Rethinking Irregular Time Series Forecasting: A Simple yet Effective Baseline

TL;DR

APN addresses irregular multivariate time series forecasting by decoupling irregularity handling from forecasting, using Time-Aware Patch Aggregation to produce high-quality, regular patch representations via Adaptive Patching and Weighted Aggregation. A lightweight Query-based Aggregation and a shallow MLP decoder then perform forecasting, yielding a simple yet effective backbone. Experiments on PhysioNet, MIMIC, HumanActivity, and USHCN show APN achieves state-of-the-art or competitive accuracy with significantly improved efficiency, across healthcare, biomechanics, and climate domains. The work provides a strong, practical baseline for IMTSF with open-source code.

Abstract

The forecasting of irregular multivariate time series (IMTS) is crucial in key areas such as healthcare, biomechanics, climate science, and astronomy. However, achieving accurate and practical predictions is challenging due to two main factors. First, the inherent irregularity and data missingness in irregular time series make modeling difficult. Second, most existing methods are typically complex and resource-intensive. In this study, we propose a general framework called APN to address these challenges. Specifically, we design a novel Time-Aware Patch Aggregation (TAPA) module that achieves adaptive patching. By learning dynamically adjustable patch boundaries and a time-aware weighted averaging strategy, TAPA transforms the original irregular sequences into high-quality, regularized representations in a channel-independent manner. Additionally, we use a simple query module to effectively integrate historical information while maintaining the model's efficiency. Finally, predictions are made by a shallow MLP. Experimental results on multiple real-world datasets show that APN outperforms existing state-of-the-art methods in both efficiency and accuracy.

Figures (4)

• Figure 1: Fixed Patching vs. Adaptive Patching.
• Figure 2: The overall framework of APN, which initially divides each univariate irregular time series into a series of unfixed patches using the Adaptive Patching Module. Then the Weighted Aggregation Module generates high-quality and regular initial patch representations. Based on the representations, the Query-based Aggregation Module is utilized to incorporate contextual information. Finally, the Forecasting Decoder outputs the final forecasting results. The Adaptive Patching Module and Weighted Aggregation Module collectively form the Time-Aware Patch Aggregation Module.
• Figure 3: Parameter sensitivity studies of main hyperparameters in APN.
• Figure 4: Comparison of computational efficiency for APN and three representative baselines. We evaluate four key metrics: (a) peak GPU memory (GB) during a single training step, (b) total number of parameters (M), (c) average training time per step (ms), and (d) average inference time per step (ms). All experiments are conducted on the USHCN dataset with a consistent batch size of 32 to ensure a fair comparison. For all metrics, a lower value indicates better performance.