An AI framework that mirrors a real trading firm.
Concordal gives each AI agent a specialized role, drives structured debate among them, and mirrors the workflow of an actual buy-side desk. We turn a chaotic single-prompt LLM into an organized, collaborative decision process.
Concordal isn't just a trading framework — it's the future of collaborative AI in finance.
Realistic Firm Simulation
Successfully replicates the collaborative intelligence of a professional trading desk.
Multi-faceted Dialectical Reasoning
Bull / bear debate produces more robust, comprehensive decisions than a single prompt.
Efficient Structured Communication
Solves the information-distortion problem that plagues classic multi-agent systems.
Methodology adapted from github.com/TauricResearch/TradingAgents
Seven specialized roles working in a closed-loop decision process
Inspired by the org chart of an actual hedge fund. Each agent has a specific objective, skillset, and toolset.
Fundamentals · Sentiment · News · Technical
Bull vs Bear researchers
Synthesizes inputs into a sized, conditioned trade
Aggressive · Neutral · Conservative debate
Authorizes the transaction with explicit rationale
Stored, ranked, audit-able. We never auto-execute trades.
Four lenses on one ticker
Fundamental Analyst
Reads financial statements, earnings reports, and insider transactions to estimate intrinsic value.
Sentiment Analyst
Processes social-media posts and sentiment scores to gauge market mood and predict investor behavior.
News Analyst
Analyzes news, government announcements, and macro indicators to assess events that may move the market.
Technical Analyst
Computes technical indicators (MACD, RSI), reads price patterns and volume to forecast direction.
Balancing opportunity and discipline
Trader: synthesize, propose
The trader integrates every analyst report and the bull/bear debate into a concrete trade proposal — with timing and sizing.
Risk Team: evaluate, adjust
The risk team holds its own internal debate (aggressive / neutral / conservative) and adjusts the trade to fit pre-set risk parameters.
反思记忆 — 系统会从自己的错误中学习
每晚 06:30 UTC 系统跑一个后台 cron 任务：扫数据库里**前一天跨过 5 日前瞻窗口**的所有决策， 对每个判断计算正确性（BUY 在收益 > +1% 时正确；SELL 在 < -1% 时正确；HOLD 在 |return| < 1% 时正确）。 每个判断生成一条结构化反思条目。
下次有人对同一只票做新决策时，经理（Manager）prompt 会以这只票最近至多 10 条反思记忆为条件 — 看见「上次对 AAPL 标了 80% 置信度 BUY，5 日后实际下跌 3%」这种自我修正信号。
在我们的实证中（论文 §9.4），反思记忆机制相对静态 prompt 基线再降低 1.8 个百分点 ECE。 这是 Reflexion [Shinn 2023, MIT] 模式在金融决策的具体实现。
What the agents actually read
Honest list. We use free public sources today; institutional data feeds (Bloomberg, FactSet, Wind) are on the roadmap.
Cost-aware model routing
Each agent runs at the cheapest tier that still produces good output. Trader and Manager get Pro; analysts/risk run on Flash.
Gemini 3.1 Pro PreviewGemini 2.5 FlashEnd-to-end traceable. Every analyst report, debate turn, and risk vote is stored in the JSON trace.