TradingAgents — Extended Fork

Multi-agent LLM trading framework. Extended fork of TauricResearch/TradingAgents, adding a web dashboard, a look-ahead-safe backtest harness, per-stage model routing, ETF support, and more.

The base framework runs a firm-like pipeline of LLM agents — fundamentals, sentiment, news, and technical analysts feed bull/bear researchers, a trader, a risk-debate team, and a portfolio manager — to produce a trading decision. This fork keeps all of that and builds new surfaces on top of it.

⚠️ Research use only. This is not financial, investment, or trading advice. See the upstream disclaimer.

---

What this fork adds

Everything below is new in this fork (i.e. on top of upstream). The broad multi-provider LLM support and Claude Sonnet 5 / Fable 5 are upstream features, not fork additions.

• Web dashboard (web/) — FastAPI + React/Vite/TypeScript/Tailwind. Launch analyses from the browser, watch live progress over WebSocket, view an agent timeline/Gantt, cancel and resume runs, browse run history and past reports, read investor summaries, with API-key preflight and sanitized errors.
• Backtest / evaluation harness (tradingagents/backtest.py) — look-ahead-safe, resumable JSONL output, null baselines (buy-and-hold, always-Buy, seeded random), a cost model, and a confidence-weighted alpha metric.
• Per-stage model routing (agent_llm_map) — assign quick vs deep models per agent stage — plus adaptive risk-debate rounds (adaptive_extra_rounds) that extend the debate only while the Research Manager and Trader disagree.
• Memory / reflection upgrades — cross-ticker lessons selected top-3 by |alpha|, with wider past-lessons injection into researchers, the trader, and the risk debators.
• ETF analysis support — dedicated ETF data tools and flows.
• Run-scoped OHLCV cache — price/volume fetched once per run and reused.
• Extra volume indicators — relative volume (rvol), volume z-score, and volume trend slope, with vendor fallback.
• Prediction-market tools — event-probability signals for the analysts.
• Headless runner + typed event protocol (tradingagents/runner.py, runner_events.py) — the shared substrate the CLI and dashboard both sit on.
• Investor briefings + structured position sizing — a dedicated briefing agent, plus stop-loss / position-size fields on the portfolio decision.
• GPT-5.6 models in the catalog — Sol (deep), Luna (cost-sensitive), Terra (balanced).

---

Running the new surfaces

Web dashboard

pip install -e ".[web]"
tradingagents-web            # serves the app; open the printed URL

For dev mode (Vite dev server + uvicorn --reload) and the full architecture, see web/README.md.

Backtest / evaluation harness

The harness is a Python API (not a CLI). It appends one JSONL row per evaluated (ticker, date) and skips already-resolved points, so a killed run resumes for free.

from tradingagents.backtest import run_backtest

rows = run_backtest(
    tickers=["NVDA", "AAPL"],
    start="2026-01-01",
    end="2026-06-01",
    cadence_days=21,        # how often to sample an entry
    holding_days=5,         # forward window per entry
    results_path="backtest.jsonl",
    config=None,            # or a DEFAULT_CONFIG.copy()
)

See tradingagents/backtest.py for summarize(), baselines, and the cost model.

Per-stage model routing & adaptive risk debate

from tradingagents.default_config import DEFAULT_CONFIG

config = DEFAULT_CONFIG.copy()
config["agent_llm_map"] = {
    # keys: analysts, researchers, research_manager, trader,
    #       risk_analysts, portfolio_manager, investor_briefing
    # values: "quick" | "deep"   (empty map = current defaults)
    "analysts": "quick",
    "research_manager": "deep",
    "portfolio_manager": "deep",
}
config["adaptive_extra_rounds"] = 2   # 0 = today's fixed cap

---

Base install & usage

For the full base install, provider API keys, Docker, and the interactive CLI, see the upstream README. Minimal quickstart so this repo runs standalone:

git clone https://github.com/klau1011/TradingAgents.git
cd TradingAgents
pip install .
export OPENAI_API_KEY=...   # or any supported provider's key
tradingagents               # interactive CLI

Programmatic use is unchanged from upstream:

from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG

ta = TradingAgentsGraph(debug=True, config=DEFAULT_CONFIG.copy())
_, decision = ta.propagate("NVDA", "2026-01-15")
print(decision)

---

Fork changelog

Notable fork commits on top of upstream (git log upstream/main..HEAD):

• Web dashboard: backend/frontend, cancellable + resumable streams, agent Gantt, run persistence, key preflight, report links, investor summaries.
• Backtest / evaluation harness with look-ahead fixes, baselines, and cost.
• Model routing + adaptive risk debate; memory relevance + wider lessons.
• ETF analysis support; run-scoped OHLCV cache; volume indicators.
• GPT-5.6 model support.

---

Attribution & citation

Forked from TauricResearch/TradingAgents. All credit for the base multi-agent framework goes to the original authors. Please cite their work:

@misc{xiao2025tradingagentsmultiagentsllmfinancial,
      title={TradingAgents: Multi-Agents LLM Financial Trading Framework}, 
      author={Yijia Xiao and Edward Sun and Di Luo and Wei Wang},
      year={2025},
      eprint={2412.20138},
      archivePrefix={arXiv},
      primaryClass={q-fin.TR},
      url={https://arxiv.org/abs/2412.20138}, 
}