SurF: A Generative Model for Multivariate Irregular Time Series Forecasting

2026-05-13

AI Summary

This work addresses the challenges of long temporal spans and numerical integration bottlenecks in generative modeling of multivariate irregularly sampled event streams by proposing SurF, a novel generative model. SurF uniquely formulates the time-rescaling theorem as a learnable bijection, establishing a mapping between event sequences and i.i.d. exponential noise. It integrates three efficient cumulative intensity parameterizations with a Transformer encoder to enable a unified pretraining framework across heterogeneous datasets. Evaluated under a rigorous leave-one-out protocol on six real-world datasets, SurF outperforms existing classical and neural autoregressive baselines on five out of six benchmarks, achieving state-of-the-art performance on both the Amazon and Earthquake datasets and significantly improving temporal prediction accuracy as measured by RMSE.

Abstract

Irregularly sampled multivariate event streams remain a stubbornly difficult modality for generative modeling: tokenization-based approaches break down when inter-event intervals vary by orders of magnitude, and neural temporal point processes are bottlenecked by window-level numerical quadrature. We (i) propose SurF, a generative model that uses the Time Rescaling Theorem (TRT) as a learnable bijection between event sequences and i.i.d.\ unit-rate exponential noise, enabling a single model to be trained across heterogeneous event-stream datasets; (ii) three efficient parameterizations of the cumulative intensity that scale to long sequences; and (iii) a Transformer-based encoder for multi-dataset pretraining. On six real-world benchmarks, SurF achieves the best reported time RMSE on Earthquake, Retweet, and Taobao, and is within trial-level noise of the strongest specialist on the remaining three. Under a strict leave-one-out protocol, the held-out checkpoint beats every classical and neural-autoregressive baseline on 5/6 datasets and beats every baseline on Amazon and Earthquake, an initial step toward foundation models over asynchronous event streams.

Problem

• irregular time series
• multivariate event streams
• generative modeling
• time rescaling
• asynchronous events

Innovation

• Time Rescaling Theorem
• generative modeling
• irregular time series
• cumulative intensity parameterization
• Transformer-based pretraining