Smorster

Someone open-sourced a hedge fund (53k stars on GitHub)

Indie Hacker News · 8m 27s · Watch on YouTube · 11 sources

Decision Card

Effort: Weekend project — git clone https://github.com/TauricResearch/TradingAgents, create a Python venv, pip install, export one LLM provider API key, and run the CLI against a single ticker (~10–15 min to first run, but expect real token spend per analysis cycle).

Honest take: The video sells this as a “this weekend” 53k-star phenomenon, but the repo is already ~80k stars and the underlying paper has been public since December 2024 — this is a year-old project, not a fresh drop. And “the system literally learns from past trades” is overstated: it appends a reflection paragraph to a prompt, it does not train or update any model weights.

Concrete next steps:

  • Read the LangGraph wiring in tradingagents/ and the paper (arXiv 2412.20138) as a reference for multi-agent decision graphs (~1–2 hrs).
  • Run one ticker with a cheap/local provider (Ollama or DeepSeek) and a low debate-round count first, to gauge token cost before scaling (~30 min).
  • Skip if you want a live-trading system — the “exchange” is a backtest, not a broker integration, and the authors explicitly label it a research framework, not financial advice.

TL;DR

TradingAgents is an Apache-2.0 Python framework that simulates a hedge-fund’s org chart with one LLM agent per role (analysts, bull/bear researchers, trader, risk team, portfolio manager) wired together on LangGraph. Its real value is as a clean, fully auditable reference implementation of a multi-agent LLM decision graph — not as a money-making trading bot.

Key Points

  • The project is a multi-agent LLM framework mirroring a Wall Street firm’s structure, pointed at a stock ticker 00:14
  • The full cast: fundamentals/sentiment/news/technical analysts, bull and bear researchers who debate, a trader, a risk team, and a portfolio manager with final approval 00:28
  • Started as a UCLA research paper in late 2024, then open-sourced; stars have been doubling every couple of months 00:57
  • The four analysts run in parallel and the system deliberately does NOT collapse them into one signal — disagreement is treated as signal, not noise 03:50
  • Bull and bear researchers argue across a configurable number of debate rounds before the trader makes a call 04:01
  • The whole orchestration runs on LangGraph, so every agent is a checkpointed node and a crashed run can resume from where it stopped 04:46
  • A persistent decision log appends each run’s reasoning to a markdown file, then injects realized-return reflection into the portfolio manager’s prompt on the next run 05:04
  • Supports nearly every major provider: OpenAI, Gemini, Claude, Grok, DeepSeek, Qwen, OpenRouter, and Ollama for local models 05:45
  • v0.2.4 (April 25) added structured-output decision agents using Pydantic schemas, plus DeepSeek/Qwen/GLM/Azure OpenAI and Docker support 06:05
  • Honest caveats: each cycle burns real LLM tokens, it’s a backtest not a live broker, and the authors call it research, not financial advice 07:26

Notable Quotes

“Real hedge funds don’t have one person picking stocks. They have entire teams, and those teams argue with each other because that argument is what produces a defensible position.” 01:09

“The disagreement between them is signal, not noise.” 03:56

“The authors explicitly say this is a research framework, not financial advice, and you should not yolo your retirement into whatever the portfolio manager spits out.” 07:42

Verified Claims

Claim: The project is a multi-agent LLM trading framework on GitHub under Apache 2.0. 00:50

Claim: It crossed ~53,000 stars and ~9,700 forks. 00:45

  • Sources: TauricResearch/TradingAgents (GitHub)
  • Verdict: Confirmed for the time of recording, but stale — current snapshots show ~80k stars, so the 53k figure reflects an earlier date than the video implies.

Claim: It started as a UCLA research paper in late 2024. 00:57

  • Sources: arXiv 2412.20138, Yijia Xiao homepage
  • Verdict: Confirmed (submitted Dec 28, 2024; lead authors Yijia Xiao and advisor Wei Wang at UCLA) — though co-authors are also affiliated with MIT, so “out of UCLA” alone undersells the collaboration.

Claim: The paper is published on arXiv as 2412.20138. 03:13

Claim: The orchestration runs on LangGraph with checkpoint/resume. 04:46

Claim: v0.2.4 shipped April 25 with structured-output Pydantic agents, new providers, Docker, and a five-tier rating scale. 06:05

Claim: The framework improves cumulative returns, Sharpe ratio, and maximum drawdown over baselines. 02:13

  • Sources: arXiv 2412.20138 abstract
  • Verdict: Inconclusive — the paper claims these improvements, but they are the authors’ own backtested results, not independently replicated, and the video presents them as established fact.

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. Does TradingAgents’ claimed Sharpe/return improvement hold up out-of-sample, and has anyone independently reproduced the paper’s backtests on tickers and time windows the authors didn’t choose?
  2. What is the realistic per-ticker token cost (and latency) of a single full run across four analysts plus multiple debate rounds, and how does cost scale with debate-round count and model choice?
  3. How much does the “persistent decision log” reflection actually change subsequent decisions — is prompt-injected reflection measurably better than no memory, or is it cosmetic?

Sources