Full opportunity report: Introducing Forezai · TradingAgents — a committee of LLMs decides paper-trades on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Forezai has launched TradingAgents, a system where multiple LLMs collaborate in a structured committee to generate paper-trading decisions. This approach aims to explore AI’s potential in financial decision-making without risking real money, marking a significant step in AI research for trading.
Forezai has introduced TradingAgents, an operational system that employs a committee of large language models to generate paper-trading decisions in a structured, multi-agent framework. This development transforms recent research into an accessible, practical research tool, emphasizing transparency and safety in AI-driven trading experiments.
The new project, Forezai · TradingAgents, is a fork of an existing multi-agent research framework that uses specialized LLM roles to analyze market data, debate, and synthesize trading signals. Unlike previous experiments that focused solely on theoretical models, this version adds operational features such as automated scheduling, paper trading via multiple brokers, and real-time monitoring through a web dashboard.
The system operates by running a daily cycle where the agent committee evaluates a watchlist of stocks or assets, generates buy, hold, or sell signals, and executes paper orders accordingly. It includes safeguards like cooldowns, sector caps, and position management to prevent unintended real-money risks. The framework also logs all decisions for later analysis, ensuring transparency and reproducibility.
Forezai emphasizes that the system does not claim LLMs predict markets reliably but instead explores whether structured, multi-voiced reasoning can produce decisions at least no worse than random, given the same data a human would see. The project is designed for research purposes, not for live trading or financial advice.
Introducing Forezai · TradingAgents — Thorsten Meyer AI
AGENTS
FOREZAI · 03
TRADINGAGENTS · LAUNCH
Research Series · Companion to Polybot Week 1-2 · 2026-05-17
Introducing Forezai · TradingAgents.
A committee of LLMs
decides paper-trades.
After two weeks of finding out most parametric strategies don’t work, the obvious next research question: can multi-agent LLM judgment do any better?
A fork of the open-source TradingAgents framework (TauricResearch): thirteen LLM agents in four stages — four parallel analysts · a bull-bear debate with research-manager arbitration · a three-voice risk team · a two-layer trader + portfolio-manager decision. The fork keeps the agent graph intact and adds the operational layer the upstream doesn’t ship: an autonomous loop · a multi-broker abstraction · a local web dashboard · Codex OAuth · MCP plug-ins · 520+ unit tests. The question is narrower than “do LLMs predict the market” — that prior is “no, with high confidence.” The narrower question is: when LLMs are structured into specialised adversarial roles, does the committee produce decisions at least no worse than a coin flip after fees? Honest priors before running: it might fail too. If it appears to work, the most likely explanation is variance.
Thorsten Meyer AI
Forezai · TradingAgents
Apache-2.0 fork · upstream cited
Companion piece · ~2,500 words
Polybot · § 03
This is not financial advice. Nothing in this announcement should be used to inform real trading decisions. The software described trades simulated money by default. If you reconfigure it to trade real money, you should expect to lose that money — regardless of how clever any individual agent’s reasoning looks. Algorithmic trading is zero-sum after fees and structurally hostile to part-time retail strategies.
The bet is on a different mechanism, not a different parameter setting. Polybot’s strategies are explicit rules — most measured as structurally negative-expectation. TradingAgents asks a separable question: does multi-agent LLM judgment, with built-in adversarial debate, survive contact with reality?
