market-status

name: market-status description: Goldman-Sachs-Weekly-Kickstart-style market-exuberance dashboard. Scores how much of a late-cycle euphoria the equity market is currently expressing — using the same 9 indicators GS uses to "evaluate exuberance" (Momentum factor returns, S&P 500 52-week market breadth, GS Speculative Trading Indicator, CBOE Put/Call ratio, median short interest, Yale Stock Market Confidence Indices, AAII Investor Sentiment, US IPO count, Net US equity issuance) plus the broader Kickstart dashboard pages (Sentiment Indicator, Financial Conditions Index, fund flows, GDP nowcast, VIX, realized correlation, SPX vs equal-weight, factor & sector returns, top movers, P/E and ERP, top-10 concentration). Each indicator gets a percentile rank vs its own ~30-year history; a single composite Exuberance Score (0–100) maps to a 5-tier verdict (Frothy / Stretched / Elevated / Neutral / Subdued). Produces a 3,000–5,000 word English markdown report mirroring the GS US Weekly Kickstart layout — Verdict box up top, 9-indicator calibration table (Dot-Com / 2021 / Current), Share-prices / Trading-activity / Investor-sentiment / Corporate-sentiment narrative blocks, plus a "Macro & Cross-Asset" dashboard panel. Use when the user asks "is the market in exuberance?", "is this dot-com / 2021 again?", "are we frothy?", "what does the Kickstart-style dashboard say?", "/market-status", "AI bubble check", "is the rally overextended?", or anything in the bull-market-peak / momentum-mania / late-cycle-euphoria family.

Market Status

Quantify where the US equity market sits on a single late-cycle-euphoria axis, using the same nine indicators Goldman Sachs uses to "evaluate exuberance" in the Weekly Kickstart (2026-06-05 issue). The motivating observation from the GS report:

"A collection of indicators generally signals exuberance that registers above historical averages today but below the levels reached at market peaks in 2000 and 2021. … we evaluate nine metrics across four categories that exhibited a median 100th percentile rank vs. history in 2000 and 95th percentile rank in 2021. Today, those measures rank in the 86th percentile since 1995."

— Goldman Sachs US Weekly Kickstart, 5 June 2026

This skill reproduces that calibration table with public-data proxies and adds the broader Kickstart dashboard pages (Financial Conditions, fund flows, GDP, vol, correlation, factor & sector returns, top movers, P/E & ERP).

This is a regime / macro lens — not a single-name call. It pairs with the project's existing [[market-complacency]] skill (credit-led under-priced-risk read), which approaches the same axis from a different angle.

When to use

The user says any of:

• "Is the market in exuberance / froth right now?"
• "Is this dot-com / 2021 again?"
• "Is the AI rally overextended?"
• "Run the GS Kickstart-style dashboard"
• "Bull-market-peak check"
• "/market-status"
• "Market sentiment dashboard"
• "Are we in late-cycle euphoria?"

The skill produces an Exuberance Verdict (5-tier) backed by a composite percentile score and a per-indicator breakdown calibrated against the two clean historical exuberance peaks (Q1 2000, Q4 2021).

When NOT to use

• The user wants a single-ticker call — use [[trader-plan]] / [[portfolio-decision]] / [[take-profit-lab]].
• The user wants a credit / under-priced risk read specifically — use [[market-complacency]] (credit-led, HY/IG/CCC OAS focused).
• The user wants a sector-specific froth read (e.g. "are semis frothy?") — use [[sector-overview]] with a stress lens.
• The user wants to know when the bull market ends — exuberance can persist for quarters (GS: "Speculative mania is a poor timing indicator"). This skill is a state read, not a timing model.
• The user wants short-bias candidates — hand off to [[idea-generation]] after the verdict prints.

Core methodology

The exuberance axis

For each of the nine GS headline indicators, define a direction such that higher complacency-percentile = more exuberant. Then the composite is a simple weighted average of those percentiles, mapped to a 0–100 scale where:

• 0–20: Subdued (bearish positioning, broad pessimism — contrarian buy)
• 20–40: Neutral
• 40–60: Elevated (signs of optimism but no broad froth)
• 60–80: Stretched (multiple GS indicators in the 80–95 percentile band — late-cycle territory)
• 80–100: Frothy (≥ 5 of 9 indicators in the top decile — 2000 / 2021-style euphoria)

The 9 GS Kickstart indicators (Exhibit 3 of the source PDF)

Anchor percentiles below were last verified against the 2026-06-05 issue (zsxq file_id 412458845521488, Exhibit 3 read visually at 300 dpi). They are a convenience copy only — see the transcription rule in Historical anchors.

