Article
Finding hidden patterns in changing systems: Adyen AI research at ICML 2026
The paper, “Identifiable Markov Switching Models with Instantaneous Effects and Exponential Families,” has been accepted at the International Conference on Machine Learning (ICML) 2026 in Seoul, South Korea.
Many real-world systems do not behave consistently over time. Financial markets, weather patterns, biological processes, and user behavior can all shift between hidden phases-or “regimes”-with different relationships and dynamics.
New research from Adyen and the University of Amsterdam introduces FlowMSM, a framework designed to identify these hidden regimes and uncover how relationships between variables change within them. In experiments using synthetic data and real-world financial-market data, FlowMSM accurately recovered both the underlying regimes and their causal structures, including in complex environments where changes occurred frequently.
The paper, “Identifiable Markov Switching Models with Instantaneous Effects and Exponential Families,” has been accepted at the International Conference on Machine Learning (ICML) 2026 in Seoul, South Korea. Adyen is also proud to sponsor ICML 2026 and support the global community advancing machine learning research.
ICML is widely recognized as one of the world's leading machine learning conferences, bringing together researchers and practitioners from academia and industry to present cutting-edge advances in AI. ICML 2026 received more than 23,000 submissions, with only 26.6% accepted, reflecting its status as one of the field's most selective and influential venues.
Our work in short
Many real-world systems change their behavior over time. Weather patterns, financial markets, biological processes, and user behavior can all switch between unobserved phases of relatively stable behavior called regimes. For example, ENSO weather patterns alternate between El Niño, La Niña, and a neutral period, with each phase displaying different but relatively stable behavior.
Detecting these regimes in time-series data is difficult because the underlying dynamics may switch frequently and could be highly complex, which particularly occurs when causal effects appear faster than the rate at which the data is measured, called instantaneous effects. For example, for diabetes patients, the effect of insulin delivery on glucose metabolism is typically faster than the five-minutes sampling frequency of glucose monitors, thus emerging as instantaneous in measurements.
FlowMSM addresses this challenge by combining a Markov Switching Model-a type of Hidden Markov Model-with modern normalizing flows capable of representing complex distributions. Alongside the framework, the research establishes theoretical guarantees for identifying regimes across a broad class of causal models based on exponential family noise.
Experiments on synthetic data and a real-world Fama-French financial-market dataset show that FlowMSM can recover hidden regimes and regime-specific causal structures, even when regimes switch frequently in complex, non-stationary environments. These results create new possibilities for understanding evolving systems whose underlying dynamics cannot be observed directly.
Research that moves in both directions
This work is part of Adyen’s broader applied-research strategy, which is focused on two connected goals: igniting innovation within the company and contributing to the wider technology community.
Over the past year, Adyen has expanded our academic collaborations by funding two PhD researchers at the University of Amsterdam, presented work at leading conferences, supported the community through conference sponsorships, and published new research and preprints.
Adyen has also developed and open-sourced DABstep, a public benchmark that helps researchers evaluate and advance AI systems.
These initiatives reflect our belief that research creates the greatest value when ideas move in both directions. Engaging with the external research community exposes our teams to new methods and perspectives. Sharing our work allows us to contribute what we have learned and participate in the collective advancement of the field.
Our contribution to ICML 2026 is another step in putting that research vision into practice.
Explore the research
Read the companion article for an accessible explanation of the motivation, approach, and findings (approx. 5 minutes, including YouTube video): Introducing FlowMSM
Read the complete paper for the methodology, theoretical results, and experiments (approx. 20 minutes): Identifiable Markov Switching Models with Instantaneous Effects and Exponential Families