Old Earnings Orchestrator

Earnings Orchestrator

Goal: Process all 8-K earnings filings for a company chronologically, building a closed-loop prediction→attribution system with accuracy feedback.

Thinking: ALWAYS use ultrathink for maximum reasoning depth.

Input: Company ticker (e.g., "GBX", "AAPL")

Architecture

Master Orchestrator (this skill)
    │
    ├─→ Query Neo4j for all 8-Ks with Item 2.02
    │
    └─→ For each filing chronologically:
            │
            ├─→ First filing: /earnings-attribution only
            │
            └─→ Subsequent filings:
                    │
                    ├─→ /earnings-prediction (predict before seeing outcome)
                    │
                    └─→ /earnings-attribution (verify & learn)

Workflow

Use TodoWrite to track progress through all filings.

Step 1: Query 8-K Universe

Query Neo4j for all 8-Ks with Item 2.02 (earnings) for the ticker:

MATCH (r:Report)-[pf:PRIMARY_FILER]->(c:Company)
WHERE c.ticker = $ticker
  AND r.formType = '8-K'
  AND any(item IN r.items WHERE item CONTAINS 'Item 2.02')
  AND pf.daily_stock IS NOT NULL
RETURN r.accessionNo AS accession_no,
       c.ticker AS ticker,
       c.name AS company_name,
       r.created AS filing_datetime,
       pf.daily_stock AS daily_return,
       pf.daily_macro AS macro_adj_return,
       r.items AS items
ORDER BY r.created ASC

Extract list of accession numbers with filing dates.

Step 2: Check Processing State

Read earnings-analysis/8k_fact_universe.csv and earnings-analysis/predictions.csv to determine:

• Which filings are already completed
• Which predictions exist (and whether they need attribution)

Step 3: Process Each Filing

For each filing in chronological order:

First Filing (no prior history)

/earnings-attribution {accession_no}

Reason: No historical baseline to predict from. Attribution only.

Subsequent Filings

1. /earnings-prediction {accession_no}
   → Outputs prediction to predictions.csv

2. /earnings-attribution {accession_no}
   → Verifies prediction accuracy
   → Updates predictions.csv with actual_* columns
   → Stores company-specific learnings

Step 4: Update Tracking

After each filing:

1. Mark completed=TRUE in 8k_fact_universe.csv
2. Update predictions.csv with actual outcomes
3. Compute running accuracy metrics

Accuracy Tracking

Direction Accuracy

correct = predicted_direction == actual_direction

Magnitude Accuracy

magnitude_correct = predicted_magnitude == actual_magnitude

Running Metrics

After processing, compute:

• Total filings processed
• Predictions made (excludes first filing)
• Direction accuracy: correct / predictions
• Magnitude accuracy: magnitude_correct / predictions

Resume Logic

If a filing is partially processed:

1. Check predictions.csv for existing prediction
2. Check Companies/{TICKER}/{accession}.md for attribution
3. Skip completed steps, resume where needed

Output Files

Run Summary Format

After completing a ticker, write summary to earnings-analysis/orchestrator-runs/{ticker}_{YYYYMMDD}.md:

# {TICKER} Orchestrator Run - {DATE}

## Summary
- Total 8-Ks found: N
- Already completed: N
- Processed this run: N
- First filing (attribution only): {accession}

## Prediction Accuracy
- Predictions made: N
- Direction correct: N/M (X%)
- Magnitude correct: N/M (X%)

## Filings Processed

| # | Accession | Date | Prediction | Actual | Correct |
|---|-----------|------|------------|--------|---------|
| 1 | xxx | 2023-01-01 | (first) | +5.2% | N/A |
| 2 | xxx | 2023-04-01 | up/medium | up/large | dir: Y, mag: N |
...

## Learnings Applied
- [List company-specific patterns discovered]

Invocation Examples

Process all filings for a ticker

/earnings-orchestrator GBX

Process with limit

/earnings-orchestrator GBX --limit 5

Processes only first 5 unprocessed filings.

Resume interrupted run

/earnings-orchestrator GBX --resume

Continues from last processed filing.

Error Handling

1. Neo4j query fails: Log error, exit gracefully
2. Skill invocation fails: Log error, mark filing as failed, continue to next
3. Missing data: Note in summary, continue with available data
4. Rate limits: Built-in retry with exponential backoff

Arguments

Version 1.0 | 2026-01-16