JPMorgan AI beats 60-40 in backtests; advisers unconcerned

JPMorgan reported that custom-built AI agents outperformed a standard 60% equities, 40% bonds allocation in hypothetical backtests covering the past two decades. In the strongest result, the AI approach posted about 0.7 percentage point a year in excess return versus the 60-40 benchmark and showed lower volatility in the tests. The agents also performed better than the bank’s rules-based market regime model, which shifts allocations across four regime categories.

Analysts at the bank tested agentic AI systems against historical market data and configured the agents to make decisions under uncertainty. They compared simulated AI-driven allocations to actual historical returns for the 60-40 portfolio and to results from the internal regime-based model.

The JPMorgan team cautioned against treating the backtests as definitive. The analysts warned the results may reflect in-sample confidence and that the language models behind the agents could implicitly recall well-known historical outcomes despite date-anonymized prompts and lagged data inputs. “We strongly caution against uncritically accepting what amounts to in-sample, overly confident answers of AI,” the JPMorgan team wrote.

Research notes say the testing approach included steps to limit direct use of known past outcomes but acknowledged limits in current model training. Analysts also flagged market-structure risks if many investors followed similar AI signals, which could crowd trades and affect asset pricing.

Financial firms have so far mainly applied AI to tasks such as answering research questions, summarizing advisor-client conversations and automating back-office work. Few institutions have handed over investment decision-making to agentic systems.

A PwC survey of more than 1,000 respondents found that 24% would rely on AI-powered tools or assistants to make financial decisions during periods of market volatility. By contrast, 50% said they would consult online research and financial news, and 48% said they would go to a financial advisor.

Reactions from advisers reflected limited alarm. Bryan Byrer, founder of Millennial Financial Planning in Indianapolis, noted that clients often act on emotion and that he has not seen clients use AI to second-guess his guidance. “Intelligence is different from emotions, and people do emotional things with money but not always intelligent things with money,” Byrer noted.

Analysts highlighted governance and validation questions for any firm that seeks to deploy AI in portfolio construction, including model oversight, data provenance and how to test forward-looking strategies trained on past events. Firms continue to run experiments while many investors and advisers express a preference for human guidance in turbulent markets.