News Impact

News Impact

Identify what drives stock price moves with maximum accuracy, comprehensiveness, and confidence.

Arguments

Received: $ARGUMENTS

Parse as: TICKER START_DATE END_DATE [THRESHOLD] [--no-perplexity]

• THRESHOLD: 3s (default), 1.5s, 2s, or fixed percent
• --no-perplexity: Skip Perplexity for gap days (faster)

If no arguments received, ask user for TICKER, START_DATE, END_DATE.

Flow

Step 0: Validate Data Availability (REQUIRED FIRST)

CRITICAL: Run this query BEFORE any other step. STOP if validation fails.

MATCH (c:Company {ticker: $ticker})
OPTIONAL MATCH (d:Date)-[r:HAS_PRICE]->(c)
WITH c, max(d.date) AS latest_date, min(d.date) AS earliest_date, count(r) AS price_count
RETURN c.name AS company_name,
       latest_date,
       earliest_date,
       price_count,
       CASE WHEN latest_date >= date($start) THEN true ELSE false END AS has_start_data,
       CASE WHEN latest_date >= date($end) THEN true ELSE false END AS has_end_data

Validation Rules:

1. If company_name is null → STOP: ERROR: Ticker {ticker} not found in database
2. If price_count = 0 → STOP: ERROR: No price data for {ticker}
3. If has_start_data = false → STOP: ERROR: No price data for {ticker} in requested range. Latest available: {latest_date}
4. If has_end_data = false → WARN but continue: WARNING: Data only available through {latest_date}, analysis will end there

DO NOT proceed to Step 1 if validation fails. DO NOT fall back to web search for missing price data.

Step 1: Get Benzinga News

Call /get-bz-news $ARGUMENTS

Returns news where |daily_adj| >= threshold (default 1.5σ) with:

• volatility: trailing adjusted volatility used
• z_score: how many sigmas this move was

If INSUFFICIENT_HISTORY returned, fall back to fixed 3% threshold.

Step 2: Analyze Each News Item

For each news item from Step 1:

1. Read title AND body - titles can be vague

2. Check market_session for timing context:

   • pre_market: News likely CAUSED the day's move → HIGH confidence
   • in_market: News aligns with intraday action → MEDIUM-HIGH confidence
   • post_market: News EXPLAINS today's move, but impacts NEXT trading day → MEDIUM confidence (reactive)

   Note: Post-market news (earnings, guidance) will move the stock at next market open. Associate post_market news with the NEXT day's return.

3. Generate driver phrase (5-15 words explaining why stock moved)

4. Assess confidence (0-100%) based on:

   • How clearly news explains the move direction
   • Timing alignment (pre-market earnings → high; post-market commentary → lower)
   • Z-score magnitude (higher z-score + clear news = higher confidence)

Step 3: Find Gap Days

Query daily returns to find big moves without news coverage.

First, get the volatility (or use the one from Step 1):

MATCH (d:Date)-[r:HAS_PRICE]->(c:Company {ticker: $ticker})
WHERE d.date >= date($start) - duration('P365D') AND d.date < date($start)
MATCH (d)-[m:HAS_PRICE]->(idx:MarketIndex {ticker: 'SPY'})
WHERE r.daily_return IS NOT NULL AND m.daily_return IS NOT NULL
WITH stdev(r.daily_return - m.daily_return) AS adj_vol
RETURN adj_vol

Then find all significant move days:

MATCH (d:Date)-[r:HAS_PRICE]->(c:Company {ticker: $ticker})
WHERE d.date >= date($start) - duration('P365D') AND d.date < date($start)
MATCH (d)-[m:HAS_PRICE]->(idx:MarketIndex {ticker: 'SPY'})
WHERE r.daily_return IS NOT NULL AND m.daily_return IS NOT NULL
WITH c, stdev(r.daily_return - m.daily_return) AS adj_vol

MATCH (d2:Date)-[r2:HAS_PRICE]->(c)
WHERE d2.date >= $start AND d2.date < $end
MATCH (d2)-[m2:HAS_PRICE]->(idx:MarketIndex {ticker: 'SPY'})
WITH d2.date AS date,
     r2.daily_return AS stock,
     m2.daily_return AS macro,
     (r2.daily_return - m2.daily_return) AS daily_adj,
     adj_vol
WHERE abs(daily_adj) >= $multiplier * adj_vol
RETURN date, stock, macro, daily_adj, adj_vol,
       abs(daily_adj) / adj_vol AS z_score
ORDER BY date

Compare with news dates from Step 1. Gap = date with big move but no news.

Step 4: Research Gaps (WebSearch → Perplexity)

Skip this step if --no-perplexity flag is set. Just list gaps as UNKNOWN with confidence=0.

Otherwise, for each gap day:

4a. WebSearch first (faster, multi-source validation):

1. Search: "{ticker} stock news {date}" or "{company name} {date} move"
2. Look for 2+ independent sources corroborating the same explanation
3. If 2+ sources agree → accept with confidence based on source quality:
   • 3+ sources: 70-85% confidence
   • 2 sources: 50-70% confidence
4. If <2 sources found → escalate to Perplexity

4b. Perplexity fallback (if WebSearch insufficient):

1. perplexity_search - "{ticker} stock news {date}"
2. perplexity_research - only for major moves (>5%) with no results

Generate driver and confidence from research. Include z-score context.

Step 5: Merge and Return

Combine news (Step 2) + gaps (Step 4), sort by date ASC.

Output Format

Pipe-delimited, one per line:

date|news_id|driver|confidence|daily_stock|daily_adj|sector_adj|industry_adj|z_score|volatility|market_session|source

Move Type Interpretation:

• If |daily_adj| >> |sector_adj| → Sector-driven (not company-specific)
• If |sector_adj| ≈ |industry_adj| ≈ |daily_adj| → Idiosyncratic (company-specific)
• If all small but daily_stock large → Market beta

Rules

• All adjusted returns and volatility calculated in Cypher, NOT by LLM
• WebSearch first for gap days — validate with 2+ sources before accepting
• Perplexity fallback only if WebSearch finds <2 sources
• Always read news body, not just title, for accurate driver extraction
• Use market_session timing to assess causation vs explanation
• Post-market news impacts NEXT trading day, not current day
• If no explanation found, driver = "UNKNOWN", confidence = 0
• Default threshold: 1.5σ of trailing adjusted returns
• Minimum 60 days history required; fall back to 3% if insufficient

CRITICAL - Date Range Enforcement:

• ONLY return results for dates within the requested START_DATE to END_DATE range
• If no significant moves found in the range, return: NO_SIGNIFICANT_MOVES: No moves exceeding {threshold} found for {ticker} between {start} and {end}
• NEVER substitute other dates - do not search for or return data from outside the requested range
• NEVER compensate for empty results by finding "interesting" moves from other periods