JPMorgan AI Agent Beats 60/40 Portfolio in Regime Detection Test

JPMorgan demonstrates AI-driven portfolio allocation

JPMorgan strategists have developed an artificial intelligence agent that identifies macroeconomic regimes and uses that information to outperform traditional 60/40 stock-bond portfolios with lower risk, according to research led by Thomas Salopek.

The approach centers on a fundamental challenge in investing: different asset classes perform differently depending on whether the economy is in expansion, contraction, high inflation, or financial stress. JPMorgan's team fed economic and market data into multiple large language models—two from Anthropic and two from OpenAI—and asked them to classify current conditions into one of four regimes: Goldilocks (above-trend growth with falling inflation), Reflation (accelerating growth and rising inflation), Stagflation (weak growth with persistent inflation), or Risk-Off (sharp slowdown or financial stress).

The models produced similar but not identical assessments. Testing across historical periods from 2001 to 2026, the AI agents correctly identified major episodes: the dot-com unwind as risk-off and stagflation, the mid-2000s as Goldilocks, the 2008 financial crisis as risk-off, and the 2010s as predominantly Goldilocks. For early 2026, all models identified a transition from Goldilocks into Reflation.

JPMorgan then applied fixed investment strategies tailored to each regime. On average, the AI-guided allocations beat the benchmark 60/40 portfolio.

Replication reveals model inconsistency

MarketWatch attempted to recreate JPMorgan's methodology using publicly available data and free AI tools. Economic indicators came from the Federal Reserve's FRED database, while market data was sourced from exchange-traded funds tracking Treasury bonds, the dollar, commodities, volatility, and corporate credit.

The experiment succeeded in obtaining regime classifications from multiple AI platforms—Claude, ChatGPT, Grok, and Gemini—all of which favored a Reflation assessment for the current period, consistent with JPMorgan's June findings. However, the models disagreed on secondary probabilities and confidence levels. Claude assigned 60% probability to Reflation with Stagflation as second-most likely, while ChatGPT gave 43% to Reflation and 31% to Goldilocks.

More problematically, repeated queries to the same model sometimes produced different answers, revealing the inherent variability in how large language models process identical inputs.

JPMorgan acknowledged a potential data leakage problem: even when prompts are date-anonymized and use lagged data, the models were trained on information from after those dates and may implicitly recognize historical episodes like the 2008 crisis or COVID-19 crash.

Why it matters

This research demonstrates both the promise and limitations of applying generative AI to investment decisions. While large language models can synthesize complex economic and market signals to identify regime shifts—a task traditionally requiring deep expertise—their outputs remain probabilistic and inconsistent. The disagreement between models and even within the same model across queries suggests AI-driven portfolio allocation requires human oversight rather than autonomous execution. For institutional investors exploring AI integration, JPMorgan's work provides a practical framework while highlighting the need for validation and risk controls.

The details of JPMorgan's AI agent research were first reported by MarketWatch.