Seven Indicators: Three for Recession, Three for Valuation & Drawdown, One for Cross-Default

The seven readings — October 2026

On the collapse prediction: the valuation family says the market has never offered a worse risk/reward starting point — that part of your thesis is fully supported by 125 years of data. The recession family says no recession trigger is present: recessions and the crashes that come with them are labor-and-credit events, labor is calm, and monthly credit is calm — but daily credit has begun to stir (FSM off zero on high-yield widening alone). The historical pattern to respect is 1998-2000: DRG-equivalent readings stayed extreme for roughly three years while the timing gauges stayed quiet, then the labor gauge turned (Dec 2000, +3 months before the peak) and the drawdown was -49%. Fragility decides how bad; inflection decides when. Right now only the first key has turned.

Logistic probability that a recession begins within 12 months, from five standardized features: 10Y-3M spread (long-history GS10 minus TB3MS), Sahm, industrial production 6m change, building permits YoY, payroll 6m growth. Accuracy: AUC 0.88 in-sample; 0.83 out-of-sample (fit 1959-1999, tested 2000-2026). Only 3.2% of non-recession months ever printed above 50%. Median probability around past peaks: 25% a year before, 29% at -6m, 28% at -3m, 33% at the peak — this model speaks in odds, not alarms, and it rarely exceeds ~35% before the fact. Honest flaw: the yield curve does nearly all the work (curve-only AUC 0.87); the other features mostly calibrate. And because the curve re-steepened in Dec 2024, RPI has already forgiven the 2022-24 inversion — if that inversion was the signal and the lag is simply unprecedented, RPI will be the last to know. The post-inversion clock: re-steepening was 2024-12, 21 months ago on current data. In the four sampled cycles, recession followed re-steepening within 1-11 months — so either this cycle's lag is the longest on record, or the inversion's signal has expired. RPI's low reading assumes the latter; the clock is printed here so the assumption stays visible.

Labor only, 0-100: Sahm scaled (0.50 = full weight), claims YoY, payroll 6m momentum, temp-help YoY (1990+). Pure labor, because every postwar recession is ultimately a labor event and labor is where false alarms go to die. Accuracy at ≥60: matched 8 of 8 recessions since 1970 (all of them), false-positive episodes 3 — 2003 (jobless-recovery echo), Jul 2024 (the supply-driven Sahm summer), and Sep 2025 (a genuine soft patch that healed). Episodes: 1970-02→1969-12 (-2m) · 1974-08→1973-11 (-9m) · 1980-04→1980-01 (-3m) · 1981-10→1981-07 (-3m) · 1990-08→1990-07 (-1m) · 2000-12→2001-03 (+3m) · 2003-02→none · 2007-09→2007-12 (+3m) · 2020-03→2020-02 (-1m) · 2024-07→none · 2025-09→none. Note the +3m leads in 2000 and 2007: claims and temp help turn before headline unemployment does, which is the lead Sahm alone never gives.

Credit channel, 0-100: a widening leg that takes the worse of Baa-Aaa widening from its 12m low and HY daily widening from its 6-month low (the HY leg exists from Oct 2023, FRED's ICE window; before that, Baa carries the history), plus spread level and NFCI. Accuracy at ≥60: matched 6 of 7 recessions since 1970 (the miss is 2001 — the one postwar recession with no financial-stress event to measure; it peaked at 50), false positives 5 (1971 and 1974 recession tails, 1984 Volcker stress, 2011 euro crisis, 2022 inflation shock — real stress, no new recession). Verdict: FSM is a coincident severity gauge, not a leader — when it fires, the question is no longer whether, but how deep. Today it reads 33, and the composition is the story: the entire reading comes from the HY leg — spreads widened off the late-August low while equities sat at record highs, the divergence the tripwire board flagged — while the Baa-Aaa and NFCI legs still read zero. Stress is being priced in the daily market before it reaches monthly credit or bank conditions, which is exactly the blindness the Oct 2026 improvement analysis fixed (the pre-fix gauge read 0). Episodes: 1971-01→none · 1973-09→1973-11 (+2m) · 1974-10→none · 1979-06→1980-01 (+7m) · 1981-10→1981-07 (-3m) · 1984-02→none · 1990-10→1990-07 (-3m) · 2007-12→2007-12 (+0m) · 2011-10→none · 2020-03→2020-02 (-1m) · 2022-07→none

