- Joao Wedson warns LLMs are not built for trading and can amplify losses without risk controls.
- Wedson advocates a quantitative trading approach with strategies backtested across market cycles.
- Researchers say AI relies heavily on patterns, raising concerns about its ability to adapt to new markets.
A new debate across crypto and prediction markets is challenging the idea that large language models (LLMs) can safely run trading strategies on their own.
Quants, researchers, and Polymarket traders are pushing back against the narrative. Their argument is that LLMs may sound intelligent, but they were not designed to trade financial markets.
“LLMs Are Not Trading Engines,” Says Joao Wedson
In a tweet, Joao Wedson, founder and CEO of Alphractal, warned traders against relying on LLMs to automate trading.
According to Wedson, LLMs are built to understand language, organize information, and predict text. They are not designed to forecast financial markets.
He said using them for trading, like in high-risk perpetual futures markets, is dangerous without strong systems behind them. Wedson stressed that poor trading setups become riskier when powered by AI.
Without proper risk controls, backtesting, and adaptation to changing market conditions, an AI model may simply scale bad decisions faster.
His message was that if profitable AI trading were as easy as connecting an LLM to a market, leading AI executives would already be making billions from it.
He added that successful trading depends on survival, validation, risk management, and repeatable systems, not intelligence that merely sounds convincing.
Quantitative Models Still Come First
Wedson’s view reflects a traditional quantitative approach to trading. That means strategies should be tested extensively using historical data. They should also undergo parameter tuning and performance checks across different market environments, bull markets, bear markets, volatile periods, and sideways conditions.
In this framework, AI still has a role. However, it should operate within proven trading systems rather than act as an autonomous trader.
Researchers Question How AI Actually Learns
On the other hand, AI researcher Rohan Paul highlighted a study suggesting that modern AI agents may not learn abstract concepts in the way humans do.
The research found that AI performance dropped sharply when structured memory systems were disrupted. However, performance changed very little when higher-level summary rules were altered.
The findings suggest that LLMs may rely more on historical pattern recognition than genuine reasoning or conceptual understanding.
For traders, that distinction matters. If AI struggles to generalize lessons across changing environments, its performance could become unreliable when market conditions shift.
Bottom Line
The AI trading debate is dividing the industry into two camps. One side believes LLM-powered bots will transform trading automation. The other argues that without rigorous quantitative discipline, risk management, and proven strategies, AI trading simply becomes a faster way to lose money.
Notably, interest in AI trading has surged amid publicly available GitHub repositories. Some platforms even connect LLMs to live markets to provide analysis and trade recommendations.
Despite the excitement, many traders remain cautious. Their view is that these tools can enhance analysis and workflow efficiency, but are not reliable standalone trading systems.
Related: Claude and TradingView Setup Shows AI Trading Assistant Workflow
Disclaimer: The information presented in this article is for informational and educational purposes only. The article does not constitute financial advice or advice of any kind. Coin Edition is not responsible for any losses incurred as a result of the utilization of content, products, or services mentioned. Readers are advised to exercise caution before taking any action related to the company.
