AI scans 40 years of factor data to flag market peaks early

Axioma by SimCorp and data analytics firm EDS have published research suggesting that AI-assisted analysis of style factor returns can identify consistent behavioural fingerprints in the weeks before equity market peaks, a capability that previously demanded days or weeks of quant analyst time.

The collaboration applied EDS's AI engine to more than 40 years of factor data drawn from Axioma's US trading-horizon risk model, examining returns for 19 style factors across 15 market peaks dating back to 1982. The central question: do factor returns shift in predictable ways as markets near a top, and can an AI engine surface those patterns faster and more reliably than traditional quantitative methods?

Two signals, 40 years of consistency

The research identified Medium-Term Momentum and Liquidity as the most durable pre-peak signals. Medium-Term Momentum posted positive 60-day pre-peak returns in 13 of the 15 cycles studied, with average returns substantially above the historical baseline for that window. Liquidity was positive in 12 of 15 equivalent periods. Axioma's Melissa Brown describes the Liquidity finding as particularly instructive: "Liquid, easy-to-trade stocks outperform heading into tops. This reads as institutional positioning, large holders quietly rotate toward names they can exit cleanly. It's not a panic signal; it's a preparation signal."

Downside Risk and Dividend Yield, by contrast, showed progressively negative returns as each peak approached, while Growth and Profitability became more positive, consistent with a market in the late stages of actively sorting winners from losers even as the headline index appears calm.

The one notable exception is the dot-com peak of March 2000, where Medium-Term Momentum produced its most negative pre-peak 60-day return of any cycle in the study. The researchers draw a cautious parallel to today: Momentum has been strong for an extended period but has recently decelerated, raising the question of whether that deceleration signals proximity to a top or merely echoes the anomalous 2000 episode. The AI's own assessment is that "the patterns that persist are better suited for characterising market regimes than for timing."

The productivity shift behind the signal

The more structurally significant finding for asset managers may be operational rather than purely analytical. Ben Lieblich, Chief Data Scientist at EDS, notes that building a 60-day cumulative factor-return window and ranking it against a 40-year historical distribution "used to be a morning's work for a quant." With the AI overlay, it becomes "a sentence and an answer." That compression does not merely save time; it lowers the friction threshold for portfolio managers to ask research questions they would previously have skipped.

This matters beyond the immediate findings. As factor-based investing grows more crowded and rotations between style regimes grow sharper, the competitive advantage is shifting from access to data toward the speed and granularity of analysis. Axioma's risk models, which underpin the research, are used by a significant share of the world's largest asset managers. The combination of a widely adopted risk model with a capable AI research engine creates a template for systematic, rapid factor-regime analysis that individual quantitative teams would struggle to replicate at the same pace.

The convergence angle for capital allocators extends beyond equities. Insurance companies, pension funds, and sovereign wealth vehicles that use factor tilts to manage liability-matching portfolios face the same regime-identification problem. If AI-assisted factor analysis can reliably compress the detection lag between a regime shift beginning and a portfolio manager acting on it, the implications reach into fixed-income factor premia, multi-asset risk budgeting, and even the alternative risk premia strategies that have proliferated across hedge fund and liquid-alternatives allocators over the past decade.

The researchers are explicit that the findings are not timing signals. They present the factor fingerprints as one lens among several, to be combined with other data points when assessing market risk. But the underlying message is pointed: the constraint on high-quality quantitative research has shifted from capability to curiosity. The bottleneck is no longer whether the analysis can be done, but what questions managers think to ask.