news-reaction-failure-analyzer

---
name: news-reaction-failure-analyzer
description: Use when judging whether a market failed to react to news favorable to a crowded position — step 2 of Shapiro's COT contrarian process — for a CONFIRMED/NOT_CONFIRMED/INSUFFICIENT_EVIDENCE verdict.
category: Performance
version: 1.0.0
tools: []
---

# News Reaction Failure Analyzer

Judge whether a market failed to react to news that should have rewarded a crowd — step 2 of Jason Shapiro's COT contrarian process (step 1 is the `cot-contrarian-detector` skill, which flags a market CROWDED_LONG or CROWDED_SHORT). A crowded-long market that doesn't rally on genuinely bullish news, or a crowded-short market that doesn't sell off on genuinely bearish news, is the core behavioral tell that the crowd has run out of buying or selling power. Confirming this only completes step 2 of 5 — price-action confirmation, entry, and exit (steps 3-5) remain manual and are still required before any position. The method generalizes beyond COT to post-earnings-announcement-drift (PEAD) and macro-crowding news-failure checks.

## Why Not a Naive Failure Ratio

An earlier design flagged "news failure" whenever fewer than half the relevant events "responded." Under pure noise, roughly 69% of individual events fail to respond by chance, so that naive rule would confirm on random noise 48-83% of the time depending on sample size. Instead, require the market to have moved *significantly against* the crowd's favorable news — a drift-significance test with a Monte-Carlo-verified null false-positive bound — never merely "didn't respond enough."

## When to Use

Trigger on requests like "did the market shrug off [event] even though [asset] is crowded long/short," "run a news-failure check on [symbol]," or "is [symbol] confirmed for a Shapiro-style contrarian setup" — including Japanese-language equivalents (asking whether a market failed to react to good news, or asking to confirm a news failure for a COT-skewed ticker). Also trigger right after a `cot-contrarian-detector` result flags CROWDED_LONG/CROWDED_SHORT and the user wants to move to step 2.

Do not use when the market isn't crowded (a NEUTRAL classification) without an explicit user-supplied direction override, or when no curated events exist yet — curate events first (see below) and never fabricate events or source URLs to force a verdict.

## Workflow

**Obtain symbol + direction.** Take these from a `cot-contrarian-detector` report the user supplies, or ask directly for symbol and direction. Refuse fail-closed (state the reason, do not guess) when: the classification is NEUTRAL and no override is given; the symbol is missing from the report; or the report is older than a reasonable staleness window (10 days is the standard default) — an explicit user-supplied direction is the only override.

**Curate the events with web_search.** Search news within the evaluation window (10 days is a typical default) using a 4-tier source hierarchy: issuer/primary source → SEC/official statistics → wire services → news portals, preferring the highest tier available. For each relevant event, record the event description, its timestamp (with explicit UTC offset), the source URL, the source tier, and the expected impact (BULLISH or BEARISH) relative to the crowd's position. Never fabricate events or URLs — if web_search turns up nothing usable, say so explicitly and either stop or, if the user accepts it, proceed to an INSUFFICIENT_EVIDENCE verdict (reason: no events found).

**Compute the reaction.** For each qualifying event, determine the market's return over a short window (a 3-trading-day window is standard) following the event and express it as a z-score against the market's normal volatility. Cluster events whose windows overlap so they aren't double-counted as independent evidence (an independence guard). Aggregate across the clustered events into a mean/drift statistic and a responded-ratio, and only call CONFIRMED when the drift statistic clears a significance threshold in the direction *against* the crowd's favorable news — not merely a low responded-ratio. A stricter threshold (raising the effective z-cutoff to roughly 1.75) is available for users who want to hold a stricter <10% false-confirm margin even under adverse correlation assumptions, at the cost of missing some genuine signals.

**Present the verdict.** Report CONFIRMED / NOT_CONFIRMED / INSUFFICIENT_EVIDENCE with a confidence level, the aggregate stats (drift statistic, responded ratio), and a full evidence table (per-event return, z-score, and reaction label), including any dropped events and why — never silently hide them. Note explicitly if a proxy instrument's price series was used in place of the actual underlying (tracking error, expense drag, and roll-timing make the reaction read approximate, not exact). Frame CONFIRMED as evidence feeding whatever price-action-confirmation step comes next in the user's process — never as a trade signal in itself.

## Guardrails

- CONFIRMED is not a trade signal — it confirms step 2 of 5; price-action confirmation, entry, and exit are still manual.
- INSUFFICIENT_EVIDENCE (never a crash, never a forced call) applies when: fewer than 3 usable relevant event clusters exist; the source detector report is missing, unparsable, dated after the as-of date, or older than 10 days; the classification is NEUTRAL without an explicit override; or no working price source exists.
- COT data is typically 3-9 days old by the time it's read — treat news-failure evidence in that lag context, not as same-day confirmation.
- Counter-direction events (ones that don't match the crowd's expected direction) are shown in the evidence table for context but excluded from the verdict itself.
- The verdict's null false-CONFIRMED rate is hard-verified under i.i.d. noise (<8%) and under a realistic residual-correlation stress (AR(1) ρ=0.1, <10%). Under an intentionally extreme correlation stress (lag-1 ρ=0.3 across non-clustered event windows — roughly 10x liquid-futures empirical autocorrelation), the measured null rate rises to ~11-13%; this is a documented limitation, not a silent gap.
- Not investment advice — research and educational purposes only.