

The framework maintains a quarterly forecast log to make Bayesian updating auditable:

Date Trigger Prior probability Evidence update New probability Reason
2026-05-12 (baseline as published) — — see §5 Initial publication
2026-08 (Q3 review) — — — — First scheduled review

This table is the framework's single source of truth for revisions and should be appended (not edited) at each review.

13.5 Out-of-scope events

The framework cannot anticipate exogenous shocks (March 2020 analogs) or sentiment-driven corrections without clean trigger mapping (Q4 2018). Realized outcomes in these categories should not be used to falsify or validate the framework; they are outside its scope. The dashboard backtest in §12.1 confirms this: VIX missed Q4 2018 (sentiment) and COVID (exogenous) by design.

13.6 Bull-case upgrade criteria

Symmetric with §13.1-13.2. The following observations should upgrade the bull case (subjective adjustment of approximately 20-30% reduction in tail-loss probability mass):

Strong bull case upgraded if:

  • Q2-Q4 2026 earnings beat rates remain ≥80% with positive guidance revisions; AND
  • AI capex / FCF ratios stabilize across Big-Four hyperscalers; AND
  • Iran-Hormuz partial reopening with sustained Brent < $85; AND
  • BOJ communicates pause / slowed normalization; AND
  • HY CDX, MOVE, VIX, and JPY basis swaps all return to/stay in normal ranges.

If 4 of 5 conditions hold, shift base-case scenario weight from ~50% to ~65-70%, reduce severe-case weight from ~15% to ~6-8%, and lengthen the monitoring cadence from daily-tier-1 to weekly-tier-1.

This makes the framework symmetric — capable of being shifted by positive evidence as well as negative — rather than only falsifiable.

14. Conclusion

The framework's central empirical claim, restated with appropriate uncertainty:

In the available monthly series (Robert Shiller through September 2023, extended through May 2026 with Multpl/GuruFocus readings), only two distinct episodes have reached CAPE ≥40: January 1999 – September 2000 (peak 44.19 in December 1999) and March 2026 – present. Across the 21 monthly observations of the 1999-2000 episode, 13 were followed by ≥25% peak-to-trough drawdowns within 24 months — a rolling-window coverage rate of 61.9%. Because these 21 monthly observations are highly overlapping within a single completed historical episode (n = 1 at the episode level), this is a descriptive historical stress signal, not a calibrated independent-event probability or confidence-bounded forecast. Current CAPE is in the 40.3-42.2 range (early May 2026, provider-dependent).

What the framework claims:

  1. Layer C-conditional history (the strongest empirical anchor): In the available sample, the only completed CAPE ≥40 regime (January 1999 – September 2000) had a 61.9% rolling-window coverage rate of ≥25% drawdowns within 24 months. n = 1 at the episode level. No statistically meaningful confidence interval is reported, for the reasons documented in §7.3 and Appendix K.

  2. Layer B scenario generator at base slope: Conditional on the frozen parameter set, the scenario-generator output is P(trigger-mediated ≥25% drawdown, 24 months) = 49.9% at slope = 6. This is a scenario output, not a forecast. Slope sensitivity 16.7% (slope=4) to 67.9% (slope=8). The framework's value here is channel identification — AI capex / oil-Fed / Treasury basis / yen carry are the channels most worth monitoring.

  3. Layer A scenario generator: Conditional on the frozen parameter set, P(≥2 modeled triggers in 24 months) = 54.6%. The factor-conditional uplift captures cluster behavior: iran_hormuz firing raises basis_trade probability 1.53×; ai_capex firing raises taiwan probability 1.80×.

  4. Dashboard (selected-sample backtest): VIX warning (≥25 sustained 5+ days) fires before peak in 3 of 8 selected historical drawdown episodes with 28-157 day lead time. False-positive rate ~1.27 per year under the paper's event-counting convention. The rule detects structural drawdowns; it misses sentiment-driven (Q4 2018), exogenous (COVID), and fast-acting (August 2024) drawdowns by design. This is selected-sample backtested, not universally validated.

  5. Prediction-market evidence (Appendix H, May 11-12, 2026 snapshots): AI bubble burst 2026 at 16-28% across market configurations; China-Taiwan military clash before 2027 at 9-12%. Used as low-grade triangulation, not calibration.

What the framework does NOT claim:

The framework does NOT claim a calibrated crash probability. The Layer C 61.9% is descriptive history from one completed episode; the Layer B 49.9% is a scenario output conditional on author-chosen parameters. Neither is an empirically estimated probability of a future market event.

The framework does NOT claim the regime-conditional rate is causal. We cannot separate "high CAPE causes drawdowns" from "the specific episode coinciding with high CAPE." The robust statement is: the only previous time the valuation regime matched today's CAPE ≥40 regime, the subsequent path was a substantial drawdown across nearly every monthly entry point in that episode.

The framework does NOT assert the September 2026 – February 2027 window is the highest-probability crash period within the 24-month horizon. It is catalyst-dense in heuristic terms — earnings, BOJ meetings, midterms, refunding, war anniversaries — but the simulation has no sub-annual temporal resolution and the framework asserts no statistical ranking.

The framework does NOT provide a basis for directional positioning advice. Author priors on trigger marginals (§5) are anchored to evidence but not estimated from history; factor loadings, severity weights, and the slope parameter are scenario assumptions, not estimated coefficients.

What the framework supports as risk-management practice:

  1. Stress-testing portfolios under adverse and severe drawdown scenarios — qualitatively, 25-35% (Adverse path) and 40-55% (Severe path), with conditional-path weights documented in §11. The Layer C historical anchor and the Layer B scenario-generator output both point to elevated fragility; both have material uncertainty; and the right use is stress-testing, not directional positioning.

  2. A risk manager using this framework would test whether liquidity reserves remain adequate under 30-40% equity drawdown scenarios and under private-credit gating assumptions.

  3. Concentration risk warrants explicit stress-testing. With official SSGA SPY top-10 weight at 39.28% and Mag 7 share-class aggregate at 34.69%, cap-weighted vs. equal-weighted broad exposure, jurisdictional diversification, and US-centric concentration are appropriate subjects of stress-test attention.

  4. A live monitoring dashboard (§12, selected-sample backtested) is the operational expression. VIX, 10-year Treasury, HY OAS, and Brent are the FRED-accessible Tier 1 indicators with selected-sample backtest support. Augment with MOVE, JPY basis, Oracle CDS, and BDC NAV discounts where commercial data is available.

  5. Trigger marginal priors are updated quarterly (§13). Real-time prediction-market evidence (Appendix H) is used as triangulation, not calibration. The forecast log (§13.4) makes the framework's revisions auditable.

The paper's contribution is a systemic-risk monitoring framework for high-valuation markets, anchored by descriptive regime-conditional history and operationalized via a selected-sample-backtested dashboard. The framework identifies elevated fragility under high-valuation conditions and maps plausible catalyst pathways. The CAPE result is a high-signal historical screen, not a calibrated independent-event forecast. The scenario-generator layers are useful for channel identification and stress-testing, but should not be read as quantitative probability estimates.

The honest summary: the only previous time markets started from CAPE ≥40, the subsequent path was a substantial drawdown across nearly every monthly entry point in that episode (n = 1 at the episode level); the eight-trigger framework identifies the most likely transmission pathways for today's configuration; the dashboard provides selected-sample-backtested early-warning signals. This is not investment advice. The author is not a registered investment adviser. Readers should consult appropriate fiduciaries for portfolio decisions.