The composite is a weighted average of the nine indicators' exuberance percentiles. Weights default to equal (1/9 each) so the score remains a faithful average of GS's own categories — the only deviation is dropping an indicator if it cannot be sourced in a given run (with re-normalisation and a disclosure in Data Used).

Broader dashboard panel (qualitative — not part of the composite)

The GS Kickstart's middle pages carry a much larger dashboard. The skill pulls and renders the following as context tables and charts that anchor the report's narrative blocks:

• GS US Equity Sentiment Indicator — combines 9 positioning measures. Proprietary; WebSearch for latest value.
• GS Financial Conditions Index (US FCI) — proprietary. Public substitute: Chicago Fed NFCI (FRED NFCI) tracks the same axis.
• Cumulative 12-month mutual-fund + ETF flows (equities / bonds / money-markets) — ICI weekly. WebSearch for the latest aggregate.
• 10-year US Treasury yield — FRED DGS10.
• Fed-funds market pricing at year-end 2026 — CME FedWatch. WebSearch for current implied rate.
• Consensus forward 4Q real US GDP growth — Atlanta Fed GDPNow (real-time nowcast) or Bloomberg consensus.
• S&P 500 3-month implied volatility (VIX3M) + spot VIX — yfinance ^VIX and ^VIX3M.
• S&P 500 3-month realized average stock correlation — derived from yfinance top-50 SPX members' rolling 3M correlation matrix.
• Concentration of top-10 names in S&P 500 market cap and earnings — derived from yfinance SPX constituents.
• Forward 12-month P/E — multpl.com monthly + factor ETFs for cross-index NTM PE.
• S&P 500 yield gap = (forward EPS yield) − (real 10Y TIPS yield) — derived. FRED DFII10 for the TIPS yield.
• EPS revision breadth — FactSet/IBES is paywall; WebSearch for the latest FactSet Earnings Insight number.
• Sector returns + sector forward P/E — yfinance sector ETFs (XLK / XLF / XLY / XLP / XLV / XLI / XLE / XLB / XLU / XLRE / XLC).
• Factor returns (Momentum / Value / Growth) — yfinance MTUM / IWD / IWF.
• Top-10 weekly movers (largest + smallest 5-day total return among SPX) — derived from yfinance.

Per-indicator percentile mapping

For each indicator with ≥ 10 years of clean history (target: 30 years, matching the GS Exhibit 3 "since 1995" base):

1. Current value — most recent valid observation.
2. Historical percentile rank — where today sits in the full available history (30y for the headline 9; 10y for the broader panel).
3. Exuberance percentile — for "low = exuberant" indicators, 100 − raw_pct; otherwise raw_pct.
4. Decile label — "1st (least exuberant)" through "10th (most exuberant)".
5. Top-decile flag — true if exuberance percentile ≥ 90.
6. GS-divergence disclosure (mandatory, every headline-9 indicator). When an own-calc percentile differs from the GS-published Current percentile (Exhibit 3, transcribed in Step 1.5) by >15 points, the Figure 1 cell MUST show both — e.g. 20 own-calc / 88 GS — with the cell colour taken from the GS value, and the section narrative must explain the divergence (different underlying series, base year, or smoothing). Silently substituting an own-calc value that flips the signal direction is forbidden — the 2026-06-07 run printed put/call at 20 (green contra-signal) where GS prints 88 (red) and built a narrative thread on the wrong cell.

Composite score

Equal-weighted mean of the 9 headline indicators' exuberance percentiles (each 0–100). If an indicator cannot be sourced, drop it from the active set and disclose the count in the Data Used manifest.

Mean vs median — one tier, two numbers. The mean is the tier-determining composite (it maps to the band table below). Wherever the composite is quoted (verdict box, Figure 1 footer row, Data Used), carry the GS-comparable median alongside it on the same line — e.g. "mean 66 → Stretched; median 86, the figure GS quotes" — and never print two different tiers as the verdict. Both numbers must be recomputed from the final Figure-1 Current column (Step 5 arithmetic gate); merge_websearch.py computes only the mean.

The composite tier bands are:

Acceleration note (calibration observation, not a timing call). Mirroring Citi BMC's "once the score breaks its band it tends to accelerate," crossing from Stretched → Frothy (≈ 80) historically precedes faster deterioration than a comparable move at lower bands. State it as a band-crossing observation, never as a forecast — see the Learning from sell-side institutional research section and the "state read, not timing model" guardrail.