13 agents
Specialised roles in four stages
Analysts · Debate · Risk · Decision
78% / -33%
Polybot prior: fleet win rate
combined with -33% bankroll
520+
Passing unit tests across engine,
services, HTTP routes (starting baseline)
€0 floor
LLM cost on Codex OAuth
(falls back to public API per token)
FOREZAI / TRADINGAGENTS·
APACHE 2.0 FORK·
UPSTREAM TAURIC RESEARCH·
LANGGRAPH·
13 AGENTS / 4 STAGES·
4 PARALLEL ANALYSTS·
BULL-BEAR DEBATE·
3-VOICE RISK TEAM·
TRADER + PORTFOLIO MANAGER·
5-TIER FINAL RATING·
ALPACA PAPER + LOCAL + SHADOW·
LIVE ENDPOINTS HARD-REFUSED·
FASTAPI + REACT VIA CDN·
CODEX OAUTH·
MCP PLUG-IN REGISTRY·
520+ UNIT TESTS·
POLYBOT WEEK 1: 21 EXPERIMENTS·
WEEK 2: -33% BANKROLL·
78% FLEET WIN RATE·
HONEST RESEARCH, NOT EDGE·
FOREZAI / TRADINGAGENTS·
APACHE 2.0 FORK·
UPSTREAM TAURIC RESEARCH·
LANGGRAPH·
13 AGENTS / 4 STAGES·
4 PARALLEL ANALYSTS·
BULL-BEAR DEBATE·
3-VOICE RISK TEAM·
TRADER + PORTFOLIO MANAGER·
5-TIER FINAL RATING·
ALPACA PAPER + LOCAL + SHADOW·
LIVE ENDPOINTS HARD-REFUSED·
FASTAPI + REACT VIA CDN·
CODEX OAUTH·
MCP PLUG-IN REGISTRY·
520+ UNIT TESTS·
POLYBOT WEEK 1: 21 EXPERIMENTS·
WEEK 2: -33% BANKROLL·
78% FLEET WIN RATE·
HONEST RESEARCH, NOT EDGE·
FIG. 01 — THE 13-AGENT COMMITTEE
Thirteen specialised roles · four stages · biases made to argue in public
The architecture forces the system to articulate its reasoning rather than relying on what a single context window happens to recall
Stage 1 · Four analysts in parallel4 agents
Market
Structure, ranges, regime indicators
News + Insider
News flow, filings, insider activity
Fundamentals
Balance sheet, earnings, ratios
Social Sentiment
Social-media tone, retail signal
↓
Stage 2 · Bull-bear debate + research-manager arbitration3 agents
Bull researcher
Argues upside thesis from analyst reports
Bear researcher
Argues downside thesis from same reports
Research manager
Arbitrates · writes single synthesis
↓
Stage 3 · Three-voice risk team3 agents
Aggressive
Looks for upside · accepts variance
Conservative
Looks for downside · protects capital
Neutral
Balances · forces downside articulation
↓
Stage 4 · Two-layer decision2 agents
Trader
Three-tier proposal · buy / hold / sell
Portfolio manager
Five-tier rating + price target + horizon · sees arguments only, never raw data
The portfolio manager only sees the arguments, never the raw data — which forces the committee to make its reasoning explicit rather than relying on a single context window’s recall. The upstream framework ships the agent graph; it does not ship the operational machinery to run that graph on autopilot, observe its results honestly, store them for later inspection, or prevent the operator from accidentally trading real money. That gap is what the Forezai fork fills.
FIG. 02 — THE POLYBOT PRIOR · WHY THIS IS A DIFFERENT BET
Two weeks of paper-trading prediction markets · the trap underneath the headline numbers
25 experiments · 78% fleet-wide win rate · -33% bankroll · most parametric strategies are structurally negative-expectation when measured honestly
The flattering number
78%
Fleet-wide win rate · week 2
“You can win four out of five trades and still go broke, because the one loss is bigger than the four wins put together.” Win rate without P&L context is a mechanical illusion.
The honest number
−33%
Fleet bankroll · week 2 close
The strongest possible demonstration of the trap. A parametric trading strategy that looks compelling in a backtest will almost always fail to survive a fresh sample. Most “edges” are mechanical artefacts.
Week 1: 21 parallel strategy experiments · early winners mostly mechanical illusions · exactly one strategy (a fair-value taker on BTC) showed the mathematical signature of real edge over a few hundred settled trades. Week 2: same fair-value strategy with more data collapsed. A separate mid-week hypothesis (market-making) also failed cleanly. Fleet ended week 2 at roughly negative thirty-three percent of bankroll. The honest research finding wasn’t on the winning side — it was on the losing side. Adding more parameters to Polybot wouldn’t change that. TradingAgents is asking a separable question.
FIG. 03 — WHAT THE FORK ADDS · THE OPERATIONAL LAYER
Six layers the upstream framework doesn’t ship
Same agent graph, intact. The fork makes it a research instrument rather than a tech demo.