Appendix A: AI Capex Reconciliation

The $650B / $725B / $800B AI capex range used in §4.1. Source-type column distinguishes filing-backed vs. management-guided vs. analyst-estimated vs. author-constructed numbers.

Company 2025 actual (AI/DC, $B) 2026 guidance/estimate (AI/DC, $B) Total capex 2026 (all-in, $B) Source type
Microsoft ~80 $120B+ FY26 (company guidance); ~$190B C2026 (analyst) ~$210-220 Company guidance + analyst extension
Alphabet ~75 $180-190B (raised from $175-185B, Q1 call) ~$200-210 Earnings-call guidance
Amazon ~83 ~$200B (Feb 2026 guidance) ~$220-240 Earnings-call guidance
Meta ~72 $125-145B (raised from $115-135B) ~$135-150 Earnings-call guidance
Big-Four subtotal ~310 ~$650-725B ~$765-820 Aggregated
Oracle ~25 ~$50B+ (Futurum/Introl) ~$50-60 Company commentary + analyst
CoreWeave ~20 ~$30B ~$30-35 IPO disclosures + analyst
Chinese hyperscalers (Alibaba, Tencent, ByteDance) ~40 ~$60-80B ~$80-100 Analyst (Bloomberg / Reuters / 21st Century Business Herald); not filing-backed
Global subtotal ~395 ~$820B ~$925-1015 Aggregated

Definitions:

  • AI/DC capex = portion of total capex attributable to AI infrastructure (servers, GPUs, AI-dedicated data centers, networking, power).
  • Total capex 2026 = full company capex including non-AI components.

Source-type categories:

  • Company filing: SEC filings (10-K, 10-Q, 8-K).
  • Earnings-call guidance: stated by management on quarterly calls; management-attributable but not formally binding.
  • Analyst estimate: published by sell-side or buy-side analysts (Bridgewater, CreditSights, Tom's Hardware, Yahoo).
  • Constructed estimate: built from cross-source triangulation; not from a single primary document. Treat as judgmental order-of-magnitude.

Reconciliation of headlines:

Headline Status
$650B (Bridgewater, Feb 23, 2026) Directly sourced analyst estimate — four Big Tech firms
$725B (Big-Four consensus, post-Q1 2026) Constructed/consensus from post-earnings guides
$750B (CreditSights, April 2026) Constructed/analyst — five hyperscalers including Oracle
$800B (global) Order-of-magnitude estimate including Oracle, CoreWeave, Chinese hyperscalers

The single most directly sourced number is Bridgewater's $650B (Big-Four scope, February 2026); higher figures reflect post-Q1 2026 upward revisions and broader scope inclusion.

Appendix B: Trigger Marginal Probability Evidence

Each marginal probability in §5 with timestamp, source, conversion method, and reproducibility flag. The base probabilities listed are author priors informed by these anchors; they are not mechanically derived from the anchors.

B.1 AI capex disappointment (base = 25%)

Anchor Source Date Implied probability Reproducibility
NVDA option left-tail skew CBOE NVDA puts 12mo ATM-20% May 8, 2026 P(-20%+ move) ≈ 18-22% Proprietary (Bloomberg surface)
AI bubble Polymarket "AI bubble burst in 2026?" May 11-12, 2026 16-28% across configurations Public, direct; multi-outcome market
Sector re-rating historical GS sector rotation, 1980-2020 n.a. 30%+ re-rating from peak in 24mo ≈ 35-45% Indirect; paywalled GS
Burry Substack positions Cassandra Unchained May 10, 2026 Notional $1.1B short via NVDA + PLTR puts (qualitative signal) Public, direct

B.2 Iran/Hormuz escalation (base = 30%)

Anchor Source Date Implied probability Reproducibility
Polymarket peace by May 31 "US-Iran peace deal by May 31, 2026?" May 11-12, 2026 13-28% across configurations Public, direct
Polymarket Hormuz reopen "Strait of Hormuz traffic returns to normal by May 15?" May 11-12, 2026 ~0.7% Public, direct
Polymarket Iran uranium "Iran agrees to end uranium enrichment by June 30?" May 11-12, 2026 25-28% Public, direct
Brent $120 call skew ICE Brent 6mo $120 strike April 2026 P(≥$120 6mo) ≈ 18-28% Proprietary (ICE)
IEA "largest disruption in history" IEA OMR April 14, 2026 April 14, 2026 qualitative Public, direct

B.3 Treasury basis trade unwind (base = 12%)

Anchor Source Date Implied Reproducibility
OFR Brief 26-01 OFR March 3, 2026 75% non-cleared hedge-fund Treasury repo Public, direct
FSB repo report FSB February 4, 2026 ~$16T outstanding; ~70% zero-haircut non-cleared Public, direct
SRF-era historical Fed record 2021-2026 0 realized Public, historical
Pre-SRF historical LTCM, GFC, March 2020 1998-2020 ~1 per 5-10 years Public, historical
MOVE index Cboe MOVE May 8, 2026 67-72 (well below stress threshold) Public, direct

B.4 Yen carry unwind (base = 18%)

Anchor Source Date Implied Reproducibility
USD/JPY 1y implied vol CME / Bloomberg May 8, 2026 P(10%+ JPY appreciation 12mo) ≈ 12-18% (lognormal flat-vol approximation, NOT risk-neutral) Proprietary; illustrative
August 2024 precedent Public market data August 5, 2024 ~5% position unwind → 12.4% Nikkei single-day drop Public, historical
BOJ April 28 dissent BOJ statement April 28, 2026 3 of 9 board members vote for 1.0% Public, direct

B.5 AI-enabled cyber incident (base = 10%)

Anchor Source Date Implied Reproducibility
SEC Item 1.05 cumulative count Debevoise tracker April 10, 2026 ~44 filings since Dec 2023; ~6 financial-sector Public, direct
IMF May 2026 blog Adrian/Gaidosch/Ravikumar May 7, 2026 Qualitative: above historical baseline Public, direct
Anthropic Project Glasswing anthropic.com April 7, 2026 Capability claim (self-reported "thousands of zero-days"), not probability Public, direct

B.6 Private credit fund forced gating (base = 15%)

Anchor Source Date Implied Reproducibility
Fitch default rate Reuters / Fitch March 6, 2026 9.2% (2025) — record in Fitch PMR series Public, direct
FSB private credit report FSB May 6, 2026 $1.5-2T market; ~$220B bank exposure Public, direct
Historical gating events BREIT 2022-23, Bear/Lehman 2007-08 2007-2023 ~1 per 3-5 years in stress Public, historical
BCRED Q1 2026 BCRED filings + CNBC + Bloomberg + AltsWire Q1 2026 $3.8B (7.9% NAV); honored with cap raise + $400M injection Public, direct
BlackRock HPS Q1 2026 Bloomberg March 11, 2026 $1.2B (9.3% NAV) requested; $620M paid Public, direct
Cliffwater Q1 2026 Investment Executive March 2026 14% of NAV; capped Public, direct
Moody's BDC outlook Moody's (Reuters) April 7, 2026 Cut to Negative Public, direct

B.7 Stablecoin de-peg with contagion (base = 5%)

Anchor Source Date Implied Reproducibility
UST/Luna 2022 Public on-chain + market May 2022 Full de-peg; contagion contained to crypto Public
USDC March 2023 Coinbase / Circle March 2023 Brief de-peg, recovered within 48h Public
Tether Q1 2026 attestation BDO Italia end-March 2026 $117B direct T-bills; $141B incl. indirect Public, direct
Stablecoin market cap DefiLlama May 10, 2026 $322.7B total; $189.6B USDT Public, direct

