Macro Warning

Macro Warning — Daily Pullback / Top-Risk Radar

Goal

Answer 3 questions every market session:

1. Is the market priced for perfection? (valuation extreme)
2. Are we euphoric or panicked? (sentiment / volatility extreme)
3. What sector should I lean into / out of? (rotation tilt)

Output is a single regime tag (🟢 GREEN / 🟡 YELLOW / 🔴 RED) plus specific position-sizing advice.

When to invoke

• Daily batch (recommended): 5pm ET (post-close) or 8am ET (pre-open) via /schedule
• Manual: User asks "macro check / regime read / 该不该入场"
• Pre-event: 24-48h before Fed / CPI / NFP / major earnings stack
• After +/-2% SPX day: regime may have shifted

The Indicator Stack (8 layers, weighted)

Layer 1 — Valuation extremes (highest weight)

Why NDX 38 matters: historical record high is 38.57 (2000 dot-com peak excluded — that was 73). Touching 38 is statistically a 95th-percentile event.

Layer 2 — Volatility (VIX / MOVE / VVIX)

Layer 3 — Sentiment

Layer 4 — Credit / Risk premium

Layer 5 — Currency / carry

Layer 6 — Market internals / breadth

Layer 7 — CTA / systematic flows

Layer 8 — Sector rotation tilt (the action layer)

Run sector-rotation-analysis skill for full breakdown. Quick check:

The composite regime tag

Score each layer 0 (worst), 1 (caution), 2 (good). Sum gives composite:

Asymmetric weighting: if NDX P/E >38 OR VIX <14 OR F&G >85, even with everything else green, regime is YELLOW minimum (top-risk override).

Output format (for batch / agent consumption)

# Macro Warning — [DATE]
## Regime: [🟢/🟡/🟠/🔴] [GREEN/YELLOW/ORANGE/RED]
**Composite score: X / 16**

## Headline (1 sentence)
[E.g., "NDX P/E at 38.1 (95th percentile) + VIX 14 + F&G 78 = late-cycle euphoria, trim AI names."]

## The 8 layers
| Layer | Reading | Score | Note |
|---|---|---|---|
| Valuation | NDX FwdPE 38.1, SPX FwdPE 22.5 | 0/2 | RED — at all-time-high zone |
| Volatility | VIX 14, MOVE 95 | 1/2 | Complacent |
| Sentiment | F&G 78, AAII bull 52% | 1/2 | Greed |
| Credit | HY OAS 285bps | 1/2 | Tight |
| FX | DXY 99, USD/JPY 154 | 2/2 | Stable |
| Breadth | 71% above 200DMA | 1/2 | Healthy but slipping |
| CTA | Max long | 0/2 | Crowded |
| Sector | XLU starting to outperform | 1/2 | Early defensive rotation |
| **Total** | | **7/16** | YELLOW |

## What changed today vs yesterday
- [Bullet: VIX +1.5, F&G +5, etc.]
- [Bullet: any threshold crossings]

## Action items (specific, sized)
- **If long-only book**: Trim X% of [overheated names]; raise cash to X%
- **If options-friendly**: Buy X% notional SPY 60DTE puts at delta 0.25
- **If looking to add**: Wait for VIX > X or F&G < Y before deploying

## Sector tilt
- **Add**: [list with reasoning]
- **Trim**: [list with reasoning]
- **Hold**: [list]

## Catalysts watch (next 7 days)
- [Date]: [Event] → impact

## Quote-worthy summary
"[Single sentence the user can paste into a chat]"

Hard rules

1. Never claim certainty about timing. Say "elevated risk", not "the top is in".
2. Always show all 8 layers even if some are GREEN — paints full picture.
3. Flag changes vs prior reading — that's where the alpha is.
4. Tag the date and time of data pull — markets move fast.
5. NDX P/E >38 OR VIX <14 = automatic YELLOW minimum even if other layers green (override rule).
6. Cite sources — every threshold has a verifiable URL. No fabricated numbers.
7. For batch run, pre-fetch all data in parallel to keep total runtime <60s.

⭐ Canonical data pull: scripts/macro_pull.py

Use this script for all batch runs. It hits direct APIs (no WebSearch, no LLM scraping) and returns a single JSON blob with raw values + deterministic 8-layer scoring.

# Full scan (~30s, includes breadth computation)
/tmp/.insider_venv/bin/python $(ls ~/.claude/{skills,plugins/claude-investment-skills}/macro-warning/scripts/macro_pull.py 2>/dev/null | head -1)

# Fast scan (~5s, skip breadth)
/tmp/.insider_venv/bin/python $(ls ~/.claude/{skills,plugins/claude-investment-skills}/macro-warning/scripts/macro_pull.py 2>/dev/null | head -1) --skip-breadth

# Pipe to jq for inspection
... macro_pull.py | jq '.scoring'

Why direct APIs not WebSearch:

• Reproducible (same input → same output)
• Batch-safe (no rate limit, no LLM token cost)
• Auditable (every value has a known source)
• WebSearch returns stale/summarized data — actual API gives the live number

Data sources used by the script

Output schema

{
  "timestamp_utc": "2026-05-08T...",
  "yf":      {SPY, QQQ, VIX, MOVE, VVIX, ^TNX, ^TYX, DXY, JPY, ETFs...},
  "fred":    {HY_OAS, IG_OAS, DGS10, DGS30, T10Y2Y, DGS2},
  "cnn":     {score, rating, prev_close, prev_1_week, prev_1_month, prev_1_year},
  "multpl":  {shiller_pe, spx_trailing_pe, spx_dividend_yield},
  "breadth": {above, total, pct_above_200dma_top50, missing_tickers},
  "scoring": {
    "layers": {valuation, volatility, sentiment, credit, currency, breadth, cta_flow, sector_rotation},
    "composite": 0-16,
    "regime": "🟢 GREEN | 🟡 YELLOW | 🟠 ORANGE | 🔴 RED",
    "triggers": ["Shiller CAPE 42.05 > 38 (extreme)", "VIX 17.19 < 18 (exit-signal threshold)", ...]
  }
}

