Test-Time Efficient Pretrained Model Portfolios for Time Series Forecasting

2025-10-07

AI Summary

In time series forecasting, a single large foundation model is not necessarily optimal. This paper proposes an “ensemble of expert models” paradigm: first pretraining a set of lightweight general-purpose models, then specializing them via post-training to yield diverse expert models, and finally employing ensemble or model selection strategies at inference time. This approach avoids the high parameter count and computational overhead inherent in monolithic large models, significantly improving both inference efficiency and generalization capability. Experiments on large-scale benchmarks demonstrate that our method achieves performance comparable to—or even surpassing—that of large models, using substantially fewer parameters; inference cost is reduced by up to several-fold, while maintaining strong scalability. The core contribution lies in empirically validating the superiority of specialized small-model ensembles in the accuracy–efficiency trade-off, and in establishing that model ensembling consistently outperforms test-time fine-tuning.

Technology Category

• Machine Learning: Ensemble Methods
• Computer Vision: Large Vision Models
• Natural Language Processing: (Large) Language Models

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Is bigger always better for time series foundation models? With the question in mind, we explore an alternative to training a single, large monolithic model: building a portfolio of smaller, pretrained forecasting models. By applying ensembling or model selection over these portfolios, we achieve competitive performance on large-scale benchmarks using much fewer parameters. We explore strategies for designing such portfolios and find that collections of specialist models consistently outperform portfolios of independently trained generalists. Remarkably, we demonstrate that post-training a base model is a compute-effective approach for creating sufficiently diverse specialists, and provide evidences that ensembling and model selection are more compute-efficient than test-time fine-tuning.

Problem

• Exploring smaller pretrained model portfolios for time series forecasting
• Achieving competitive performance with fewer parameters via ensembling
• Demonstrating compute-efficient specialist creation and selection strategies

Innovation

• Portfolio of smaller pretrained forecasting models
• Ensembling or model selection over specialist collections
• Post-training base model for diverse specialists