B.8 Taiwan kinetic escalation (base = 9%)

Anchor Source Date Implied Reproducibility
Polymarket China invade June 2026 Polymarket May 11-12, 2026 ~0-1.8% Public, direct
Polymarket China invade end-2026 Polymarket May 11-12, 2026 ~7% Public, direct
Polymarket China-Taiwan military clash before 2027 Polymarket May 11-12, 2026 ~9-12% Public, direct
Polymarket China invade end-2027 Polymarket May 11-12, 2026 ~17-18% Public, direct
Bloomberg Economics scenario Bloomberg February 10, 2026 $10.6T cost / 9.6% global GDP (magnitude, not probability) Public, direct
ODNI Annual Threat Assessment ODNI mid-March 2026 "PLA likely lacks plans for 2027 offensive or sooner" Public, direct

B.9 Reproducibility summary

  • Public, direct, reproducible: 13 anchors (Polymarket, FRED, SEC EDGAR/Debevoise, BIS, OFR, FSB, MBA, IEA, BEA, BLS, anthropic.com, tether.io, BOJ).
  • Public, historical/indirect: 9 anchors (historical events, MBA maturity wall, OFR/FSB reports, ODNI ATA).
  • Proprietary / illustrative: 5 anchors (Bloomberg/ICE/CME option surfaces).

The base-case probabilities sit inside the cluster of empirical anchors for every trigger. They are not formally derived from them — they remain author priors informed by external evidence.

Appendix C: Yen Carry Exposure Disaggregation

"Yen carry" is not a single trade. Approximate gross exposures (order-of-magnitude only — see note below):

Investor type Estimated gross (USD) Funding Likely liquidation asset
Leveraged macro funds $100-200B JPY-denominated repo + JPY-funded swap US Treasuries, Mag 7
Retail FX margin (Japan) $80-150B JPY-margin against USD/JPY long-USD Direct USD/JPY closeout
FX-hedged Treasury portfolios (Japanese institutional) $250-400B (gross hedged) JPY-denominated hedging against USD bonds Treasuries (forced sale if hedge breaks)
Offshore dollar-funding $50-100B Cross-currency basis swaps USD funding stress, basis blowout
Equity-funded through yen $40-80B JPY-denominated equity finance Asian equities, Mag 7
Total approximate gross $520-930B Mixed Multi-asset

The August 2024 tremor primarily affected leveraged macro fund + retail FX margin (~$200-300B subtotal), liquidating an estimated 5% of these positions. A material (>15%) unwind across all segments would imply roughly $80-150B of forced liquidation, with heaviest pressure on Treasuries (from FX-hedged portfolios) and Mag 7 (from leveraged macro and equity-funded carry).

Important note: gross exposure is not equal to forced-sale amount. Many positions can be unwound gradually; only a fraction would be forced-sold in a stress event. The August 2024 precedent suggests ~5% of relevant positions liquidating produced the -12.4% Nikkei single-day move.

Sources (order of magnitude only): Morgan Stanley, Goldman Sachs FX, Bank of Japan FSR, BIS quarterly review.

Appendix D: Code — Layer A Factor-Copula Monte Carlo

"""
Layer A: Factor-copula Monte Carlo for trigger co-occurrence.
Frozen parameter file, May 12, 2026.

Four-factor latent correlation structure:
  F1: Risk-off / global risk aversion    (affects all triggers)
  F2: Inflation / Fed                     (affects oil-driven and credit triggers)
  F3: Liquidity / funding                 (affects Treasury, repo, carry, stablecoin)
  F4: Tech / AI                           (affects AI-specific and Taiwan-via-chips)

Reproducibility: seed = 42, 200,000 trials.
Package versions: numpy 1.26+, scipy 1.12+.
"""

import numpy as np
from scipy.stats import norm

TRIGGER_NAMES = [
    "ai_capex", "iran_hormuz", "basis_trade", "yen_carry",
    "cyber", "private_credit", "stablecoin", "taiwan",
]

# 12-month marginal probabilities (frozen May 12, 2026)
P_BASE = np.array([0.25, 0.30, 0.12, 0.18, 0.10, 0.15, 0.05, 0.09])

LOADINGS = np.array([
    #  RiskOff  Inflation  Liquidity  TechAI
    [   0.40,    0.10,      0.30,      0.65],   # 0 ai_capex
    [   0.40,    0.70,      0.20,      0.10],   # 1 iran_hormuz
    [   0.45,    0.25,      0.70,      0.05],   # 2 basis_trade
    [   0.40,    0.30,      0.55,      0.30],   # 3 yen_carry
    [   0.30,    0.05,      0.20,      0.20],   # 4 cyber
    [   0.40,    0.25,      0.55,      0.20],   # 5 private_credit
    [   0.30,    0.10,      0.50,      0.15],   # 6 stablecoin
    [   0.50,    0.20,      0.20,      0.45],   # 7 taiwan
])


def scale_prob(p_12, horizon_months):
    return 1.0 - (1.0 - p_12) ** (horizon_months / 12.0)


def simulate_factor_copula(probs_12m, loadings, horizon_months, n_sims, rng):
    n_triggers, n_factors = loadings.shape
    probs = np.array([scale_prob(p, horizon_months) for p in probs_12m])
    sumsq = (loadings ** 2).sum(axis=1)
    eps_std = np.sqrt(np.maximum(1 - sumsq, 0))
    F = rng.standard_normal((n_sims, n_factors))
    eps = rng.standard_normal((n_sims, n_triggers))
    X = F @ loadings.T + eps_std[None, :] * eps
    U = norm.cdf(X)
    return (U < probs).astype(int)


if __name__ == "__main__":
    rng = np.random.default_rng(42)
    fired = simulate_factor_copula(P_BASE, LOADINGS, 24, 200_000, rng)
    n_fired = fired.sum(axis=1)
    print(f"P(>=1)={(n_fired>=1).mean():.4f}")  # 0.7665
    print(f"P(>=2)={(n_fired>=2).mean():.4f}")  # 0.5456
    print(f"P(>=3)={(n_fired>=3).mean():.4f}")  # 0.3831
    print(f"E[N]={n_fired.mean():.4f}")          # 2.2407

Verified output (frozen file May 12, 2026, seed=42, 200K trials):

Metric Value
P(≥1) 0.7665
P(≥2) 0.5456
P(≥3) 0.3831
P(≥4) 0.2574
P(≥5) 0.1600
E[N] 2.2407

Average implied binary pairwise correlation: 0.25. Range 0.10 to 0.42.

Appendix E: Code — Layer B Second-Stage Severity Monte Carlo

"""
Layer B: Second-stage severity Monte Carlo for trigger-mediated drawdown.
Frozen parameter file, May 12, 2026.

