Modeling Dynamic Mixtures of Time-Delay Systems from Streaming Time Series

Researchers present DelayMix, an online machine learning framework that models streaming time series as dynamic mixtures of time-delay systems, enabling rapid adaptation to regime shifts while maintaining memory efficiency. The method uses tensor decomposition to capture system dynamics and input delays, demonstrating superior forecasting accuracy on non-stationary data compared to existing approaches.

DelayMix addresses a fundamental challenge in adaptive modeling: systems that change rapidly degrade prediction performance when using static models. The framework's innovation lies in treating incoming data streams as compositions of multiple time-delay systems rather than monolithic models, allowing it to switch between learned patterns as environmental conditions shift. This approach is particularly relevant for financial markets, sensor networks, and other domains where both system behavior and temporal relationships evolve unpredictably.

The technical contribution centers on representing historical regime information as a fixed-length tensor derived from Markov parameters, which encode both dynamic characteristics and input-output delays. By maintaining this compact summary rather than storing complete historical models, the system achieves memory efficiency—a critical constraint for streaming applications. The tensor decomposition mechanism enables rapid identification and retrieval of relevant past models when the current regime changes, supporting real-time decision-making.

For practitioners in financial analytics, forecasting, and control systems, DelayMix's demonstrated performance on non-stationary data suggests meaningful practical value. The method's computational efficiency and adaptation speed could enhance algorithmic trading systems, predictive maintenance applications, and risk modeling platforms that operate under concept drift. The framework handles a pervasive real-world problem: many time series exhibit both changing dynamics and varying latencies between inputs and outputs, yet most production systems use static lag assumptions.

Future applications may extend DelayMix to high-frequency trading scenarios, IoT sensor fusion, and automated control systems requiring sub-second adaptation. The approach's generality suggests potential integration into existing machine learning pipelines serving non-stationary data.

• →DelayMix uses tensor decomposition to dynamically model streaming time series with regime shifts and varying input delays.
• →The fixed-length tensor representation reduces memory requirements while capturing both system dynamics and temporal relationships.
• →Method demonstrates superior forecasting accuracy on non-stationary data compared to baseline approaches.
• →Framework enables rapid adaptation to environmental changes and delay variations in real-time applications.
• →Practical applications span financial forecasting, sensor networks, and automated control systems requiring adaptive modeling.