TradingAgents v0.4.0 fixes look-ahead data leaks and adds GPT-5.6 and GLM-5.3

TradingAgents v0.4.0 arrives with point-in-time data fixes

TradingAgents, the open-source framework that assigns large language models to the roles of a trading firm, shipped version 0.4.0 in August, and the project drew fresh attention after its GitHub repository appeared on Hacker News on September 8. According to the project's changelog, maintained by TauricResearch, the update is largely about data honesty: agents now operate on information that would genuinely have been available on the date under analysis.

What the new version changes

The 0.4.0 entry lists look-ahead and point-in-time corrections in three areas: FRED macroeconomic data, the social-sentiment feed, and the persistent decision log that serves as agent memory. Look-ahead bias, where a simulated strategy unknowingly draws on data published after the decision moment, is a classic way for backtests to overstate performance, so aligning these surfaces to a strict timeline makes replayed runs a fairer measure of what the agents would actually have decided.

Beyond the timeline fixes, the release brings clearer decision signals, a working checkpoint-resume path in the command-line interface so an interrupted run can continue where it stopped, and price grounding for the Trader agent, which ties its trade sizing and timing to observed prices rather than the model's own recall. The supported model catalog also grows to include GPT-5.6 and GLM-5.3.

A trading firm assembled from agents

The framework splits trading work into specialist teams. An analyst group covers fundamentals, sentiment aggregated from news headlines plus StockTwits and Reddit chatter, worldwide news with macroeconomic interpretation, and technical indicators such as MACD and RSI. A research team of one bullish and one bearish agent then debates those findings to weigh potential gains against risk.

A Trader agent turns the resulting reports into decisions about when and how much to trade. A risk-management layer evaluates volatility, liquidity and other exposures and reports to a Portfolio Manager, which approves or rejects each proposal; approved orders are executed on a simulated exchange. The repository is explicit that the framework exists for research, that results vary with the underlying model, temperature, trading period, data quality and other non-deterministic factors, and that it is not financial, investment or trading advice.

Wide provider and market coverage

TradingAgents accepts API keys for OpenAI, Google, Anthropic, xAI, DeepSeek, Qwen, GLM, MiniMax and OpenRouter, alongside Azure OpenAI and AWS Bedrock deployments. Local inference works through Ollama or any OpenAI-compatible server such as vLLM, LM Studio or llama.cpp, with the model defined by whatever the server hosts. Configuration can be driven by environment variables, with API-key auto-detection.

On the market side, the framework handles anything Yahoo Finance covers, using exchange-suffixed tickers for the United States, Hong Kong, Tokyo, London, India, Canada, Australia and mainland China A-shares, plus crypto pairs such as BTC-USD and ETH-USD. Alpha Vantage, FRED and Polymarket are among the supported data vendors.

A fast release cadence

The project has iterated quickly through 2026. Version 0.2.0 in February introduced multi-provider support; April's 0.2.4 added structured-output agents, LangGraph checkpointing, a persistent decision log and Docker images; May's 0.2.5 grounded the Sentiment Analyst and hardened ticker handling; June's 0.3.0 brought a verified data-access contract and an expanded provider registry including NVIDIA, Kimi, Groq and Mistral; and July's 0.3.1 added Alpha Vantage look-ahead filtering, working crypto sentiment sources and support for Claude Sonnet 5 and Fable 5. A Trading-R1 technical report appeared in January, with a companion Terminal described as expected soon.

Why it matters

Point-in-time correctness is unglamorous but decisive for anyone evaluating LLM agents on financial tasks: a framework that leaks future data produces impressive-looking results that collapse in live use. By fixing look-ahead issues across macro data, sentiment feeds and its own memory, TradingAgents makes its replayed decisions a more credible benchmark, and the fixes land in an open-source project that anyone can inspect and reproduce.

The release also shows how quickly agentic frameworks now absorb each new frontier model, with GPT-5.6 and GLM-5.3 arriving here a month after Claude Sonnet 5 and Fable 5, across a provider list that spans hyperscalers and self-hosted servers. For researchers comparing models in a multi-agent, debate-driven setting, that combination of breadth, provider neutrality and reproducibility is the real draw, provided the project's own caveat is kept in mind: it runs against a simulated exchange and is built for study, not for trading.