Takes Layer A trigger vectors and maps through:
  1. Identity-weighted standalone severity
  2. Liquidity coupling
  3. Loop activation (additive boost, capped at 1.5)
  4. Trigger-specific stochastic policy effectiveness
  5. Severity-to-drawdown mapping (slope is calibration parameter)

Reproducibility: seed = 42, 200,000 trials.
"""

import numpy as np

# Identity weights (1-5 scale)
STANDALONE_SEVERITY = np.array([3, 4, 4, 3, 4, 3, 2, 5])
LIQUIDITY_COUPLING  = np.array([3, 4, 5, 4, 4, 3, 3, 5])

# Trigger-specific policy effectiveness (base case)
POLICY_EFF_BASE  = np.array([0.10, 0.20, 0.70, 0.50, 0.25, 0.50, 0.50, 0.10])
POLICY_EFF_STD   = np.array([0.07, 0.12, 0.15, 0.18, 0.12, 0.15, 0.18, 0.07])

# Loop activation rules
LOOPS = {
    "A": {"triggers": [0],    "boost": 0.30},  # AI capex
    "B": {"triggers": [0, 5], "boost": 0.40},  # AI capex + private credit
    "C": {"triggers": [1],    "boost": 0.30},  # Iran/oil/Fed
    "D": {"triggers": [3],    "boost": 0.30},  # Yen carry
    "E": {"triggers": [2],    "boost": 0.50},  # Basis trade
    "F": {"triggers": [5],    "boost": 0.20},  # Private credit
}


def compute_severity_score(fired):
    n_sims = fired.shape[0]
    severity_matrix = fired * STANDALONE_SEVERITY
    primary = severity_matrix.max(axis=1).astype(float)
    n_fired = fired.sum(axis=1)
    additional = 0.5 * np.maximum(n_fired - 1, 0)
    coupling_sum = (fired * LIQUIDITY_COUPLING).sum(axis=1).astype(float)
    n_fired_safe = np.where(n_fired == 0, 1, n_fired)
    avg_coupling = coupling_sum / n_fired_safe
    coupling_term = 0.3 * avg_coupling * (n_fired > 0).astype(float)
    loop_boost = np.zeros(n_sims)
    for ld in LOOPS.values():
        active = np.all(fired[:, ld["triggers"]] == 1, axis=1)
        loop_boost += np.where(active, ld["boost"], 0)
    loop_boost = np.minimum(loop_boost, 1.5)
    return primary + additional + coupling_term + loop_boost


def sample_policy_offset(fired, rng):
    n_sims = fired.shape[0]
    eff = rng.normal(loc=POLICY_EFF_BASE, scale=POLICY_EFF_STD, size=(n_sims, 8))
    eff = np.clip(eff, 0.0, 0.95)
    w = fired * STANDALONE_SEVERITY
    wsum = w.sum(axis=1)
    wsum_safe = np.where(wsum == 0, 1, wsum)
    return np.where(wsum == 0, 0.0, (eff * w).sum(axis=1) / wsum_safe)


def severity_to_drawdown(net_severity, rng, slope=6.0):
    expected = np.minimum(slope * net_severity, 80.0)
    std = np.minimum(2.5 + 0.5 * net_severity, 9.0)
    realized = rng.normal(expected, std)
    return np.clip(realized, 0.0, 90.0)


def run(fired, slope=6.0, seed=42):
    rng = np.random.default_rng(seed)
    gross = compute_severity_score(fired)
    offset = sample_policy_offset(fired, rng)
    net = gross * (1.0 - offset)
    return severity_to_drawdown(net, rng, slope=slope)

Verified output (frozen file May 12, 2026, slope=6.0, seed=42):

Threshold P(drawdown ≥ threshold over 24mo)
≥10% 0.7486
≥15% 0.7163
≥20% 0.6393
≥25% 0.4992
≥30% 0.3380
≥40% 0.1040
≥50% 0.0147

Slope sensitivity: P(≥25%) is 0.1669 (slope=4), 0.3399 (slope=5), 0.4992 (slope=6), 0.6136 (slope=7), 0.6790 (slope=8).

Trigger-count distribution sanity check: P(0) = 0.2335 = 1 − P(≥1) = 1 − 0.7665 ✓

Appendix F: Code — Layer C Historical Coverage Frequency (Monthly Shiller Data)

"""
Layer C: Monthly rolling-window drawdown coverage frequency.
Uses monthly Shiller CAPE (not annual averages).

Data:
  - Shiller monthly file: http://www.econ.yale.edu/~shiller/data/ie_data.xls
    Columns: Date (YYYY.MM), P (S&P composite monthly price), CAPE (column 12)
    Coverage: 1871-01 through 2023-09 in published Yale file.
  - Forward extension Oct 2023 - May 2026 from Multpl/GuruFocus monthly readings.

Reproducibility: any reader can re-run with file downloaded from Shiller's page.
"""

import pandas as pd
import numpy as np

# Parse Shiller's monthly file (Yale download)
raw = pd.read_excel("ie_data.xls", sheet_name="Data", header=None)
data = raw.iloc[8:].copy()
data = data[data[0].notna()].reset_index(drop=True)

def parse_date(d):
    d = float(d)
    year = int(d)
    month = int(round((d - year) * 100))
    if month == 0:
        month = 10  # 1871.10 may parse as 1871.1
    return pd.Timestamp(year=year, month=month, day=1)

shiller = pd.DataFrame({
    "date":  data[0].apply(parse_date),
    "price": pd.to_numeric(data[1], errors="coerce"),
    "cape":  pd.to_numeric(data[12], errors="coerce"),
}).dropna(subset=["price"])
shiller = shiller[shiller["date"] >= "1881-01-01"].reset_index(drop=True)

# Forward-extend Oct 2023 - May 2026 (monthly S&P close + CAPE from Multpl/GuruFocus)
post_2023 = [
    ("2023-10-01", 4275.20, 29.19), ("2023-11-01", 4467.71, 30.50),
    ("2023-12-01", 4682.85, 31.66), ("2024-01-01", 4769.83, 32.32),
    ("2024-02-01", 4953.86, 33.85), ("2024-03-01", 5170.93, 35.32),
    ("2024-04-01", 5099.20, 34.34), ("2024-05-01", 5187.70, 34.55),
    ("2024-06-01", 5360.79, 35.66), ("2024-07-01", 5538.01, 36.85),
    ("2024-08-01", 5523.18, 36.30), ("2024-09-01", 5648.40, 37.07),
    ("2024-10-01", 5762.48, 37.74), ("2024-11-01", 5917.11, 38.78),
    ("2024-12-01", 6040.53, 38.40), ("2025-01-01", 5994.57, 37.93),
    ("2025-02-01", 6037.42, 37.94), ("2025-03-01", 5705.45, 35.34),
    ("2025-04-01", 5440.42, 33.40), ("2025-05-01", 5811.07, 35.13),
    ("2025-06-01", 6022.05, 36.03), ("2025-07-01", 6244.28, 37.06),
    ("2025-08-01", 6395.78, 37.74), ("2025-09-01", 6519.45, 38.10),
    ("2025-10-01", 6720.31, 38.93), ("2025-11-01", 6850.92, 39.28),
    ("2025-12-01", 6980.43, 39.45), ("2026-01-01", 7050.81, 39.65),
    ("2026-02-01", 7140.21, 39.87), ("2026-03-01", 7195.55, 40.11),
    ("2026-04-01", 7325.92, 40.55), ("2026-05-01", 7398.93, 40.42),
]
post = pd.DataFrame(post_2023, columns=["date", "price", "cape"])
post["date"] = pd.to_datetime(post["date"])
shiller = pd.concat([shiller, post], ignore_index=True).sort_values("date").reset_index(drop=True)


def max_dd(p):
    cm = np.maximum.accumulate(p)
    return ((cm - p) / cm).max()


def regime_coverage(prices, capes, window, thresholds, cape_min=None, cape_max=None):
    """Rolling-window coverage frequency, optionally filtered by CAPE regime."""
    n = len(prices)
    dds, n_w = [], 0
    for i in range(n - window):
        c = capes[i]
        if np.isnan(c):
            continue
        if cape_min is not None and c < cape_min:
            continue
        if cape_max is not None and c >= cape_max:
            continue
        dds.append(max_dd(prices[i:i+window+1]))
        n_w += 1
    dds = np.array(dds)
    return {t: (dds >= t).mean() for t in thresholds}, n_w

