JPMorgan's AI Agents Beat Classic 60/40 Portfolio in Backtests

JPMorgan Chase has published the results of an internal study in which AI agents making capital allocation decisions outperformed both the classic 60/40 portfolio and the bank's own rule-based model in historical backtests. It is one of the first major banks to publicly demonstrate that agentic AI systems can manage assets better than strategies that have been trusted for decades.

How the Agents Work

JPMorgan's cross-asset systematic strategy team designed a system that classifies the market into one of four macroeconomic regimes based on economic growth and inflation: Goldilocks, reflation, stagflation, and risk-off. The AI agents were tasked with deciding how to allocate capital across asset classes in each of these scenarios.

During periods of strong growth, the agents increased their exposure to equities, and when the economic outlook worsened, they shifted capital toward bonds and fixed-income instruments. This logic resembles classic macro management, but the decisions on asset weightings were made by a language model rather than a team of analysts following rigid rules.

Backtest Numbers

The key result is a 0.7 percentage point advantage in annual return for the best agent over a portfolio made up of 60 percent stocks and 40 percent bonds, a strategy that has served for decades as the benchmark for pension funds and investment advisors worldwide. Equally significant is the lower volatility, since the agent achieved this result with volatility 2.8 percentage points lower per year.

Sharpe ratios, which measure risk-adjusted returns, ranged from 0.74 to 0.95 for the agents, while the 60/40 portfolio achieved 0.61 over the same period. All eight agent variants, which differed in factors such as the underlying model and parameters, beat both the market benchmark and JPMorgan's own internal regime-based model.

The Bank's Caveats

JPMorgan explicitly notes that the results come from historical simulations rather than live investing, and warns against treating them as proof that AI will consistently beat the market in the future. Historical market data differs from current conditions, and agents trained and tested on the past may not perform as well in new, unforeseen macroeconomic scenarios.

The results are based on historical simulations, not live investing, and should not be treated as proof that artificial intelligence can consistently beat the market - cross-asset systematic strategy team, JPMorgan

What This Means for Finance

JPMorgan's study fits into a broader trend of agentic investing, in which large financial institutions are testing language models not just as analytical tools but as entities making real allocation decisions. Similar experiments were previously run by smaller fintech platforms, but the involvement of one of the world's largest banks gives the topic a different weight.

For fund managers and investment advisors in Poland, the result is a signal that agentic AI models are ceasing to be a curiosity and becoming a real reference point in designing asset allocation strategies. Polish financial institutions, including pension funds and TFIs (investment fund companies), are increasingly examining similar tools, though production deployments remain cautious for now due to regulatory requirements and the risk of model errors.

JPMorgan has not disclosed whether or when it plans to deploy AI agents in the actual management of client capital. The bank stresses that the study is exploratory in nature, and that any decision to deploy the technology will require further testing under market conditions that differ from the historical data on which the agents were evaluated.

Sources: JPMorgan Builds AI Agents That Beat 60/40 Portfolio in Backtests (bloomberg.com), JPMorgan's AI Agents Beat 60/40 Portfolio, Its Own Rule-Based Regime in Backtests (seekingalpha.com)