System Architecture
Agentic Trading System — 4-Layer Design
1 Perception — "What's happening now?"
Specialized agents parse heterogeneous data sources into structured analyses (Pydantic-typed outputs).
Price
OHLCV + charts → technical analysis
News
Articles + social → sentiment
Filings
SEC / earnings → financial factors
Macro
Economic data → macro analysis
structured analyses
2 Memory — "Has this happened before?"
Retrieval layer grounding current analysis in historical context via Qdrant vector DB and RAG.
Pattern Search
Similar price patterns
Event Impact
Event → price correlations
Company
Sector / competitor context
Self-Reflection
Past signal outcomes
In development Qdrant RAG
analysis + context
3 Cognition — "What should we believe?"
Thesis → Antithesis → Synthesis reasoning pattern forces structured deliberation and prevents premature convergence.
Analyst
Thesis — directional view + evidence
Critic
Antithesis — risks + counterarguments
Judge
Synthesis — conclusion + confidence
report + confidence
4 Action — "What do we do?"
Converts cognition report into trading signal (buy / sell / neutral) with confidence score and evidence-linked rationale.
Signal generation LangGraph orchestration Pydantic schemas
sungjun han · 2025