Verified output (monthly 1881-2026, 1,745 observations):

Regime N P(≥10%) P(≥25%) P(≥40%)
All 1,722 63.0% 19.9% 5.2%
Low CAPE (<15) 690 59.7% 14.2% 1.0%
Medium (15-25) 775 65.0% 24.0% 7.1%
High (25-35) 208 59.6% 19.7% 13.0%
Very high (≥35) 49 91.8% 36.7% 2.0%
Extreme (≥40) 21 100.0% 61.9% 4.8%
Very extreme (≥42) 13 100.0% 69.2% 0.0%

Distinct CAPE ≥40 episodes:

Episode Start End Months Max CAPE Max CAPE date
Dot-com 1999-01 2000-09 21 44.20 1999-12
Current 2026-03 2026-05+ 3 (continuing) 40.55 2026-04

Episode 1 forward 24-month max drawdowns (all 21 monthly observations): range 10.4%-40.9%, mean 26.8%, median 29.7%, 13 of 21 ≥25%, 1 of 21 ≥40%.

Episode 2: May 12, 2026 is the publication date; all 3 starting observations have incomplete 24-month forward windows.

Stationary block bootstrap (mean block = 24 months, B = 2000):

Regime / threshold Point 5th pct 50th pct 95th pct
CAPE ≥40, P(≥25%) 61.9% 57.7% 97.4% 100.0%
CAPE ≥40, P(≥40%) 4.8% 13.4% 74.4% 100.0%
CAPE ≥38, P(≥25%) 55.6% 63.0% 91.7% 100.0%

Non-overlapping windows (sample every 24 months): CAPE ≥40 has 1 truly independent observation (1/1 hit ≥25%, 1/1 hit ≥40%). CAPE ≥35 has 2 independent observations (1/2 hit ≥25%).

Appendix G: Code — Dashboard Threshold Backtest

"""
Dashboard backtest using FRED VIX data.

Tests VIX>=25 sustained 5 days against pre-specified historical drawdowns.
Threshold sensitivity (22.5, 25, 27.5, 30) and duration sensitivity (3, 5, 7, 10).
"""

import urllib.request, io
import pandas as pd
import numpy as np

def fetch_fred(series_id, start="1990-01-01"):
    url = f"https://fred.stlouisfed.org/graph/fredgraph.csv?id={series_id}"
    with urllib.request.urlopen(url, timeout=30) as r:
        df = pd.read_csv(io.StringIO(r.read().decode()))
    df.columns = ["date", series_id]
    df["date"] = pd.to_datetime(df["date"])
    df[series_id] = pd.to_numeric(df[series_id], errors="coerce")
    df = df.dropna()
    return df[df["date"] >= pd.to_datetime(start)].set_index("date")[series_id]

EPISODES = [
    ("LTCM 1998",            "1998-07-17", "1998-08-31", "1998-04-01"),
    ("Dot-com 2000-2002",    "2000-03-24", "2002-10-09", "1999-10-01"),
    ("GFC 2008",             "2007-10-09", "2009-03-09", "2007-01-01"),
    ("Europe 2011",          "2011-04-29", "2011-10-03", "2011-01-01"),
    ("Q4 2018",              "2018-09-20", "2018-12-24", "2018-07-01"),
    ("COVID 2020",           "2020-02-19", "2020-03-23", "2019-12-01"),
    ("2022 cycle",           "2022-01-03", "2022-10-12", "2021-10-01"),
    ("August 2024",          "2024-07-16", "2024-08-05", "2024-05-01"),
]

def detect_pre_peak(series, threshold, duration, episodes):
    for name, peak_s, _, ctx_s in episodes:
        peak = pd.to_datetime(peak_s)
        ctx = pd.to_datetime(ctx_s)
        window = series[(series.index >= ctx) & (series.index <= peak)]
        if len(window) < duration:
            yield (name, None, None); continue
        above = (window >= threshold).astype(int)
        sus = above.rolling(duration, min_periods=duration).sum() == duration
        if sus.any():
            first = sus[sus].index[0]
            yield (name, first, (peak - first).days)
        else:
            yield (name, None, None)

if __name__ == "__main__":
    vix = fetch_fred("VIXCLS", start="1990-01-01")
    print("VIX>=25 sustained 5 days:")
    for name, first, lead in detect_pre_peak(vix, 25.0, 5, EPISODES):
        if first is None:
            print(f"  {name}: not detected")
        else:
            print(f"  {name}: {first.date()}, {lead}d before peak")

Verified output: Dot-com 2000 (-49% drawdown): sustained ≥25 from 1999-10-19, 157 days before peak. GFC 2008 (-57%): sustained ≥25 from 2007-08-15, 55 days before peak. 2022 cycle (-25.4%): sustained ≥25 from 2021-12-06, 28 days before peak. Other 5 episodes not detected pre-peak.

Full-period false positive count 1995-2024: 80 sustained-crossing events (separated by ≥30 days for distinct counting); 42 inside episode windows; 38 false positives; FP rate ≈1.27/year over 30 years.

Sensitivity table: see §12.1 main body.

Appendix H: Code — Real-Time Polymarket Anchor Fetch

"""
Fetch live Polymarket prices for §5 trigger anchors.
Treats prediction-market prices as triangulation, not calibration.
"""

import urllib.request, json

POLYMARKET_API = "https://gamma-api.polymarket.com/markets"
UA = "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36"
MIN_24H_VOLUME = 1000  # liquidity filter

def fetch_universe(limit=1000):
    universe = []
    for offset in range(0, limit, 100):
        params = (f"closed=false&limit=100&offset={offset}"
                  "&order=volume24hr&ascending=false")
        req = urllib.request.Request(f"{POLYMARKET_API}?{params}",
                                     headers={"User-Agent": UA})
        with urllib.request.urlopen(req, timeout=15) as r:
            batch = json.loads(r.read().decode())
        if not batch: break
        universe.extend(batch)
    return universe

def extract_yes_price(market):
    outcomes = market.get("outcomes"); prices = market.get("outcomePrices")
    if isinstance(outcomes, str):
        outcomes = json.loads(outcomes); prices = json.loads(prices)
    if not outcomes: return None
    for o, p in zip(outcomes, prices):
        if o and o.lower() == "yes":
            return float(p)
    return None

# Usage filter example:
# universe = fetch_universe()
# liquid_ai = [m for m in universe
#              if any(kw in m.get("question", "").lower()
#                     for kw in ["ai bubble", "openai", "nvidia"])
#              and float(m.get("volume24hr") or 0) >= MIN_24H_VOLUME]

Snapshot (May 11-12, 2026):

Trigger Polymarket question Implied YES (range across market configs)
ai_capex "AI bubble burst in 2026?" 16-28%
iran_hormuz "US-Iran peace deal by May 31" 13-28%
iran_hormuz "Iran agrees to end uranium enrichment by June 30" 25-28%
iran_hormuz "Strait of Hormuz traffic returns to normal by May 15" ~0.7%
taiwan "China-Taiwan military clash before 2027" 9-12%
taiwan "Will China invade Taiwan by 2026" 7%
taiwan "Will China invade Taiwan by end of 2027" 17-18%

Settlement-text mapping: Each Polymarket market's resolution criteria are documented on the market page. The "military clash before 2027" market resolves YES if there is any direct military exchange between PRC and ROC forces (not gray-zone activity). The "AI bubble burst" market settlement criterion depends on the specific outcome chosen (December 31, 2026 outcome is the leading one with $3M total volume).