Historical anchors

The skill must explicitly anchor today's reading against three clean historical peaks plus today, mirroring GS Exhibit 3 columns:

• Dot-Com peak (Mar 2000)
• Post-COVID peak (Dec 2021)
• Today

For each headline indicator, the calibration table shows the exuberance percentile at each anchor. Historical anchors come from the GS PDF directly when the percentiles are unambiguous — the script does not need to back-fill 1995 history for every indicator. When the script can compute the historical percentile from a long-history series (Momentum / VIX / put-call / 10Y), it does so; when only a current value is reliably fetchable (AAII / Yale / IPO count / net issuance), the script quotes the GS percentile for the historical columns and computes only the current column directly.

Transcription rule (hard). Historical anchor percentiles MUST be transcribed from the actual GS Exhibit 3 each run (Step 1.5) — never trusted from this file, the Figure 1 template, or build_status.py's HEADLINE_INDICATORS dict. The authoritative copy of each week's Kickstart is in db/zsxq.db. If the hardcoded values disagree with the PDF, the PDF wins and the stale copies must be corrected in the same commit. (Past failure: the Momentum and Breadth 2021 anchors sat transposed — 76↔95 — across all three hardcoded copies and flowed verbatim into the published 2026-06-07 report.)

Learning from sell-side institutional research

The GS Kickstart is the primary analog, but the broader sell-side exuberance literature (GS, Citi, Morgan Stanley, J.P. Morgan, Bernstein) sharpens several blocks. Apply these named moves on top of the 9-indicator core — they deepen the read without adding new data sources.

• Earnings-led vs multiple-led diagnostic (the call's spine). Mirror the GS "Evaluating exuberance" core argument — its whole "not yet a systemic top" verdict rests on fwd-EPS being up more than price (e.g. fwd-EPS +16% YTD vs index +8% YTD → "rally is earnings-led, not theme-led"). Report fwd-EPS revision % YTD against index price return YTD as an explicit named diagnostic, and label the froth: earnings-backed = more durable; multiple-driven = more fragile. Compute it in Step 3 and let it color the verdict — this is the one diagnostic that changes what the composite means, not just its level.

• Flows decomposed by holder type (biggest structural gap). Mirror the Morgan Stanley China/HK Flows & Positioning Tracker and the GS Kickstart flows page: replace the single ICI line with a holder-type table — domestic institutional, retail margin/leverage (FINRA margin debt + put/call as the retail-leverage read), ETF growth-vs-dividend rotation, and equity-futures net positioning where a public proxy exists (foreign-active-vs-passive too, where applicable). Each line carries a percentile-vs-history and a deep-URL source. Per J.P. Morgan "Positioning Looks Stretched Only in Isolated Pockets," express positioning as a single history percentile and then decompose into crowded pockets vs under-owned pockets (e.g. mega-cap-AI / leveraged single-stock ETFs crowded vs broad-market underowned) — and frame stretched-but-not-extreme as "buy-the-dip until proven otherwise," not a sell signal.

• Breadth as cross-sectional DISPERSION, not just % above 200dma. Mirror J.P. Morgan Quant "Calm Surface, Violent Undercurrents": a quiet index can mask a 10-year-high in single-stock dispersion. Add a median-stock-vs-index return and single-stock realized vol vs index realized vol read (both derivable from the top-50 SPX members already pulled for the correlation calc). State the insight explicitly when the spread is wide — index flat while the median stock is down is a froth tell that % above 200dma misses.

• IPO COUNT vs PROCEEDS, and gross vs net-of-buyback issuance. Mirror the GS nuance: IPO proceeds at records but IPO count near average, and net issuance low because buybacks exceed new supply. Encode both sub-metrics on the corporate-financing leg with their own percentiles so the "supply pressure" flag reflects the count/value divergence rather than a single read — this is what GS uses to say supply pressure is still contained.

• Region / segment split of the composite. Mirror Citi's Bear Market Checklist region decomposition (US 11.5 vs Europe 5): add a "where is the froth concentrated" line splitting mega-cap AI vs the broad market, so the single US composite isn't read as uniform. Tie concentration to its 2000 benchmark explicitly — quote top-10 weight at each historical anchor (GS: today's concentration is high but still below 2000), not just today.

• Forward-path context — the level the Street is underwriting. Every analog couples the exuberance read to an explicit forward index path. Report the published Street / GS year-end + 12m SPX point target and the EPS-growth assumption behind it (sourced via WebSearch with a date), framed as "the level of exuberance the Street is currently underwriting" — explicitly NOT the skill's own forecast (preserves the "no timing / price call" guardrail and the "analyst's model is not a source" rule; cite GS / consensus, never "our model").