01 · Loop
An autonomous loop
Scheduler · watchlist · auto-trader maps ratings to paper orders · allow-list filtering · per-ticker cooldowns · sector caps · cash checks · position manager evaluates open positions every 60s for TP / SL / max-hold. Append-only audit logs.
02 · Brokers
Multi-broker abstraction
Three modes: local Python broker (yfinance fills, JSON-persisted) · Alpaca paper-trading adapter · “shadow” mode running both in parallel with divergence view. Real Alpaca live endpoints are hard-refused at multiple layers.
03 · Dashboard
A local web dashboard
FastAPI backend · React via CDN, no Node toolchain · SVG equity curve · rolling-peak drawdown · win-rate by rating / ticker / model · exit-reason breakdown · LLM cost vs realised P&L joined by run ID. Runs locally; nothing sent to a cloud service.
04 · Codex
Codex OAuth
Runs the engine on a ChatGPT Pro subscription via the Codex backend. LLM cost floor effectively zero if you already have ChatGPT Pro. Token stored encrypted locally. Falls back to the regular OpenAI API if you’d rather pay per token.
05 · Alerts
Multi-channel alerts
Slack · Discord · SMTP email · configurable filter on rating events and order fills · append-only history kept locally. Webhook URLs masked in API responses so a screenshot can’t accidentally leak credentials.
06 · MCP
MCP plug-ins
Registry for adding Anthropic Model Context Protocol servers (Kensho · Aiera · FactSet · Morningstar · LSEG) as analyst tools. Plug-ins advertise category (fundamentals · news · market data · social) · probe endpoint tests credentials.
Honest-by-design touches: every generated report prepends “Research, not advice” and appends a footer with version, commit, provider, models used, run ID, and cost. Closed trades carry the same metadata. 520+ passing unit tests across engine, services, and HTTP routes. The intent: when the system loses money, the journal makes it impossible to pretend it didn’t.
FIG. 04 — HONEST PRIORS · BEFORE RUNNING THIS IN ANGER
Three priors stated before the data starts arriving
The bias of the project: when the data says no, the dashboard says no, the article says no
1
It might fail too. LLMs are not oracles, and a sophisticated framework around language-model outputs does not change the underlying error rate of the model. Sample is still everything. The framework’s outputs are subject to the same statistical noise as any prediction system over small samples.
Highest likelihood
2
If it appears to work, the most likely explanation is variance. The same trap that caught the first article’s candidate edge applies here. A high win rate over fifty trades means much less than it looks. Without out-of-sample confirmation, a flattering early sample tells you almost nothing about whether the system has real edge.
Second-most likely
3
If it appears to work for the right reasons — empirical win rate matches stated confidence, and alpha-versus-benchmark persists across non-overlapping samples — that would be a meaningful research finding. Whether that happens, I don’t know. The point of putting it in the open is that the data will say.
Genuinely open
This is explicitly not a launch announcement for a product anyone should connect a real brokerage account to. The Alpaca live endpoints are hard-refused at multiple layers in the code, and the design choice is deliberate. The right next step is data, not deployment. The bias of the whole project is straightforward: when the data says no, the dashboard says no, the article says no, and no one tries to retroactively rescue the thesis. That’s the contribution.
FIG. 05 — WEEK THREE · WHAT THE METHODOLOGY WILL MEASURE
Four concrete measurements before publishing findings
The hope: write the week-three article from a position of “here’s what the data says”. The fear: another candidate falsified at higher sample. Both outcomes are publishable.
M1 · Sample discipline
Small watchlist for a few weeks before publishing
A handful of tickers across two or three sectors. Long enough to gather sample, narrow enough to keep attention on what’s actually happening per agent. Avoid the noise of a 65-ticker autonomous loop until the smaller version has been read carefully.
M2 · Calibration view
Stated confidence vs. realised win rate
When the system says “75% confident”, do the trades actually win 75% of the time? Same measurement applied to Polybot’s fair-value model. If the model is systematically over-confident, that bias dominates everything downstream.