Triggers without liquid prediction markets: basis_trade, yen_carry, cyber, private_credit, stablecoin (anchors fall back to historical base rates).

Caveat: Prediction-market prices are sensitive to liquidity, settlement wording, manipulation risk, and participant composition. Different market URLs with similar names produced materially different probabilities for the US-Iran peace and AI-bubble questions. Treat as triangulation, not calibration.

Appendix I: Source Register and Quality Grading

I.1 Source quality grades

Grade Source type Examples
A Primary official FRED, SEC, FSB, BIS, IMF, IEA, OFR, OECD, MBA, BEA, BLS, BOJ, Treasury Fiscal Data, company SEC filings
A-/B+ Market data provider Cboe (VIX), FactSet, ICE (Brent), CME, FRED
B+ Major press with named primary source Reuters, CNBC, Bloomberg, AP citing primary entities
B Analyst research Apollo Academy (Slok), Bridgewater, GS, MS, Bloomberg Economics
C+/B- Prediction markets Polymarket (public settlement criteria)
C Aggregators / interpretive sites Motley Fool, AltsWire, AhaSignals, industry summaries
(D for reproducibility) Proprietary surfaces Bloomberg option surfaces (acknowledged as illustrative)

I.2 Source register

ID Source Supports Grade Type Date accessed
S1 FRED (SP500, VIXCLS, BAMLH0A0HYM2) S&P 500 7,398.93 May 8; VIX 17.19; HY OAS 2.81% May 8 / 2.79% May 7 A Primary May 11, 2026
S2 CNBC live blogs Intraday May 8/10/11; oil prints; cloture vote B+ Major press / primary May 11, 2026
S3 TheStreet / Trading Economics Nasdaq/Russell/DJIA May 8 closes; commodity intraday B+ Major press May 11, 2026
S4 IEA Oil Market Report April 2026 10.1 mb/d supply fall; "largest disruption in history" A Primary May 11, 2026
S5 CNBC; Atlantic Council Birol 12 mb/d (Apr 1) → 13 mb/d (Apr 13) statements B+ Major press / primary May 11, 2026
S6 Reuters / CNBC / Roll Call Powell term ends May 15; Warsh cloture May 11; vote May 12 B+ Major press / primary May 11, 2026
S7 Bank of Japan, Statement on Monetary Policy April 28, 2026 BOJ held 0.75% in 6-3 vote with three dissenters for 1.0% A Primary May 11, 2026
S8 Reuters / MSN HSBC raised S&P target to 7,650 on May 11 B+ Major press May 11, 2026
S8b Benzinga / CNBC Yardeni raised target to 8,250 on May 11; Calvasina (RBC) to 7,900 May 8; JPM 7,600 April 21 B+ Major press May 11, 2026
S9 FactSet Earnings Insight (Butters), May 8, 2026 Q1 2026 84% beat / 18.2% surprise; forward P/E 21.0 A- Market data May 11, 2026
S10 Bridgewater (Jensen letter Feb 23, 2026); CreditSights April 2026 $650B four-firm 2026 AI capex; $750B five-firm B Analyst May 11, 2026
S11 SSGA SPY Holdings (official) Top 10 = 39.28%; Mag 7 share-class aggregate = 34.69% from the rounded holding table (May 8, 2026) A- Issuer official May 11, 2026
S11b Motley Fool Research (Lyle Daly) Mag 7 = 33.7% S&P (April 14, 2026 — historical context only) C Aggregator May 11, 2026
S12 BIS Bulletin 120 (Aldasoro/Doerr/Rees), bis.org/publ/bisbull120.htm, published January 7, 2026 AI financing shift to debt A Primary May 11, 2026
S12b BIS Quarterly Review March 2026 (Aldasoro/Doerr/Rees), bis.org/publ/qtrpdf/r_qt2603u.htm AI on- and off-balance-sheet borrowing follow-on A Primary May 11, 2026
S13 Reuters Alphabet euro/yen bond market activity B+ Major press May 2026
S14 Apollo Academy / Torsten Slok, March 24, 2026 $14T 2026 IG supply estimate B Analyst May 11, 2026
S15 OECD Global Debt Report 2026 (DOI 10.1787/e9d80efd-en) Sovereign / corporate borrowing context A Primary May 11, 2026
S16 MBA, February 9, 2026 / April 27, 2026 $875B 2026 maturities; 4.02% Q1 2026 delinquency A Primary May 11, 2026
S17 OFR Brief 26-01 (Hempel/Kahn/Mann/Paddrik, March 3, 2026) 75% non-cleared hedge-fund Treasury repo A Primary May 11, 2026
S18 FSB Vulnerabilities in Government Bond-backed Repo Markets, February 4, 2026 $16T repo; concentration risks A Primary May 11, 2026
S19 SEC Press Release 2025-43 (Feb 25, 2025) Treasury clearing dates Dec 31, 2026 / Jun 30, 2027 A Primary May 11, 2026
S20 Anthropic, April 7, 2026 (anthropic.com/glasswing) Project Glasswing / Claude Mythos Preview A Primary May 11, 2026
S21 IMF blog Adrian/Gaidosch/Ravikumar, May 7, 2026 AI-enabled cyber financial-stability framing A Primary May 11, 2026
S21b Debevoise & Plimpton tracker, April 10, 2026 ~44 cumulative Item 1.05 8-Ks since Dec 2023; 6 financial-sector B+ Law firm tracker May 11, 2026
S22 Reuters / Fitch, March 6, 2026 US private-credit defaults 9.2% in 2025 (record in PMR series) B+ Major press May 11, 2026
S23 FSB Report on Vulnerabilities in Private Credit, May 6, 2026 $1.5-2T market; ~$220B bank exposure A Primary May 11, 2026
S24 DefiLlama / Cryptonews May 10, 2026 Stablecoin total $322.7B; USDT $189.6B C+ Aggregator May 11, 2026
S25 Tether Q1 2026 attestation (BDO Italia), tether.io $117B direct T-bills; $141B incl. indirect B+ Issuer attestation May 11, 2026
S26 White House / Federal Register, July 18, 2025 GENIUS Act signed A Primary May 11, 2026
S27 Polymarket Real-time prediction-market anchors C+ Prediction market May 11-12, 2026
S28 BEA, April 9, 2026 (third estimate) Q4 2025 GDP +0.5% annualized A Primary May 11, 2026
S29 BLS, May 8, 2026 April 2026 nonfarm payrolls +115k; unemployment 4.3% A Primary May 11, 2026
S30 Shiller monthly data (Yale; multpl.com; GuruFocus) Shiller monthly CAPE through September 2023, extended through May 2026 A Primary plus extension May 12, 2026
S31 Treasury Fiscal Data; JEC Debt Dashboard $38.91T federal debt May 5, 2026 A Primary May 11, 2026
S32 CNBC / Bloomberg / AltsWire BCRED Q1 2026: 7.9% NAV ($3.8B); cap 5%→7%; $400M injection B+ Major press May 11, 2026
S33 Bloomberg / Reuters / Alternative Credit Investor BlackRock HPS, Morgan Stanley North Haven, Cliffwater, Blue Owl Q1 2026 stress; Moody's BDC outlook to Negative April 7, 2026 B+ Major press May 11, 2026
S34 FAU CRE screener (Rebel Cole, Q3 2025 vintage) 59 of 155 largest banks CRE >300% equity C Academic screener May 11, 2026
S35 FDIC FIL-23-2023 300% CRE/capital threshold from 2006 joint guidance A Primary May 11, 2026
S36 Bloomberg Economics February 10, 2026 Taiwan war scenario $10.6T / 9.6% global GDP B Major press / analyst May 11, 2026
S37 Barclays Fixed Income Research (Andrew Keches), November 11, 2025 Oracle debt to Underweight C Analyst (secondary press) May 11, 2026
S38 Moody's Credit Outlook, December 15, 2025 / September 2025 Oracle Baa2, outlook Negative A Primary May 11, 2026
S39 TradingKey / Foreign Policy Journal Burry Substack May 10, 2026 disclosure ~80% Scion in NVDA + PLTR puts B+ Major press May 11, 2026
S40 AP via US News, August 5, 2024 Nikkei -12.40% single-day B+ Major press May 11, 2026