The tier below HY is flashing louder than HY itself. ICE BofA CCC & lower OAS closed Oct 8, 2026 at 12.52%, +3.48pp off its May 2026 low of 9.04% in five months, while BB widened ~0.44pp and BBB ~0.10pp over the same span; the CCC−HY quality gap of 9.37pp is the widest in FRED's window. FRED truncates this series to ~3 years, but TradingView's chart of the same series carries the full history (verified 2026-10-11): since Dec 1996 the all-time high is 36.7% (Nov 2008) and the stress peaks were 23.9% (Dec 2000), 14.5% (Sep 2011), 19.8% (Feb 2016), 17.9% (Mar 2020) and 12.81% (Sep 2022). Today's 12.52% sits 0.29pp under the entire 2022 rate-shock peak, around the 85th percentile of months since 1996 — and in every prior episode, a momentum crossing of ~12% overshot by +1 to +24pp rather than topping (the first >12% monthly close in 2008 came in March, eight months before the peak, with a head-fake dip back below 12% in between). CCC and BBB widening wires joined the tripwire board on 2026-10-11; both CCC wires seeded tripped and BBB is calm, so the stress is so far contained to the weakest tier — BBB joining is the contagion tell.

Yes — this is the Shiller PE indicator, rebuilt as a return gauge rather than a ratio chart. Expected next-10-year real total return implied by log-CAPE, fit on 1365 start months (1900-2016, R² = 0.30 on overlapping data). The quintile ladder is the accuracy claim — monotone across 125 years:

Today's CAPE of 41.8 implies -0.7%/yr real — the lowest expected-return reading in the 1900-2026 sample (1th percentile). Even the richest historical quintile managed +4.0% medians, because it contained 1990s starts rescued by the bubble's second half; a start at 41.9 has no such cohort. Honest precision: the regression's residual SD is 4.6pp, so the ±1 SD band runs -5.4% to +3.9%/yr. The band widens the uncertainty without rescuing the outlook — even its optimistic edge sits below the ~+6.5% long-run real average.

Mean of five extremity percentiles (expanding, no look-ahead): CAPE, Buffett trend deviation, household equity allocation (37.8% of total household assets, Fed Z.1 — the positioning gauge), top-10 concentration (range-normalized across sourced anchors; no continuous official series exists), and margin leverage (FINRA/NYSE margin debt YoY growth, 1959+). The leverage leg today: margin debt hit a record $1,502B in Jun 2026 and stood at $1,454B in Aug 2026, +37.2% YoY (88th percentile since 1959), with debit balances at 7.0x free credit cash balances — the thinnest cash cushion in the dataset. Leverage is the amplifier: it does not start declines, it converts them into forced selling. Accuracy — as a return gauge, excellent:

Accuracy — as a crash predictor, deliberately shown mixed. DRG at the start of every major drawdown:

Valuation-driven tops (1968, 2000, 2022) started from DRG 74-93. Policy/exogenous crashes (1980, 1987, 1990, 2020) started from 29-50. DRG measures how much air is under the market, not whether something will push it. At 95 today, the air is thinner than at any of those peaks.

Your concentration point, quantified with sourced anchors only (J.P. Morgan, S&P Dow Jones Indices, FactSet/RQI, Goldman Sachs, constituent-cap calculations) — no free continuous series exists, and none is fabricated here:

Two facts, and they pull in opposite directions. Cap concentration is unprecedented: 39.2% (Sep 30, 2026), vs. 26.6% at the 2000 peak — and the index's leaders are the same names across the S&P 500, Nasdaq-100, and growth indexes, with a 36-month S&P/Nasdaq return correlation of 0.96. (That correlation's percentile is only 38th since 1971 — the two indexes have moved as one for most of two decades, so sync alone is a weak discriminator; the concentration level is the real break with history.) But the earnings support is real this time: the top 10's cap-vs-earnings gap is roughly 40% of cap vs. ~37% of expected earnings (Goldman, Oct 2026) — a ~3pp unsupported-weight gap, vs. ~12pp in 2000 (27% of cap, ~15% of earnings). Concentration here is a single-point-of-failure risk (a handful of earnings reports now move the index), not the hollow froth of 2000. If those earnings crack, there is no diversified index left underneath — that is the collapse mechanism this gauge watches: top-10 earnings revisions, which the other five indicators would only catch after the fact.

