Community Discussion · Policy

When Large Models Start Reading Candlestick Charts: The True Capability Boundaries of LLM Technical Analysis

Mo MoJul 202026/07/20 49 views

2 AM. I'm staring at the BTC/USD 15-minute candlestick chart on my screen: RSI oversold, MACD bullish divergence, shrinking volume—a textbook rebound signal. But I didn't place an order because just the day before, I got burned by a "precise" analysis generated by an LLM. It confidently told me, "Double bottom pattern confirmed, support level is valid," and then the price broke down and crashed. Where's the problem? It's not that the LLM doesn't understand technical analysis; it's that it doesn't understand the "failure boundaries of technical analysis in real markets."

Policy

1 replies

Ctrl + Enter to reply

Sister Liang on ValuationJul 29(edited)

[quote="moyan, post:1, topic:1132"]

At 2 AM, I stared at the BTC/USD 15-minute candlestick chart on my screen: RSI oversold, MACD bottom divergence, shrinking volume—a textbook rebound signal. But I didn't place an order because the day before, I had been burned by a "precise" analysis generated by an LLM: it confidently told me "double bottom pattern confirmed, support level valid," and then the price broke down and crashed. Where did the problem lie? It wasn't that the LLM didn't understand technical analysis, but that it didn't understand the "failure boundaries of technical analysis in real markets."

At the same time, this paper on Arxiv (AI Trading: Evalu...

[/quote]

I've tracked this direction. The core issue is that LLMs lack probabilistic thinking. If you put the same signal into different volatility environments, its statistical significance varies completely, and the model doesn't perform conditional probability adjustments. The ceiling for this track is clear: it's the gap between data frequency and confidence levels.