COT Contrarian Detector

Overview

Implements step 1 of Jason Shapiro's COT (Commitment of Traders) contrarian process: detect when large speculators are crowded into one side of a futures market. Crowded positioning is a precondition for a contrarian trade, not a trade signal — a market only becomes tradable once crowding is confirmed by a news failure and price-action reversal (steps 2-3), which this skill guides the user through manually.

Core thesis (Shapiro): Large speculators (hedge funds, CTAs, momentum traders) tend to be maximally positioned at trend exhaustion, not trend inception. When they are already crowded onto one side, the next big move is statistically more likely to run them over than to reward them further. Fade the speculators, not the commercials (commercials hedge for structural reasons and are not a crowd-psychology signal).

When to Use This Skill

English:

• "What markets are the speculators crowded into right now?"
• "Run a COT report analysis" / "Show me COT positioning extremes"
• "Is anyone 'trapped' in gold / the dollar / bonds right now?"
• User wants to find contrarian futures setups
• User asks for a Jason Shapiro-style COT screen

Japanese:

• 「COTレポートで買われすぎ・売られすぎのポジションを調べて」
• 「投機筋が偏っている市場は？」
• 「ジェイソン・シャピロ式の逆張り分析をして」

Do NOT use when:

• The user wants a trade signal right now — crowding alone is not actionable; see Guardrails below
• The user is asking about individual equities — COT reports cover CFTC futures markets only (indices, rates, FX, metals, energy, agri, crypto), not single stocks

Prerequisites

• FMP API Key: Required. Set FMP_API_KEY environment variable or pass --api-key. COT endpoints require an FMP Premium+ plan — a free-tier key will not have access.
• Python 3.9+ with requests installed.
• API Budget: One call per market (23 for --core, up to ~65 for the full universe), plus one call for the market list when neither --symbols nor --core is given.

Workflow

Phase 1: Run the crowding screen

# Curated core futures universe (23 liquid/representative markets)
python3 skills/cot-contrarian-detector/scripts/screen_cot_crowding.py --core --output-dir reports/

# Explicit symbols
python3 skills/cot-contrarian-detector/scripts/screen_cot_crowding.py --symbols "ES,GC,CL" --output-dir reports/

# Full universe (all ~65 markets FMP's COT list covers)
python3 skills/cot-contrarian-detector/scripts/screen_cot_crowding.py --output-dir reports/

The script fetches each market's weekly legacy COT report (large-speculator long/short positions), computes a 156-week (3-year) and 26-week COT Index per market, and classifies extremes:

• CROWDED_LONG — COT Index >= 90 (near the 3-year net-long high)
• CROWDED_SHORT — COT Index <= 10 (near the 3-year net-short high)
• NEUTRAL — everything in between

Markets with insufficient history to compute the index are never silently dropped — they appear in a skipped list with the reason (e.g. "insufficient history: 40/156 weeks").

Phase 2: Present the crowding report

Present the generated Markdown report, highlighting:

• Which markets are CROWDED_LONG / CROWDED_SHORT and by how much
• The 26-week index for context (is the crowding fresh or aging?)
• Week-over-week net-position swings (fast-moving crowds are more fragile)
• The methodology note and disclaimer — crowding is not a trade signal

Phase 3: Guide steps 2-5 manually (Shapiro process)

For any CROWDED_LONG / CROWDED_SHORT market the user wants to pursue, load references/shapiro-methodology.md and walk through the remaining steps — these are not automated:

1. Crowding detection (done — this skill)
2. News failure — use WebSearch to check whether recent news favorable to the crowd's direction failed to move price the way the crowd would expect (e.g. crowded-long market doesn't rally on bullish news). This is the core edge and the most important manual confirmation.
3. Price-action confirmation — check the weekly chart for a reversal pattern or a failure at a new high/low.
4. Entry — against the crowd, with a stop at the recent swing extreme and small, fixed-risk sizing (see position-sizer skill).
5. Exit — when positioning normalizes toward neutral (COT Index back toward 50) or the stop is hit.

Never recommend an entry from crowding alone — steps 2 and 3 must both confirm first.

Output

• JSON: reports/cot_crowding_<as-of-date>.json — machine-readable, with a run_context block (schema_version, params, universe, data_date) plus markets (ranked results) and skipped (never silently dropped).
• Markdown: reports/cot_crowding_<as-of-date>.md — human-readable report with Crowded Long / Crowded Short / Full Ranking / Week-over-Week Swings / Skipped Markets / Methodology sections.

Cadence

CFTC publishes the COT report Fridays ~3:30pm ET, with positions as of the prior Tuesday — data is always 3+ days old by the time it's published, and up to 9 days old by the following Friday. Run this skill:

• Weekly, after Friday's publication or over the weekend, for a fresh read
• Ad hoc, when the user asks about a specific market's positioning — the underlying data will be from the most recent Friday release either way

Guardrails

• Crowdedness alone is NOT a trade signal. It is a precondition. Never suggest an entry without steps 2 (news failure) and 3 (price action) from references/shapiro-methodology.md also confirming.
• Data is lagged. COT positions are 3-9 days old by the time they're read; do not treat them as a real-time signal.
• Fade speculators, not commercials. This skill only looks at non-commercial ("large speculator") positioning — commercial hedging flows are structurally different and not a crowd-psychology signal.
• Not investment advice. All output is for research/educational purposes.

Resources

references/shapiro-methodology.md

The full 5-step process (crowding → news failure → price action → entry → exit), why speculators (not commercials) are the fade target, the 3-day publication lag caveat, and a table of what this skill automates vs. what stays manual. Load this whenever guiding a user past step 1.

references/cot-index-calculation.md

The COT Index formula, lookback rationale (156w primary / 26w context), extreme threshold sensitivity, open-interest normalization rationale, the legacy-vs-disaggregated report distinction (this skill uses the legacy report's non-commercial = large-speculator fields), and a glossary of the FMP COT API field names consumed by scripts/cot_index.py.

Vendored client: scripts/fmp_client.py (generated) owns FMP transport — rate limiting, single 429 retry, header auth (the key never appears in a URL), and apikey redaction on (data, error) results (Issue #495).

When to Load References

• First use / explaining the methodology: Load references/shapiro-methodology.md
• Explaining a specific number in the report: Load references/cot-index-calculation.md
• Regular execution: References not needed — the script handles the crowding computation