Claims that depend on lower-grade sources are flagged in-text where used. Source-type column distinguishes whether the source supports a fact (e.g., FRED for VIX), an estimate (e.g., Bridgewater for capex), or a probability (e.g., Polymarket for prediction-market priors).

Appendix J: Reproducibility Manifest

This manifest documents the data, code, and parameter file used to generate the framework's numerical outputs. Any reader can reproduce the published tables from these inputs.

J.1 Data sources and download dates

Dataset Source URL Download date File / endpoint
Shiller monthly CAPE 1881-2023 http://www.econ.yale.edu/~shiller/data/ie_data.xls May 12, 2026 ie_data.xls sheet "Data"
Shiller monthly CAPE Oct 2023 - May 2026 (extension) https://www.multpl.com/shiller-pe; https://www.gurufocus.com/economic_indicators/56/sp-500-shiller-cape-ratio May 12, 2026 Tabular series, monthly closes
S&P 500 monthly close 1881-2023 Same as Shiller ie_data.xls column 1 May 12, 2026 —
FRED VIX (VIXCLS) https://fred.stlouisfed.org/series/VIXCLS May 12, 2026 CSV daily 1990-01-02 to 2026-05-08
FRED HY OAS (BAMLH0A0HYM2) https://fred.stlouisfed.org/series/BAMLH0A0HYM2 May 12, 2026 CSV daily 1996-12-31 to 2026-05

J.2 Frozen parameter file (May 12, 2026)

Parameter Value
Random seed (numpy default_rng) 42
Monte Carlo trials per run 200,000
Layer A 12-month marginals (P_BASE) [0.25, 0.30, 0.12, 0.18, 0.10, 0.15, 0.05, 0.09]
Layer A factor loadings (LOADINGS) 8×4 matrix, see Appendix D
Layer B severity weights (STANDALONE_SEVERITY) [3, 4, 4, 3, 4, 3, 2, 5]
Layer B liquidity coupling [3, 4, 5, 4, 4, 3, 3, 5]
Layer B policy effectiveness mean (POLICY_EFF_BASE) [0.10, 0.20, 0.70, 0.50, 0.25, 0.50, 0.50, 0.10]
Layer B policy effectiveness std (POLICY_EFF_STD) [0.07, 0.12, 0.15, 0.18, 0.12, 0.15, 0.18, 0.07]
Layer B base slope 6.0
Layer B slope sensitivity range {4.0, 5.0, 6.0, 7.0, 8.0}
Layer C window sizes {12, 18, 24, 36} months
Layer C drawdown thresholds {10%, 15%, 20%, 25%, 30%, 40%, 50%}
Layer C CAPE regime cuts <15, [15,25), [25,35), ≥35, ≥38, ≥40, ≥42
Bootstrap method Stationary block bootstrap (Politis-Romano)
Bootstrap block length (mean) 24 months
Bootstrap replications 2,000

J.3 Package versions

Package Version
Python 3.11+
numpy 1.26+
scipy 1.12+
pandas 2.0+
xlrd 2.0+ (Shiller .xls parser)

J.4 Script execution order

To regenerate all published tables:

  1. Download ie_data.xls from Shiller's Yale page → place in working directory.
  2. Run cape_step1_load.py (Appendix F base) → produces shiller_monthly.pkl and CAPE episode table.
  3. Run cape_step2_rolling.py → produces unconditional and regime-conditional tables (§7.2, §7.3).
  4. Run cape_step3_bootstrap.py → produces block-bootstrap CIs.
  5. Run layer_a_mc.py (Appendix D) → produces fired_layer_a.npy and Layer A tables.
  6. Run layer_b_mc.py (Appendix E) → consumes fired_layer_a.npy, produces Layer B tables.
  7. Run dashboard_backtest.py (Appendix G) → produces VIX backtest table.

All scripts are deterministic given the seed and inputs.

Reproducing the §7.5 global sensitivity table. Each variant in the §7.5 table is generated by editing one parameter family in the frozen parameter file and rerunning Layers A and B with seed 42 and 200,000 trials. Specifically:

  • Trigger marginals ±20%: scale the P_BASE array in layer_a_mc.py by 0.8 or 1.2 (clipped to [0, 1]), rerun both Layer A and Layer B.
  • Factor loadings ±20%: scale the LOADINGS matrix in layer_a_mc.py by 0.8 or 1.2 (re-clipping idiosyncratic variance to [0, 1]), rerun both layers.
  • Policy offsets ±20%: scale POLICY_EFF_BASE in layer_b_mc.py by 0.8 or 1.2 (clipped to [0, 0.95]), rerun Layer B only (Layer A unchanged — this is why the table's P(≥2 triggers) column equals 54.6% across policy-offset rows).
  • Loop boosts set to zero / +25%: edit the LOOPS dictionary in layer_b_mc.py to multiply each boost value by 0.0 or 1.25, rerun Layer B only (same Layer A invariance).

Each sensitivity run takes approximately the same wall-clock time as the base run. The published table is the deterministic output of these edits; no additional code changes are required.

J.4a Forward-extension table (October 2023 – May 2026)

The post-Shiller extension is reproduced here so the CAPE-regime filter is auditable without opening the code block in Appendix F. Values are the monthly price/CAPE pairs used in the paper's rolling-window computation.

Tiebreaking rule between providers. Two providers were consulted for the post-Shiller extension: Multpl (multpl.com/shiller-pe) and GuruFocus (gurufocus.com/economic_indicators/56/sp-500-shiller-cape-ratio). Where the two providers agreed on a monthly reading (rounded to two decimals), that value was used. Where they differed by less than 0.10 CAPE points (the typical case), the Multpl reading was used as the primary because Multpl's series anchors to month-end closes consistent with Shiller's published methodology. Where they differed by 0.10 or more (rare — observed in only 2 of 33 monthly observations, both in highly volatile months), the lower of the two readings was used to avoid upward bias in the regime filter. The S&P 500 price input column uses the official month-end S&P 500 close from CBOE/S&P Dow Jones Indices and does not depend on provider choice. No tiebreaking decision changed which months crossed the CAPE ≥40 regime threshold.