• Behavioral acceleration rule (calibration observation, not a timing call). Mirror Citi BMC's memorable one-liners ("once the checklist breaks its band it tends to accelerate"; "when more flags turn red, stop buying the dip"). For this skill's 0–100 composite, carry a one-line note that crossing from Stretched → Frothy historically precedes faster deterioration — clearly labeled as a calibration observation, consistent with the "state read, not timing model" guardrail.

• House idiom — valuation in standard-deviations, gauges with absolute levels. Quote every valuation in standard-deviations-from-history (and PEG), never a bare multiple — "fwd 12m P/E 22.3x = +3.4 s.d. vs 10y" is the desk idiom and exactly the project's "no absolute threshold without percentile context" rule. Name proprietary gauges explicitly with an absolute level + percentile every time (GS Speculative Trading Indicator, Citi BMC score, MS MSASI); when proxying a proprietary series, say so and show the proxy's own percentile.

Data sources

Hard data (no WebSearch required)

WebSearch fallback (no clean public series)

The build script flags these as requires_websearch: true in the JSON output. The agent then runs WebSearch and patches the JSON before writing the report.

The build script writes a _websearch_queue.json alongside the data JSON. The agent picks it up in Step 2 and resolves each entry.

Database write rule

Read-only against db/indicators.db. The skill does not write to any project DB — every cache lives in oneoff/market_status_<DATE>_*.csv. See CLAUDE.md § "Database Safety".

Workflow

Step 0 — Sanity check the request

Confirm the user wants a macro / regime read, not a single-ticker call. If the prompt names a ticker ("is NVDA exuberant?"), redirect to [[take-profit-lab]] or [[company-research]].

If the user wants a credit-led under-priced-risk read specifically (HY OAS / VIX-driven), use [[market-complacency]] instead.

Step 1 — Run the build script

# As-of today (default)
python3 .claude/skills/market-status/scripts/build_status.py

# Specific date
python3 .claude/skills/market-status/scripts/build_status.py --date 2026-06-07

# Skip the heavy per-constituent breadth + concentration calc (faster)
python3 .claude/skills/market-status/scripts/build_status.py --date 2026-06-07 --quick

What it does:

1. Pulls all hard-data indicators (yfinance + FRED + multpl).
2. Computes the 30-year and 10-year percentile ranks per indicator.
3. Writes the headline 9-indicator calibration table (with GS-anchor values for historical columns when reliable; live values for the current column).
4. Renders charts for the GS Kickstart's main exhibits (Momentum + AI overlay, breadth, Speculative Trading proxy, put/call, short interest, IPO count, net issuance, AAII, Yale, FCI, GDP nowcast, VIX, correlation, factor returns, sector returns, top movers).
5. Emits _websearch_queue.json with the list of indicators that need WebSearch fallback.
6. Writes a structured JSON summary to oneoff/market_status_<DATE>_summary.json.

Outputs (idempotent — same date → same outputs):

• oneoff/market_status_<DATE>_indicators.csv — per-indicator current value, percentile, decile, top-decile flag.
• oneoff/market_status_<DATE>_calibration.csv — 9-row calibration (Dot-Com / 2021 / Current) percentile table.
• oneoff/market_status_<DATE>_panel.csv — broader dashboard panel (FCI / NFCI / GDP / VIX / etc.) current values.
• oneoff/market_status_<DATE>_top_movers.csv — weekly top/bottom 5 SPX movers.
• oneoff/market_status_<DATE>_sectors.csv — sector ETF 1W / 1M / 3M / 12M / YTD returns.
• oneoff/market_status_<DATE>_summary.json — composite score, tier, active indicator count, output paths.
• oneoff/market_status_<DATE>_websearch_queue.json — fallback queue for the agent.
• reports/charts/market_status_<DATE>_*.png — 12-15 charts.

Step 1.5 — Resolve the source PDFs from the zsxq library

The GS Kickstart is the report's primary source — read the real PDF; never cite a guessed external URL.

1. Locate the latest issue (and the Citi BMC cross-check) in the local library — read-only against db/zsxq.db:

   /opt/anaconda3/bin/python3 zsxq_fts.py --query "Kickstart exuberance" --limit 5
   /opt/anaconda3/bin/python3 zsxq_fts.py --query "Bear Market Checklist" --limit 5
   