Fallback / gaps the script does NOT cover

Example execution for batch agent

# Pseudo-code for the agent
indicators = {
    'NDX_PE':       fetch_yfinance('QQQ').get('forwardPE'),
    'VIX':          fetch_yfinance('^VIX').get('regularMarketPrice'),
    'MOVE':         fetch_yfinance('^MOVE').get('regularMarketPrice'),
    'VVIX':         fetch_yfinance('^VVIX').get('regularMarketPrice'),
    'TNX':          fetch_yfinance('^TNX').get('regularMarketPrice'),
    'TYX':          fetch_yfinance('^TYX').get('regularMarketPrice'),
    'DXY':          fetch_yfinance('DX-Y.NYB').get('regularMarketPrice'),
    'USDJPY':       fetch_yfinance('JPY=X').get('regularMarketPrice'),
    'F&G':          webfetch_cnn_fear_greed(),
    'HY_OAS':       webfetch_fred_BAMLH0A0HYM2(),
    'breadth_200':  websearch_breadth(),
    'CTA':          websearch_cta_flows(),
    'sectors':      [fetch_yfinance(s) for s in ['XLU','XLK','XLY','XLP','SPY','RSP']],
}

# Score each layer
scores = score_layers(indicators)
total = sum(scores.values())
regime = regime_from_total(total, indicators)  # apply override rules

# Compare to yesterday's snapshot (saved to ~/.claude/projects/.../macro_history.jsonl)
yesterday = load_yesterday()
delta = compute_delta(indicators, yesterday)

# Render report
print(format_report(regime, total, scores, indicators, delta))
save_today(indicators)  # for tomorrow's delta

Recommended scheduling (use /schedule skill)

Option A: Pre-market alert (recommended)

Cron: 0 12 * * 1-5 (8am ET = 12 UTC, weekdays)

Routine name: daily-macro-warning-premarket
Repo: claude-investment-skills (this repo)
Model: claude-sonnet-4-6
Allowed tools: WebSearch, WebFetch, Bash, Read
MCP: yfmcp (for ^VIX, ^MOVE, ^VVIX, sector ETFs)
Prompt: "Run the macro-warning skill. Output the structured report. If regime
flipped from GREEN→YELLOW or YELLOW→ORANGE/RED since yesterday, emphasize
the change. If NDX FwdPE crossed 38 in either direction, lead with it.
Save today's snapshot to memory for tomorrow's delta."

Option B: Post-close summary

Cron: 0 21 * * 1-5 (5pm ET = 21 UTC, weekdays)

Same prompt but appends:

• "What broke today (single-day moves >2σ)"
• "Tomorrow's catalysts (CPI/Fed/earnings)"

Option C: Both

Pre-market for action, post-close for reflection. Doubles the cost but catches both windows. Recommended only if user is active trading.

Memory integration

After each run, write a one-line entry to:

~/.claude/projects/-Volumes-workplace-invest/memory/macro_history.jsonl

Format:

{"date":"2026-05-08","regime":"YELLOW","score":7,"NDX_PE":38.1,"VIX":14.3,"FG":78,"HY_OAS":285,"USDJPY":154.2}

This enables:

• Trend tracking ("VIX rising 3 days in a row")
• Threshold crossing alerts ("NDX PE crossed above 38 today")
• Backtesting historical signals

When NOT to use this skill

• Single-stock decisions → use analyze-stock instead
• Pre-earnings positioning → use earnings-prep
• Portfolio audit → use portfolio-audit
• New idea hunting → use find-untapped-thesis / find-alpha

This skill answers ONE question: "Should I be aggressive, neutral, or defensive RIGHT NOW?"

Companion skills

• sector-rotation-analysis — full 11-GICS heat map (this skill only does Layer 8 quick check)
• macro-risk-check — overlapping but more news-driven; this skill is more quantitative
• portfolio-audit — once regime is determined, audit current book against it

Hard-learned context (verified examples)

• 2000 March: NDX PE 73, F&G 95, VIX 18 → 6 months later NDX -78%
• 2020 Feb: NDX PE 32, F&G 60, VIX 13 → 5 weeks later -35%
• 2021 Nov: NDX PE 35, F&G 80, VIX 17 → 12 months later -33%
• 2022 Jan: VIX <17 in low-rate environment + Fed hawkish pivot → -27%
• 2026 May (now): NDX PE 38.15 (essentially at all-time non-bubble high), VIX low, AI capex narrative max-extended

The signal is loudest when valuation is extreme AND sentiment is greed AND VIX is low AND breadth is narrowing. Any 3 of 4 = serious YELLOW. All 4 = RED.

Year-specific overlays (update annually)

For 2026:

• AI capex names (NVDA / AVGO / AMD / MRVL) are sentiment-coupled to AI Power names (CEG / VST / GEV) and AI receiver names (FN / TSEM / GFS / NOK / GLW)
• A correction in any of the three sub-themes will likely spread to all three within 5-10 trading days
• Yen carry (USD/JPY) is the single highest-correlation external trigger for AI / semi names
• Watch for: NVDA earnings (5/20 next), CPI prints, Fed dots, BOJ rate path, China CXMT IPO (June 2026 — could pressure MU/SK Hynix valuations)