The first six indicators watch the whole market. CDG watches one chain: in the AI buildout, a small number of private hubs (today, effectively OpenAI) have signed purchase commitments now measured in the hundreds of billions; those commitments are financed in part by the hub's own suppliers and investors, and public neoclouds have borrowed against them. The exposure chain runs both directions: hub → neocloud contracted revenue → neocloud bonds, loans and equity → public equity and credit indexes. It is concentration plus circularity plus a funding gap, with every leg exposed to the same counterparties. Spec: cross-default-gauge-spec-2026-10-14.md. Headline discipline: only definitive and closed deals are ever summed; LOIs are never added to totals (announced-but-unfunded headline figures circulating for this chain exceed $1.4T and mix statuses).

Composite level: 76 / 100 (0.35·CON + 0.30·CIR + 0.35·GAP). State: CONFIRMED — the level alone can only reach ARMED; CONFIRMED additionally requires the market layer MKT ≥ 60 plus a defined confirmation event in the trailing 6 months; FIRED requires MKT ≥ 80, a Type-4 hub event, or distress CDS bands. Clearing steps down one rung at a time with hysteresis (two consecutive monthly readings; quarterly for ARMED→NORMAL).

CON 81: CoreWeave top-2 customers = 62% of revenue (Q2 2026 10-Q; top-1 was 67% in FY2025, so direction is improving at the anchor); Applied Digital top-2 = 84% (FY2026 10-K). CON = 0.5·median + 0.5·max of the banded subscores (70, 85). Nebius is excluded — it does not disclose customer revenue shares; it is carried as a commitment memo only (~$32B firm of >$40B reported commitments from Microsoft and Meta). Reverse concentration memo: S&P's Jul 2026 Oracle downgrade named OpenAI at nearly half of Oracle's $638B backlog.

CIR 70 (provisional): 8 qualifying circular-financing events in the trailing 12 months (window from 2025-10-01): C-02, C-03, C-06b, C-08, C-09, B-02, B-03, B-04. Dollar intensity is not computable today — warrant values, backstop utilization, and Nvidia's share of the OpenAI round are undisclosed — so the layer is velocity-scored and flagged provisional until filings disclose the values. The one computable pair, Microsoft/OpenAI (~$13.8B invested vs a $250B Azure commitment, ~6%), is a memo, not the layer.

GAP 75: OpenAI commitments of $750B through 2030 (Reported Jul 2026) = $150B/yr against a ~$50B annualized run rate (net basis, FT, 2026-10-08) — ratio 3.0, band 75. Every figure is a reported estimate, not an audited financial; bases are never blended. Sensitivity (commitment / run-rate basis → ratio (score)): $600B / gross $70B → 1.71 (55) · $600B / net $50B → 2.4 (75) · $750B / gross $70B → 2.14 (75) · $750B / net $50B → 3.0 (75). Even the most favorable pairing scores 55. Anthropic memo (not scored): ratio ~1.4-1.6, band 55.

MKT 73: Oracle 5Y CDS ~261 bps (2026-10-08, two independent reports, band confirmed) → 65; CoreWeave 5Y CDS ~855 bps (2026-08-15) → 80 SINGLE SOURCE, provisional; stale (>45d); primary market → 70 (CoreWeave senior notes at 9.75% Apr 2026 and 9.625% Jun 2026; the Aug 2026 $2.6B term loan flexed ~+100-125 bp wider); CRWV equity drawdown 49.2% from its 52-week high → 75. A CDS band may change only on two independent reports; a single report sets a provisional band with a flag. Readings go stale at 45 days and drop out at 90. The suite's credit tripwires (CCC/HY/BIZD/Fitch) are displayed alongside CDG in the cockpit but are never scored into MKT — counting them twice would double-count the same evidence.

Confirmation events already in the trailing 6 months (why the state is CONFIRMED, not merely ARMED):

• Type 1 (rating action citing hub/concentration exposure), 2026-07-09: S&P downgraded Oracle to BBB- citing, among other things, OpenAI at nearly half of Oracle's $638B backlog.
• Type 2 (new-issue flex >= +50 bp), 2026-08-15: CoreWeave $2.6B first-lien term loan flexed wider (about +100-125 bp) between launch talk and final pricing.
• Type 3 (CDS band crossing, two independent reports), 2026-10-08: Oracle 5Y CDS crossed upward through the 200 and into the 200-300 band across Jul-Oct 2026, confirmed by two independent reports.