2. Render and read. The Kickstart PDFs are image-only (fitz.get_text() returns empty) — render the Exhibit pages to /tmp PNGs with fitz (page.get_pixmap(dpi=200-300)) and Read them visually, per the project's three-tier OCR flow (never Tesseract; no DB writes needed for a read).
3. Transcribe Exhibit 3's full grid — all 9 indicator rows × 3 columns PLUS the Median and Average rows and the inversion footnote ("Historical percentiles are inverted for market breadth, CBOE put/call ratio, and short interest. CBOE put/call ratio percentile rank is evaluated vs. history since 1997"). These transcribed values override the hardcoded copies in this file / the template / build_status.py (see the transcription rule under Historical anchors).
4. Citation rule. Cite the GS issue via the zsxq direct-download URL — http://xs-macbook-air.local:5001/zsxq/pdf/<file_id>/<urlencoded-name> (paste find_pdf.py's pdf_url) — plus the issue date and lead author. An external gspublishing.com URL may be added ONLY if it HTTP-validates 200 with a real-browser UA; the guessed deep-URL pattern returns 403 (the 2026-06-07 report shipped it 4× before this rule).

Step 2 — Resolve the WebSearch queue (the agent's job)

The agent (Claude in this conversation) reads _websearch_queue.json and runs WebSearch for each entry. For each result, the agent:

1. Captures a current numerical value with a source URL and a publication date.
2. Patches oneoff/market_status_<DATE>_summary.json — fills value, percentile, source_url, source_date for that indicator.
3. Re-computes the composite if any of the headline 9 had been missing.

Do not skip this step. A composite computed only from the hard-data indicators understates the exuberance score (because the WebSearch indicators tend to be among the most stretched in a Frothy regime). The Data Used manifest must record every WebSearch citation.

Step 3 — Inspect for surprises

Before writing the report, scan the calibration table and the broader panel for:

1. Cross-category divergence — e.g. share-price indicators all in the top decile but corporate-sentiment indicators (IPO count, net issuance) mid-decile. The narrative blocks must call this out.
2. A single indicator carrying the composite — if the composite is 78 but 8 of 9 indicators are mid-50s and one is 100, the composite is misleading.
3. Stale data on a WebSearch indicator — if the WebSearch result is older than 30 days, prefer a more recent secondary source.
4. Earnings-led vs multiple-led diagnostic — compute fwd-EPS revision % YTD against index price return YTD (the GS "rally is earnings-led, not theme-led" check). Label the froth: earnings-backed (EPS ≥ price) = more durable; multiple-driven (price ≫ EPS) = more fragile. This colors the verdict's meaning, not just its level — see the Learning from sell-side institutional research section.
5. Colour-band disagreement with GS Exhibit 3 — any Figure 1 cell that disagrees in colour-band with the Step-1.5 transcription is a surprise that must be called out in the narrative, not absorbed.

Step 4 — Write the report

Save to reports/market-status/market_status_<YYYY-MM-DD>.md. Target 3,000–5,000 words plus 10–15 charts plus 3 tables. Structure mirrors the GS Kickstart:

