Earnings Attribution

Earnings Attribution Analysis

Goal: Determine WHY a stock moved after an 8-K filing by calculating the SURPRISE (actual vs expected).

Thinking: ALWAYS use ultrathink for maximum reasoning depth.

End Goal: Build company-specific driver understanding for real-time prediction accuracy.

Core Principles

1. Comprehensiveness Over Speed: Query ALL relevant data sources
2. Evidence-Based Claims Only: Every claim must cite a source
3. Surprise-Focused: Stock moves on actual vs expectations, not absolute results
4. Learn from History: Each analysis builds company-specific knowledge
5. Self-Audit: Validate all claims have sources before completing

Confidence Levels

• High: Multiple sources agree, clear fundamental reason, unambiguous evidence
• Medium: Sources partially agree, conclusion fits with caveats
• Insufficient: Sources conflict, missing critical data, no clear driver

If evidence is insufficient, say so. Never fabricate certainty.

Resources

• Output format: output_template.md
• Evidence audit checklist: evidence_audit.md
• Usage examples: examples.md
• Self-improvement: update-skills.md
• Known data gaps: data_gaps.md

Neo4j Subagents

CRITICAL: Do NOT write Cypher queries. Describe what data you need in natural language. Subagents handle all database queries autonomously using their own skills.

Use Task tool to spawn these specialized subagents:

Iterative Refinement: Subagents return agent_id with results. For follow-up queries on the same company, use resume: <agent_id> instead of spawning new—agent preserves context (ticker, dates, previous queries). Spawn new agent only for different company or unrelated domain.

Resume prompt examples: "Only 2 articles, expand to ±5 days", "Also get EX-99.1 exhibit text", "What did analysts ask in Q&A?"

Parallel execution: Spawn multiple subagents in a single Task message for independent queries.

PIT filtering (prediction mode): Prefix prompt with [PIT: {filing_datetime}] to restrict results to before that time.

Skill Updates: If a subagent response contains SKILL_UPDATE_NEEDED block, execute the update immediately using Edit tool—no permission needed. This includes schema corrections, property name fixes, and data gaps. Do not skip or defer these updates.

Response Validation: When receiving subagent data, verify each value has citation format: (Source: {NodeType}:{identifier}, {date}). If citation missing:

1. Use resume: <agent_id> to ask subagent to re-provide with source
2. Do NOT add unsourced data to Evidence Ledger
3. Note "Data unavailable" if subagent cannot cite source

Workflow (11 Steps)

Use TodoWrite to track progress. Mark each step in_progress before starting, completed immediately after.

Rules: Step 1 always first. Steps 2-5 case-by-case sequencing. Steps 6-11 sequential. Step 11 is MANDATORY.

Step 1: Data Inventory

Query what data exists before making claims. Use neo4j-report subagent.

Check for: News count, Transcript exists, XBRL reports count, Dividends, Splits.

Step 2: Get Report and Returns

Get filing details and price reaction. Use neo4j-report subagent.

Returns to capture:

• Macro-adjusted (vs SPY) - primary
• Sector-adjusted - context (may be null)
• Industry-adjusted - context (may be null)

Step 3: Get Consensus Estimates

Always query Perplexity - do not rely on Neo4j News for consensus.

mcp__perplexity__perplexity_search:
  query: "{ticker} Q{quarter} FY{year} EPS revenue estimate consensus before {date}"

Source preference for Evidence Ledger:

1. Best: Perplexity pre-filing search (explicit consensus before event)
2. Acceptable: News headline citing estimate (note: post-filing, weaker provenance)
3. Unacceptable: Unsourced or inferred values

If using News as consensus source, cite it accurately in Evidence Ledger—do not claim Perplexity in Data Sources Used.

Step 4: Query Neo4j Sources

Query based on Data Inventory results. Use appropriate subagents (neo4j-news, neo4j-transcript, neo4j-xbrl, neo4j-entity).

