Finter Insight Skill

Generate novel research hypotheses by analyzing past research and avoiding duplicates.

When to Use

Use this skill when generating new research topics for the Fund Manager.

Workflow

code

1. Think of topic idea
2. Run search_research.py to check for similar past research
3. Based on results:
   - High similarity (>0.7): IMPROVE existing or pick different topic
   - Medium similarity (0.4-0.7): Review context, differentiate approach
   - Low similarity (<0.4): Proceed as novel topic
4. Run finalize_insights.py to validate and save insights.json

CRITICAL: Use finalize_insights.py

ALWAYS use finalize_insights.py to save insights.json. Do NOT use Jupyter or manual Write - use the script to ensure validation.

Scripts

finalize_insights.py - Save and Validate insights.json (REQUIRED)

ALWAYS run this to save insights.json. This ensures proper validation.

bash

# From draft file
python .claude/skills/finter-insight/scripts/finalize_insights.py --file draft_insights.json

# With inline JSON (all go to completely_new)
python .claude/skills/finter-insight/scripts/finalize_insights.py --json '[
  {"topic": "Multi-Coin Momentum", "universe": "crypto_test", "hypothesis": "...",
   "category": "momentum", "approach": "...", "novelty_score": 8,
   "related_research": {"checked": true, "similar_count": 0, "max_similarity": 0.0}}
]'

# With separate categories
python .claude/skills/finter-insight/scripts/finalize_insights.py \
  --improve '[...]' \
  --resurrect '[...]' \
  --new '[...]'

search_research.py - Check for Similar Research

ALWAYS run this before finalizing a topic idea.

Run locally using Bash tool (NOT Jupyter):

bash

# Basic search
python .claude/skills/finter-insight/scripts/search_research.py "momentum strategy on kr_stock"

# Filter by universe
python .claude/skills/finter-insight/scripts/search_research.py "value investing" --universe us_stock

# Get more results
python .claude/skills/finter-insight/scripts/search_research.py "volatility" --top 10

Output interpretation:

•similarity > 0.7: Very similar - consider improving instead of new
•similarity 0.4-0.7: Related - review for context and differentiate
•similarity < 0.4: Novel - but still check the results for useful insights

Output Format

When generating insights.json, include related_research for each hypothesis:

json

{
  "completely_new": [
    {
      "topic": "Acoustic Damping in Price Discovery",
      "universe": "kr_stock",
      "hypothesis": "...",
      "related_research": {
        "checked": true,
        "similar_count": 2,
        "max_similarity": 0.45,
        "differentiation": "Prior work focused on volatility magnitude, this focuses on propagation speed"
      }
    }
  ]
}

Key Rules

•ALWAYS search before generating - No exceptions
•High similarity = pivot or improve - Don't suggest duplicate research
•Include related_research field - Show you checked
•Learn from failures - If similar research failed, explain why yours is different