1. Exuberance Verdict — single bold blockquote (≤ 200 words). First line is the verdict ("Stretched — exuberance score 68 / 100, 86th percentile vs history since 1995"). Then 2-3 sentences naming the most-stretched and most-contra indicators with absolute levels. Then one line of action (3-5 verbs). Mirror Citi-BMC and GS-Kickstart "summary box" style.
   • Bull-market-ending triggers checklist (mini-table, exempt from the ≤ 200-word prose limit — like Figure 1). Mirror the GS US Weekly Kickstart "Evaluating exuberance" lead-in and Citi's Bear Market Checklist: a 3-row table checking the three classic termination triggers, each marked Not-yet / Marginal / Flagged with its sourced metric: (a) fundamentals/earnings weakening (fwd-EPS revision breadth turning negative); (b) surge in new equity supply (IPO proceeds + net-of-buyback issuance); (c) Fed / financial-conditions tightening (NFCI + fed-funds path). GS's June-2026 read marks all three only "marginally elevated → not yet a systemic top" — the checklist makes the verdict a checklist, not just a score.
   • Counter-signal line (mandatory, one sentence). Mirror Citi BMC's "what is NOT yet flagged" and GS's note that median short interest sits at a multi-decade high (lots of money still short = not full euphoria): name the headline indicators currently in the bottom / mid decile so the read is never one-sided. Reinforces the existing "never assert this looks like 2000 / 2021" guardrail.
2. Figure 1. Kickstart-style 9-indicator calibration table — exactly the layout of GS Exhibit 3. Columns: Dot-Com Bubble / 2021 / Current. Rows grouped by Share-prices / Trading-activity / Investor-sentiment / Corporate-sentiment. Cells coloured: red if exuberance pct ≥ 80, amber 60–80, green ≤ 30, plain otherwise. Cite the public source URL inline per row.
   • Basis tag per Current-column cell: 98ᴳ = GS-published (since-1995, from the Step-1.5 transcription); 100ᵖ = own proxy calc (state the window — e.g. MTUM exists only since 2013, so a "10y proxy" tag, never bare under a "since 1995" header). Mixed-basis columns without tags are forbidden.
   • Mandatory table footnote mirroring GS's: "Percentiles inverted for market breadth, CBOE put/call ratio, and short interest; put/call base year 1997; ᵖ rows computed vs own proxy history (window stated)."
   • Sub-metric rows added beyond GS Exhibit 3 (IPO proceeds, gross issuance) must leave their historical anchor cells blank (—) — GS publishes no anchor for them, and an invented anchor attributed to GS is a fabrication.
3. § Share-price action — narrative paragraph + Momentum chart (*_momentum.png) + Breadth chart (*_breadth.png). 200-300 words.
4. § Trading activity — narrative paragraph + Speculative Trading proxy chart + Put/Call chart + Short-interest chart. 300-400 words.
5. § Investor sentiment — narrative paragraph + AAII chart + Yale chart + GS Sentiment Indicator chart. 300-400 words.
6. § Corporate sentiment — narrative paragraph + IPO chart + Net-issuance chart. 200-300 words.
7. § Macro & cross-asset panel — single block with sub-sections for: 10Y + fed-funds path / GDP / VIX & VVIX / flows decomposed by holder type (institutional / retail margin-leverage / ETF growth-vs-dividend / futures net positioning — mirror the MS China/HK Flows & Positioning Tracker, with crowded-vs-under-owned pockets per JPM "Stretched Only in Isolated Pockets") / breadth incl. cross-sectional dispersion (median-stock-vs-index, single-stock vs index realized vol — JPM "Calm Surface, Violent Undercurrents") & top-10 concentration quoted at each historical anchor / S&P 500 P/E + ERP in standard-deviations-from-history / sector returns + factor returns + top movers. 600-900 words. See the Learning from sell-side institutional research section.
8. § Historical anchor — "What does today look like?" — a short close that names 2-3 historical analogs (with their forward 12-month SPX returns) and explicitly states the calibration limit ("the dashboard described the regime N of N historical times; it predicted the path 0 of N times").
   • Forward-path context line — report the published Street / GS year-end + 12m SPX point target and the EPS-growth assumption behind it (WebSearch with a date), framed as "the level of exuberance the Street is currently underwriting" — explicitly NOT the skill's own forecast. Cite GS / consensus, never "our model".
   • One-line earnings-led note — restate the Step-3 fwd-EPS-vs-price diagnostic ("earnings-backed froth is more durable than multiple-driven froth").
   • Paired 利好 / 利空 (catalysts vs downside-risks) bullet block — mirror every GS regional Kickstart's closing pair: a compact, scannable two-column set of bullets, each dated and carrying a deep-URL source (satisfies paragraph-level citation). Replaces diffuse prose with the desk format.
9. ## Data Used / 数据来源清单 — manifest of every source URL + the WebSearch citations.

Further viewing — explainer videos (optional, but default to including)

When this report covers something a reader would struggle to picture from prose alone — what an exuberance indicator actually measures (how the GS Speculative Trading Indicator is constructed, what the CBOE equity put/call ratio captures, how net-of-buyback equity issuance is derived, how a momentum factor is built), a market-structure concept, an unfamiliar positioning measure, or any mechanism behind a number in the dashboard — attach 1–3 short explainer videos (YouTube and/or Bilibili) so the reader can see it, not just read about it. Default to including them on any topic; omit only when the report is purely numeric with nothing worth visualizing.

Videos are a teaching aid, NOT a citation — they live in their own slot, never enter the citation chain, and never carry a number.

• Where: a **Further viewing** bullet list at the end of the section the concept lives in, or a single 📺 note beside the hard concept.
• Durable sources only: the company's own product / IR / engineering channel, an OEM or reputable teardown / cutaway channel, or a well-known explainer channel — not a low-view re-upload that will be deleted or is clearly pirated.
• Validate before committing — 200 OK only. YouTube / Bilibili return 403 to bare urllib, so HTTP-check each URL with a real-browser User-Agent; drop dead / private / region-gated links (a 404 link is worse than none). Flag Bilibili that may need login/VPN outside CN: (Bilibili — may require login/VPN outside CN).
• Label honestly: [<what it shows> — <why it helps>](URL). No statistic, price target, share figure, or growth rate is ever attributed to a video (a video can't be string-matched against its source).

