TradingAgents

TradingAgents is a research project from TauricResearch that uses language-model agents to analyze financial-market information. It assigns different roles to fundamentals, sentiment, news and technical analysis, then combines their reports through a research debate and risk-review process. The repository explicitly describes the system as research software, not financial or trading advice. Its generated decisions and simulated exchange behavior do not establish a profitable strategy.

Examine the roles and the actual evidence

The repository describes analysts working in parallel before bullish and bearish researchers debate their reports. A trader role synthesizes the analysis, followed by risk reviewers and a portfolio manager. These are software roles, not licensed professionals assessing suitability for an individual.

The project uses LangGraph and supports multiple model providers, with local and compatible-server options also documented. Model choice and data availability affect the output. Compare the underlying reports and sources when evaluating the system; agreement among several agents can still reflect shared inputs or a repeated error. A staged debate is not independent verification by itself.

Check when data became available

The README describes SEC EDGAR statements as filed, using publication timing for historical analyses. It withholds historical Yahoo fundamentals where the provider cannot establish when figures became public, and notes similar limits for insider-trade data. Those distinctions matter when studying hindsight bias: a number belonging to a fiscal period was not necessarily known during that period.

The release notes claim point-in-time improvements, but the reproducibility section still warns that live news and social sources can reflect current content even for a historical analysis date. That caveat limits how a dated run should be interpreted. Review the actual source timestamps and saved configuration rather than assuming one date setting freezes the entire information environment.

Treat backtests as experiments

The project describes a backtest runner over ticker-and-date grids, using a separate memory log and scoring against regional benchmarks. It also offers optional checkpoints for interrupted runs and persistent decision memory. Keeping experimental state separate helps avoid mixing an evaluation with unrelated prior runs, but it does not eliminate data leakage or model variability.

The README warns that identical tickers and dates can produce different outputs because language-model sampling and inputs vary. It says published returns are not guaranteed to replicate. Record the version, provider, data sources and relevant settings when comparing experiments, and inspect failed or withheld-data cases alongside completed ones. This is a tool for studying multi-agent analysis, not a recommendation to transact on its signals or assume an automated review can prevent financial loss.