What fires it (FIRED): MKT ≥ 80; OR a Type-4 hub event (payment delay, renegotiation, commitment reduction, missed funding milestone, flat/down round, or failed raise — none has occurred); OR Oracle 5Y CDS > 300 bps or CoreWeave 5Y CDS > 1,000 bps on two independent reports. What the gauge is not: not a probability, not a timing tool, not a claim that default is likely; CON and CIR describe an exposure structure, GAP a funding arithmetic, and only MKT says anything about likelihood. It also says nothing about whether the buildout is economically justified — only about what happens to the chain's creditors and shareholders if it is not.

• LIG crossing 60 — the historical "when" signal; watch claims and temp help first, they led in 2000 and 2007.
• RPI pushing through 25-30% and rising — its pre-recession median path; a print above 50% outside a recession has happened in only 3.2% of months.
• FSM pushing through 60 — it has already left zero on the HY leg alone (33); if the Baa-Aaa and NFCI legs join, a market divergence becomes broad financial severity.
• Margin debt rolling over — two consecutive monthly declines from a peak after a +30% YoY surge (the tripwire now in the daily watch). That deleveraging pattern, not the surge itself, is what preceded the ugly stretches in 2000, 2007, and 2022. Jul 2026 dipped once off the record, then Aug rebounded — not yet.
• Top-10 earnings revisions turning down (CSG's tripwire) — the one channel with no historical precedent at this scale.
• CDG state transitions — a step to FIRED (Type-4 hub event, MKT ≥ 80, or distress CDS bands) would mean the AI chain's cross-default risk has market confirmation at distress level; a genuine step-down to ARMED requires MKT < 40 on two consecutive monthly readings. Level moves alone change nothing.
• Until at least one fires, history says extreme valuation alone has coexisted with further gains for years at a time (1997-99, 2020-21). That is not reassurance — it is the documented behavior of the setup you are describing.

• All seven are my constructions (this project), built only from published ingredients; thresholds were set with the historical record in view, so episode statistics are partly in-sample. The split-sample AUC on RPI and the expanding (no-look-ahead) percentiles on DRG are the out-of-sample-honest pieces.
• CDG has no backtest and no threshold history: no modern episode of this structure exists to calibrate against (the closest analogs, vendor-financed telecom 1999-2002 and Enron's off-balance-sheet vehicles, differ in counterparty quality and disclosure). Its thresholds are structural judgments, stated explicitly so they can be disputed. Its CDS inputs are episodic reported readings, not a feed; every figure carries its source and date in the layer files.
• CDG revision triggers (spec Section 10): add Anthropic as a second scored hub when its commitment and run-rate figures are comparably sourced; score CIR intensity when warrant values, backstop utilization, or round allocations are disclosed; add a private-credit/BDC transmission leg if BDC filings begin naming AI-infrastructure exposure; retire or re-weight the gauge if the hub IPO produces audited financials that replace reported estimates with filings.
• Recession samples are small (8-11 episodes). Probabilities and medians are orientation, not law.
• Backtests use revised data; the Oct 2025 unemployment print is missing in FRED (shutdown) and interpolated in labor gauges.
• Concentration history is 11 sourced anchors, not a series; the 2007 anchor does not exist in public sources and is not estimated.
• Margin debt is spliced NYSE legacy (1959-1999) + FINRA (2000+); FINRA publishes with a ~3-week lag, so the leverage leg carries its last computed percentile forward up to 3 months (Aug 2026 data in the current reading).
• FSM credit widening takes the worse of monthly Baa-Aaa and daily HY widening; the daily HY leg is computable only from Oct 2023 (ICE licensing truncates FRED's history), so it is validated on the live 2026 divergence and spot-checked against published ICE BB history (2.02% in Feb 2007 widening to 3.17% by Aug 15, 2007 — credit cracked months before the equity peak that year too), not backtested over decades. Pre-2023 FSM history is unchanged: the Baa leg carried it, and Sep 2008 still reads 100.
• 2020 is included in episode tables but was exogenous; no indicator is graded on it.

Scripts: ~/workspace/valuation-model/work/analysis3.py (RPI/LIG/FSM/RPS/MSI), analysis4.py (DRG), cdg.py (CDG; data: cdg-deals.json, cdg-con.json, cdg-gap-log.json, cdg-mkt-log.json). Concentration research: work/concentration-research.md. Prepared for Michael Davis, 2026-10-10; CDG added 2026-10-14.