Step 5 — Verify and clean up

• Spot-check ≥ 3 numbers in the report against the indicators CSV.
• Confirm every percentile in the calibration table matches the script's output.
• Arithmetic gate (mandatory). Recompute the mean AND median of the final Figure-1 Current column after all WebSearch patches (merge_websearch.py computes only the mean — do the median yourself). Every quoted composite (verdict box, Figure 1 footer row, Data Used) must equal those recomputed values exactly; any top-decile / bottom-decile COUNT stated in prose must be re-counted against the final column. (Past failure: the 2026-06-07 report quoted "median 88" against a column whose median was 68, and claimed "five of nine in the top decile" when only three cells were ≥ 90.)
• Confirm every WebSearch citation has both a URL and a publication date.
• Append the verification log (mandatory — the report is incomplete without it). End the report with <details><summary>Verification log — YYYY-MM-DD</summary>…</details> listing: (a) every Figure-1 cell vs the Step-1.5 GS Exhibit 3 transcription (not just vs the script CSV — the CSV can carry the error); (b) the recomputed mean/median consistency check; (c) 3–5 random number→URL string-matches; (d) the HTTP status of every new external URL; (e) charts rendered N of M planned, with reasons for misses.
• Stop any test servers used during chart rendering.
• Commit and push per the project's standard workflow (fix: / feat: Conventional Commits, no Co-authored-by footer).

Output Format (mandatory blocks)

Mirror the GS Kickstart's structure precisely — the report is the daily-scan artifact, and matching the format institutional readers already know is the point.

Mandatory blocks, in order:

1. Exuberance Verdict (blockquote, ≤ 200 words) + bull-market-ending triggers mini-table (fundamentals / new-supply / Fed, each Not-yet / Marginal / Flagged) + counter-signal line (≥ 1 indicator in the bottom / mid decile)
2. ## Figure 1. Exuberance — 9-indicator calibration (HTML table, raw HTML — not markdown with emoji dots)
3. ## Share-price action (narrative + 2 charts)
4. ## Trading activity (narrative + 3 charts)
5. ## Investor sentiment (narrative + 3 charts)
6. ## Corporate sentiment (narrative + 2 charts)
7. ## Macro & cross-asset panel (sub-sections + 4-6 charts)
8. ## Historical anchor (short close)
9. Further viewing — 1–3 validated explainer videos (optional, default-include; placed at the end of the section whose concept they explain — see Further viewing)
10. ## Data Used / 数据来源清单 — must carry a Charts rendered: N of M planned (missing: X, Y — reason) line
11. Verification log — <details><summary>Verification log — YYYY-MM-DD</summary> block per Step 5

Chart counts are conditional, not absolute. The per-section counts above apply to hard-data indicators. For each WebSearch-only indicator, either (a) build a sparse chart from whatever history the fetched source provides (Cboe's downloadable daily put/call stats support a 21-day-MA series; Renaissance publishes annual IPO counts by year — a 9-bar chart beats no chart), or (b) state inline *chart unavailable — indicator covered in Figure 1 only* at the section opener. A section silently shipping zero charts against a mandated count is a spec violation. Every chart title carries an as-of date in the GS exhibit style ("as of May 31, 2026") — the source footers already do.

Figure 1 styling

Like the [[market-complacency]] Figure 2 BMC table: raw HTML with inline-style cell backgrounds — not markdown with emoji dots. Required class palette:

• .ms-red — background: #f4a8a8; color: #5a0000 (exuberance pct ≥ 80)
• .ms-amber — background: #ffd17a; color: #5a3300 (60–80)
• .ms-green — background: #b6e3b6; color: #1a4d1a (≤ 30 — counter-signal: bearish positioning)
• .ms-now — border-left: 2px solid #888 (column separator on the Now column)

Each indicator row gets a markdown link to its public source.

Chart styling (mandatory — applies to every chart-producing call)

Per the global CLAUDE.md rules:

1. Every chart has a source annotation in the bottom margin.
2. X-axis is clipped to the intersection of all plotted series' valid data — not the union.
3. Latest data point must cover "now" — spot-check before saving.
4. For derived charts, plot the source/component series in the same chart.
5. Multi-series charts: first = max(d.min() for d in all_series); last = min(d.max() for d in all_series).

What NOT to include

Forbidden blocks (rejected in prior similar skill iterations):

• ❌ Long composite-decomposition tables — implicit in Figure 1.
• ❌ Multiple postscript / version-history blocks — those belong in git commit messages.
• ❌ A separate "Action Implications" prose section — the verdict line carries the action.
• ❌ "What Would Invalidate This Read" multi-paragraph block — collapse into a 2-sentence note in the Historical anchor section.
• ❌ Per-chart narrative paragraphs between charts within the same section — keep the narrative at the section opener.
• ❌ The composite score quoted as a single number without the per-indicator breakdown.