Month S&P price input CAPE input
2023-10 4,275.20 29.19
2023-11 4,467.71 30.50
2023-12 4,682.85 31.66
2024-01 4,769.83 32.32
2024-02 4,953.86 33.85
2024-03 5,170.93 35.32
2024-04 5,099.20 34.34
2024-05 5,187.70 34.55
2024-06 5,360.79 35.66
2024-07 5,538.01 36.85
2024-08 5,523.18 36.30
2024-09 5,648.40 37.07
2024-10 5,762.48 37.74
2024-11 5,917.11 38.78
2024-12 6,040.53 38.40
2025-01 5,994.57 37.93
2025-02 6,037.42 37.94
2025-03 5,705.45 35.34
2025-04 5,440.42 33.40
2025-05 5,811.07 35.13
2025-06 6,022.05 36.03
2025-07 6,244.28 37.06
2025-08 6,395.78 37.74
2025-09 6,519.45 38.10
2025-10 6,720.31 38.93
2025-11 6,850.92 39.28
2025-12 6,980.43 39.45
2026-01 7,050.81 39.65
2026-02 7,140.21 39.87
2026-03 7,195.55 40.11
2026-04 7,325.92 40.55
2026-05 7,398.93 40.42

J.5 Exact list of CAPE ≥40 monthly starting dates

1999-01-01  1999-02-01  1999-03-01  1999-04-01  1999-05-01  1999-06-01
1999-07-01  1999-08-01  1999-09-01  1999-10-01  1999-11-01  1999-12-01
2000-01-01  2000-02-01  2000-03-01  2000-04-01  2000-05-01  2000-06-01
2000-07-01  2000-08-01  2000-09-01     # Episode 1 ends (21 months)
2026-03-01  2026-04-01  2026-05-01     # Episode 2 ongoing as of publication

J.6 Episode 1 (1999-2000) forward 24-month max drawdown outcomes

Full 21-row output from cape_step2_rolling.py, rounded to one decimal point:

Start month Forward 24m max drawdown ≥25%?
1999-01-01 10.4% No
1999-02-01 12.1% No
1999-03-01 20.2% No
1999-04-01 20.2% No
1999-05-01 20.2% No
1999-06-01 20.2% No
1999-07-01 20.2% No
1999-08-01 20.7% No
1999-09-01 29.7% Yes
1999-10-01 29.7% Yes
1999-11-01 29.7% Yes
1999-12-01 29.7% Yes
2000-01-01 29.7% Yes
2000-02-01 29.7% Yes
2000-03-01 29.7% Yes
2000-04-01 29.7% Yes
2000-05-01 29.7% Yes
2000-06-01 31.7% Yes
2000-07-01 39.2% Yes
2000-08-01 39.2% Yes
2000-09-01 40.9% Yes

Mean 26.8%, median 29.7%, range 10.4%-40.9%. 13 of 21 ≥25%; 1 of 21 ≥40%.

J.7 Publication record

Date Status
May 12, 2026 Working paper, first public version.

Appendix K: Statistical Caveats — Why the Bootstrap CI Is Not Reported

This appendix documents in full why the paper does not report a confidence interval for the Layer C regime-conditional drawdown coverage rate at CAPE ≥40.

K.1 The bootstrap output

For completeness, the stationary block-bootstrap output (mean block length = 24 months, 2,000 replications, applied to the 1881-2026 monthly series, consisting of Shiller through September 2023 plus the documented extension) is:

Regime Threshold Point estimate 5th pct 50th pct 95th pct
CAPE ≥ 35 ≥25% 36.7% 60.7% 85.1% 100.0%
CAPE ≥ 38 ≥25% 55.6% 63.0% 91.7% 100.0%
CAPE ≥ 40 ≥25% 61.9% 57.7% 97.4% 100.0%
CAPE ≥ 35 ≥40% 2.0% 38.6% 77.0% 100.0%
CAPE ≥ 40 ≥40% 4.8% 13.4% 74.4% 100.0%

The naive reading would be to present the [57.7%, 100%] interval (CAPE ≥40 / ≥25%) as a "90% confidence interval." That reading is wrong. The interval does not have a frequentist confidence-interval interpretation.

K.2 The diagnostic that exposes the problem

In a well-behaved bootstrap, the point estimate should fall inside the bootstrap distribution — typically near the median. Here, for the CAPE ≥35 regime, the bootstrap 5th percentile (60.7%) lies above the point estimate (36.7%). The 5th percentile being above the point estimate is not a quirk; it is a signature of bootstrap failure. The same pathology appears, less dramatically, in the CAPE ≥35 / ≥40% row, where the point estimate is 2.0% and the bootstrap 5th percentile is 38.6%.

K.3 Why this happens

Stationary block bootstrap on a rare regime concentrated in one historical cluster has a degenerate failure mode:

  1. The regime "CAPE ≥40" exists, in the available history, only in the connected cluster January 1999 – September 2000 (21 consecutive months).
  2. Each bootstrap replication resamples blocks from the full 1,745-month time series with replacement.
  3. Many replications draw zero blocks that overlap the 1999-2000 cluster. In those replications, there are no CAPE ≥40 starting points and the coverage rate is undefined (0/0).
  4. Replications with one or more overlapping blocks tend to include multiple overlapping copies of the same cluster (because resampling with replacement preserves and amplifies dense clusters).
  5. When the cluster is present, the coverage rate within it is empirically very high; when present in multiple copies, the bootstrap statistic concentrates near the empirically observed within-cluster rate.

The resulting bootstrap distribution is bimodal-or-truncated: heavy mass near the within-cluster rate, with the lower tail effectively discarded as undefined. The reported "5th percentile" is therefore not the 5th percentile of a sampling distribution for the true population coverage rate — it is the 5th percentile of a degenerate distribution conditional on the cluster being resampled at all. This is not the kind of interval that supports the standard frequentist interpretation.

K.4 The Wilson interval as a sanity check

If the 21 monthly observations within the 1999-2000 episode were treated as independent Bernoulli trials (which they are not; they overlap heavily), the 90% Wilson interval for 13 successes in 21 trials would be approximately [44.1%, 77.0%]. This is a reasonable order-of-magnitude bound for what a "calibrated" interval might look like under an idealization. But even this is not appropriate, because:

  • The 21 windows are overlapping monthly entry points into the same 24-month-look-ahead horizon, not independent draws.
  • They all come from one historical regime (the dot-com peak), so any underlying generative-process variation is captured by n = 1 episode, not 21 trials.
  • The relationship between starting CAPE level and forward drawdown was almost certainly different in this episode than it would be in a different macro regime.

K.5 What the paper reports instead

The paper reports:

  • The point estimate 13/21 = 61.9% as descriptive within-episode coverage.
  • The non-overlapping-window count: 1 of 1 (the 1999 episode contributes one non-overlapping 24-month window with a 40.9% max drawdown).
  • The episode-level count: n = 1 (one completed episode of CAPE ≥40 in 145 years of monthly data).
  • The explicit caveat: this is descriptive history, not a calibrated probability.

The framework's conclusion does not rest on a statistical confidence interval for this figure. It rests on the substantive fact that the only previous time this configuration appeared, the subsequent path was a substantial drawdown across nearly every monthly entry point in that episode. That fact is informative even though it is unrepeatable history.

K.6 What would change this

A revised analysis could:

  1. Extend the regime threshold downward to CAPE ≥35 or ≥30 to obtain more historical observations, at the cost of moving further from the current configuration.
  2. Use a Bayesian framework with an explicit prior on the future coverage rate, treating the 1999-2000 episode as one observation updating a beta-distribution prior. This would produce a posterior with explicitly stated subjectivity.
  3. Augment with cross-country evidence (Japan 1989, Asian valuation peaks) — at the cost of cross-country regime-comparability assumptions that themselves require defense.

Each of these is a different framework choice with different costs. The paper makes none of them; it reports the within-sample observation honestly and lets the reader draw the appropriate inferences.

End of working paper.