M3 · Cost accounting
Cost per ticker · per rating · per profitable trade
With Codex OAuth the marginal LLM cost is effectively zero. With the public OpenAI API, each run is hundreds of agent turns. The honest question: does this scale economically if you ever did run it at real cost?
M4 · Non-overlapping windows
Alpha vs benchmark · out-of-sample
Not within-sample alpha — trivially inflatable. Hold out one period entirely, run the system on the next, then check whether the held-out result matches the in-sample stats. If they diverge sharply, the in-sample was curve-fit.
Open under Apache-2.0 with upstream cited from every relevant surface. Not open: the operator’s running results, the specific watchlist, the per-agent prompt customisations, the alert channels, the trade journals — kept local for the same reason Polybot’s per-experiment data is kept local. Publishing exact configurations encourages people to copy them with real money, which is the opposite of what an honest research project should do. Summary findings will be published. Recipes will not.
The bet is on a different mechanism, not a different parameter setting. The point is not to find a money-printing AI. The point is to put honest measurements of these systems into the public record — so the next person looking at the space starts a step further along than the last.
Thorsten Meyer AI · Introducing Forezai · TradingAgents · § 03
Potential for AI-Driven Market Research Tools
This development demonstrates a concrete step toward operationalizing AI research in finance, allowing systematic testing of multi-agent LLM frameworks in simulated trading environments. It highlights the potential for AI systems to articulate reasoning through structured debate, which could inform future AI-assisted decision-making tools. However, it remains a research prototype, not a commercial trading system, and its broader implications for market behavior or AI reliability are still uncertain.
From Parametric Strategies to Multi-Agent AI Frameworks
Previous research by Thorsten Meyer and colleagues showed that simple parametric trading strategies often fail in live simulations despite promising backtests, revealing the prevalence of overfitting and mechanical artifacts. This led to questions about whether less rule-bound AI approaches, like multi-LLM committees, could perform better.
The underlying research framework, originally published by TauricResearch, involves specialized roles—analysts, debate agents, risk teams, and decision syntheses—that argue and justify trading decisions explicitly. The new Forezai fork extends this framework into an operational environment, enabling automated, repeatable testing of these AI committees in simulated markets.
“This project turns theoretical multi-LLM debate frameworks into a practical research tool, allowing us to test AI decision-making in simulated trading environments.”
— Thorsten Meyer
Limitations and Unanswered Questions in AI Trading Research
It is still unclear how well the AI committee’s decisions will generalize beyond simulated environments or whether they can outperform traditional strategies in live markets. The system explicitly does not predict market movements but explores reasoning processes, so its practical trading utility remains to be validated.
Additionally, the long-term stability, robustness under different market conditions, and potential biases of the AI agents are still being studied. The project is primarily a research prototype, and operational risks or unintended behaviors are not fully understood yet.
Next Steps for Testing and Validation of AI Committee Trading
Future work will involve extensive backtesting and live simulation to assess the AI committee’s decision quality over longer periods and diverse market conditions. Researchers plan to analyze decision logs to identify strengths and weaknesses of the approach, and possibly refine agent roles or debate structures.
Further integration with real broker APIs and enhanced safeguards could enable limited live testing, though the primary focus remains on research and understanding AI reasoning in trading contexts. The project aims to establish benchmarks for AI decision-making in finance.
Key Questions
Can Forezai TradingAgents be used for real trading now?
No. The system is currently designed for simulated, paper trading environments and is intended solely for research purposes.
What makes this AI approach different from traditional trading algorithms?
Instead of relying on fixed rules or pattern recognition, it employs a committee of specialized LLMs that debate and justify their decisions explicitly, aiming to explore AI reasoning rather than prediction accuracy.
How transparent are the AI decisions in this system?
The framework logs all agent arguments and decision rationales, making the reasoning process explicit and reviewable for research and analysis.
Does this project aim to replace human traders?
No. It is a research tool to understand AI decision-making in trading contexts, not a commercial trading system or advice platform.
What are the main limitations of this system?
Its performance in live markets remains untested, and it does not predict market movements. Its decisions are based on structured debate rather than market forecasts, and risks of unanticipated behaviors exist.
Source: ThorstenMeyerAI.com