Guardrails

• Never call the Claude API or any LLM API. Per CLAUDE.md: the agent (this conversation) does the analysis directly. The script is pure pandas + yfinance + urllib.
• Never write to any db/*.db file. Read-only against db/indicators.db. New indicator caches live in oneoff/.
• Never silently drop a headline-9 indicator from the composite without disclosure. If GS Speculative Trading proxy fails AND WebSearch returns no result, the report must say "composite computed from 8 of 9 GS indicators" in the verdict and the Data Used block.
• Never assert "this looks like 2000 / 2021". The calibration table does that work — let the reader pattern-match. Naming a specific historical analog in the verdict is over-confident.
• Always carry the counter-signal line. The Verdict must include one sentence naming the headline indicators currently in the bottom / mid decile (the "not yet flagged" evidence — mirror Citi BMC and GS's multi-decade-high short-interest note). A one-sided read that lists only the stretched indicators is rejected.
• Never use absolute thresholds without percentile context. "VIX is 18" means nothing without "which is the 32nd percentile of the last 10 years". The indicator tables must always carry both.
• Stale data — explicit removal rule (applies to EVERY table and chart, not just the composite set). If any source (multpl or otherwise) is more than 60 days stale, the value is removed: composite indicators drop out of the active set with re-normalization and a note; a Macro-panel row whose source is >60 days stale must be replaced with a fresher equivalent (e.g. FactSet Earnings Insight forward 12-month P/E, which the WebSearch queue already fetches) or dropped — a stale value with a parenthetical apology is still a stale value. Every report run must spot-check the publication dates.
• Historical base rates ("N of N times since YYYY") may appear ONLY if (a) computed by build_status.py with the output CSV path cited inline, or (b) carried by an external citation that contains the figure. The Historical anchor table may contain only anchor rows present in the cited source (GS Exhibit 3 = Dot-Com and 2021 only; adding 2007 requires a separate cited source). Forward 12-month SPX returns in that table must be computed from yfinance ^GSPC and labeled as a calc (= close[t+12m]/close[t] − 1). Every paragraph quoting GS-exhibit numbers cites the zsxq PDF URL in that paragraph.
• No "Source: our model" anywhere. Cite the script path for derived calcs, the underlying FRED / yfinance / multpl / Yale / AAII / Renaissance / SIFMA URLs for the inputs.
• Never claim the dashboard predicts the timing of a regime turn. It is a state read. Exuberance can persist for quarters — GS quotes Greenspan's "irrational exuberance" speech of Dec 1996 as preceding the Dot-Com peak by more than 3 years.
• WebSearch citations must include the publication date. A bare URL without a date is not a citation. The Data Used manifest lists each WebSearch result as [Title, YYYY-MM-DD](URL).
• Numerical accuracy: per CLAUDE.md, every number in a paragraph must trace to a URL cited in that same paragraph where the number literally appears.

Output location

Save to reports/market-status/market_status_<YYYY-MM-DD>.md under the project root (create the directory if missing — first report establishes it). The viewer at http://xs-macbook-air.local:5001/reports will surface it under a new "MARKET-STATUS" type (or as "OTHER" until the viewer's bucket map is updated).

Supplementary deliverables:

• Charts: reports/charts/market_status_<DATE>_*.png + *.html (interactive Plotly time-series).
• Build script: .claude/skills/market-status/scripts/build_status.py (reusable via --date YYYY-MM-DD).
• Generated CSVs / JSON: oneoff/market_status_<DATE>_*.{csv,json}.

Update-in-place rule

One report per date. If reports/market-status/market_status_<YYYY-MM-DD>.md already exists for today's date, update it in place rather than creating a parallel copy. Across dates, keep separate files — the historical sequence of verdicts is itself useful context.

Related skills

• [[market-complacency]] — credit-led under-priced-risk read; complement, not replacement.
• [[take-profit-lab]] — single-ticker exit discipline; takes the macro exuberance read as one input.
• [[sector-overview]] — sub-sector stress.
• [[idea-generation]] — turns a "frothy" read into specific short / hedge candidates.
• [[portfolio-decision]] — final position rating; uses the verdict as one macro input.

Source PDF reference

Goldman Sachs US Weekly Kickstart, "Evaluating exuberance", 5 June 2026 (Ben Snider et al.) — used as the indicator catalog and the 2000 / 2021 percentile anchor source. Reproduction restricted; do NOT redistribute the PDF; do cite the issue date and lead author in every report run.