4A: News

• Start ±2 trading days; expand to ±5 if <3 items
• Filter: WHERE r.daily_stock IS NOT NULL (data quality)
• Extract: EPS vs estimate, guidance, analyst reactions

4B: Transcript (Item 2.02 only)

• Get Prepared Remarks + Q&A exchanges
• Look for: analyst concerns, management hedging, specific numbers

4C: Historical XBRL

• Last 4 quarters of 10-K/10-Q data
• Look for: EPS trend, revenue trend, margin changes

4D: Dividends

• Cuts = bearish, Raises = bullish, Special = one-time

4E: Splits

• Rarely explain fundamental reactions

4F: Verify Sub-Agent Results

Check Coverage for adequate search. If Coverage missing or unclear, ask sub-agent to clarify. Re-query if gaps seem wrong (max 2 follow-ups). Check cross-domain consistency before synthesis.

Step 5: Query Perplexity

Use when Neo4j doesn't provide complete picture.

Query Principles:

• Be specific about time (exact date, not "recently")
• Include company name AND ticker
• For consensus: specify fiscal quarter AND year
• For "why stock moved": include date AND direction

Step 6: Synthesize

Calculate SURPRISE for each metric using these exact formulas:

Surprise Calculation Formulas

EPS Surprise % = ((Actual - Consensus) / |Consensus|) × 100
Revenue Surprise % = ((Actual - Consensus) / |Consensus|) × 100

Guidance Surprise:

1. Calculate midpoint of guidance range: (Low + High) / 2
2. Compare to consensus: ((Midpoint - Consensus) / |Consensus|) × 100
3. If range width changed significantly, note this separately

Rounding: All percentages to 2 decimal places.

Example Calculation

Actual EPS: $2.03
Consensus EPS: $1.98
Surprise = (($2.03 - $1.98) / |$1.98|) × 100 = +2.53%

Guidance Range: $12.00 - $13.50
Midpoint: $12.75
Consensus: $13.22
Surprise = (($12.75 - $13.22) / |$13.22|) × 100 = -3.56%

Surprise Analysis Table

Rank surprises by absolute magnitude. Largest surprise typically explains the move.

Attribution Structure

Primary Driver (High Confidence)

• What: [Descriptive name]
• Surprise: Expected $X, Got $Y (Z% surprise)
• Evidence: [Source citation]

Contributing Factor(s) (if applicable)

• What: [Name]
• Evidence: [Source]

Evidence Extraction

Capture evidence immediately when reading each source. Populate the Evidence Ledger in output_template.md as you go; every numeric claim must appear there.

Strict Rules for Numbers (Used in Surprise Calculations)

Any number used to calculate surprise MUST have:

1. Exact value (not rounded or paraphrased)
2. Explicit source with timestamp/date
3. Format: {metric}: ${exact_value} (Source: {source_name}, {date})

Examples:

Consensus EPS: $1.98 (Source: Perplexity, pre-filing)
Actual EPS: $2.03 (Source: 8-K EX-99.1, 2023-11-02)
Guidance midpoint: $12.75 (Source: 8-K EX-99.1, calculated from $12.00-$13.50)

If you cannot cite exact value + source, do NOT use the number in surprise calculations.

Flexible Rules for Qualitative Evidence

For non-numeric observations, paraphrasing is acceptable with source:

Management tone: Cautious on near-term demand (Source: Transcript Q&A #3)
Analyst concern: Margin pressure (Source: News "ROK Beats but Guides Lower")

Sources: News: "{headline}", Transcript Q&A #N, Transcript Prepared Remarks, Exhibit EX-99.1, Perplexity search, XBRL: {report_id}

Output

See output_template.md for full report format.

Save to: earnings-analysis/Companies/{TICKER}/{accession_no}.md

Required sections:

1. Report Metadata
2. Returns Summary
3. Evidence Ledger
4. Executive Summary
5. Surprise Analysis
6. Attribution (Primary + Contributing)
7. Data Sources Used
8. Confidence Assessment
9. Historical Context

Core Rules

1. Data Inventory First: Know what data exists before making claims
2. Surprise-Based: Stock moves on actual vs expected
3. Evidence Required: Every claim needs a cited source
4. Perplexity for Consensus: Always query for market expectations
5. Self-Audit: Validate all claims have sources
6. Learn and Record: Store company-specific learnings
7. Honest Confidence: State "Insufficient" when evidence doesn't support
8. No XBRL for 8-K: Use historical 10-K/10-Q XBRL for trends

Perplexity Tools

Note: For SEC filings search, use utils/perplexity_search.py:perplexity_sec_search().

