Systems 101: Overconfidence

Summary

The article argues that modern LLMs (like Claude) exhibit the same overconfidence and lack of learning as inexperienced junior analysts, making them unreliable for live trading without human oversight. Using his own failed hedge decision and coding frustrations, the author warns that AI tools can amplify overconfidence and produce confident but wrong outputs, stressing that true machine learning for investing requires solving the fundamental learning problem, which frontier models do not yet handle. For markets, the piece is a cautionary note on over-reliance on AI, but offers no direct actionable market call beyond the disclosed short credit position.

•Author was short credit via HYG puts and stock short at 300% net liq, missed a squeeze due to overconfidence and not hedging overnight.
•Claude reversed its advice five times in 20 minutes without checking facts, mirroring the author's input rather than providing independent judgment.
•Building a gold signal system with Claude Code took 10x longer to verify outputs than doing the work manually, defeating the purpose of automation.
•LLMs are likened to first-year analysts: articulate, fast, but often wrong by an order of magnitude and unable to reflect on past mistakes.
•Lehman and Bridgewater trained humility through hazing and constant questioning, a culture that AI cannot replicate because it resets each session with no memory of prior errors.
•Research concepts like test-time training and activation steering are noted as potential paths to give AI 'scar tissue,' but none are production-ready.

Read time 14 min

Length 14,978 chars

Category finance

Trade Ideas

More from Campbell Ramble