Conflict Resolution Guidelines

When sources conflict, use these principles (guidelines, not rigid rules):

Source Reliability Hierarchy

Driver Priority (Typical, Not Absolute)

• Forward-looking (guidance) usually > Backward-looking (EPS)
• BUT: If EPS miss is >10%, it can dominate
• AND: Validate against actual price reaction direction

Conflict Handling

If News says "beat" but Transcript reveals "guidance cut":
→ Trust Transcript (higher reliability)
→ Note the conflict in analysis

If two News sources give different consensus numbers:
→ Use the one with explicit source attribution
→ If neither has clear attribution, note both values and the discrepancy
→ Do NOT average (averages have no direct source)

Rule: When in doubt, note the conflict explicitly rather than silently choosing one.

Data Quality Guardrails

News Anomaly Filter (REQUIRED)

1,746 News→Company relationships (0.9%) have daily_industry but no daily_stock. Always filter:

WHERE r.daily_stock IS NOT NULL

Missing Returns Handling

If a report has no return data (pf.daily_stock IS NULL):

• Flag as incomplete data
• Do NOT use for driver sensitivity calculations
• Can still analyze qualitatively

Duplicate News Detection

If multiple news items have nearly identical titles (>80% similarity):

• Use the earliest timestamp
• Treat as single data point, not confirmation

Stale Consensus Warning

When Perplexity returns consensus estimates:

• Verify the date/quarter matches your target
• "Q3 2023 consensus" is wrong if you need Q4 2023
• Re-query with more specific date if unclear

Company-Specific Learning

After each analysis, update: earnings-analysis/Companies/{TICKER}/learnings.md

See output_template.md for learnings format.

Gating rule: Only update learnings.md after the Evidence Ledger is complete and evidence_audit.md passes.

Session & Subagent History (Shared CSV)

History file: .claude/shared/earnings/subagent-history.csv (shared with earnings-prediction)

Format: See subagent-history.md for full documentation.

accession_no,skill,created_at,primary_session_id,agent_type,agent_id,resumed_from
0001514416-24-000020,attribution,2026-01-13T10:30:00,0415feb7,primary,,
0001514416-24-000020,attribution,2026-01-13T10:31:05,0415feb7,neo4j-entity,abc12345,

On analysis start:

1. Read CSV (create with header if doesn't exist)
2. Append primary row: {accession},attribution,{timestamp},{session_id},primary,,

Before calling a subagent:

1. Query latest agent ID: grep "{accession}" | grep ",{agent_type}," | tail -1 | cut -d',' -f6
2. If agent ID exists → can use resume: <id> in Task call
3. If want fresh session → proceed without resume

After each subagent completes:

1. Extract agentId from Task response
2. Append row: {accession},attribution,{timestamp},{session_id},{agent_type},{agent_id},{resumed_from}

Step 10: Mark Completed

After successful analysis (report saved, audit passed, learnings updated):

1. Update tracking CSV: Set completed=TRUE for this accession_no in earnings-analysis/8k_fact_universe.csv
2. Update predictions CSV: If a prediction exists for this accession_no in earnings-analysis/predictions.csv, fill in:
   • actual_direction: up/down based on daily_stock return
   • actual_magnitude: small/medium/large based on |return|
   • actual_return: the actual daily_stock percentage
   • correct: TRUE if direction matches, FALSE otherwise
3. Verify: Confirm rows are updated correctly

Step 11: Build Thinking Index (MANDATORY - DO NOT SKIP)

This step is REQUIRED. Execute this exact command:

python3 scripts/build-thinking-index.py {accession_no}

Replace {accession_no} with the actual accession number from your analysis (e.g., 0001234567-24-000001).

This extracts thinking from all sessions and sub-agents and saves to Obsidian.

Version 2.2 | 2026-01-16 | Added Step 11: mandatory thinking index build