Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LTSM-Bundle: A Toolbox and Benchmark on Large Language Models for Time Series Forecasting

Yu-Neng Chuang, Songchen Li, Jiayi Yuan, Guanchu Wang, Kwei-Herng Lai, Songyuan Sui, Leisheng Yu, Sirui Ding, Chia-Yuan Chang, Qiaoyu Tan, Daochen Zha, Xia Hu

TL;DR

The paper addresses the challenge of training universal large time series models (LTSMs) across heterogeneous datasets and proposes LTSM-Bundle, a modular toolbox and benchmark. It systematically evaluates design choices across prompting, tokenization, training paradigms, base models, data quantity, and dataset diversity, using a suite of TSF benchmarks and a TDengine-backed end-to-end pipeline. The study finds that time series prompts, linear tokenization, fully fine-tuned GPT-2-Medium backbones, and moderate data down-sampling with diverse datasets yield strong zero-shot and few-shot performance, often matching or surpassing full-data baselines with only 5% of the data. These findings offer a practical baseline and a scalable platform for future research and real-world TSF tasks, enabling reproducible, component-wise analysis of LTSM training strategies.

Abstract

Time Series Forecasting (TSF) has long been a challenge in time series analysis. Inspired by the success of Large Language Models (LLMs), researchers are now developing Large Time Series Models (LTSMs)-universal transformer-based models that use autoregressive prediction-to improve TSF. However, training LTSMs on heterogeneous time series data poses unique challenges, including diverse frequencies, dimensions, and patterns across datasets. Recent endeavors have studied and evaluated various design choices aimed at enhancing LTSM training and generalization capabilities. However, these design choices are typically studied and evaluated in isolation and are not benchmarked collectively. In this work, we introduce LTSM-Bundle, a comprehensive toolbox, and benchmark for training LTSMs, spanning pre-processing techniques, model configurations, and dataset configuration. It modularized and benchmarked LTSMs from multiple dimensions, encompassing prompting strategies, tokenization approaches, training paradigms, base model selection, data quantity, and dataset diversity. Furthermore, we combine the most effective design choices identified in our study. Empirical results demonstrate that this combination achieves superior zero-shot and few-shot performances compared to state-of-the-art LTSMs and traditional TSF methods on benchmark datasets.

LTSM-Bundle: A Toolbox and Benchmark on Large Language Models for Time Series Forecasting

TL;DR

The paper addresses the challenge of training universal large time series models (LTSMs) across heterogeneous datasets and proposes LTSM-Bundle, a modular toolbox and benchmark. It systematically evaluates design choices across prompting, tokenization, training paradigms, base models, data quantity, and dataset diversity, using a suite of TSF benchmarks and a TDengine-backed end-to-end pipeline. The study finds that time series prompts, linear tokenization, fully fine-tuned GPT-2-Medium backbones, and moderate data down-sampling with diverse datasets yield strong zero-shot and few-shot performance, often matching or surpassing full-data baselines with only 5% of the data. These findings offer a practical baseline and a scalable platform for future research and real-world TSF tasks, enabling reproducible, component-wise analysis of LTSM training strategies.

Abstract

Time Series Forecasting (TSF) has long been a challenge in time series analysis. Inspired by the success of Large Language Models (LLMs), researchers are now developing Large Time Series Models (LTSMs)-universal transformer-based models that use autoregressive prediction-to improve TSF. However, training LTSMs on heterogeneous time series data poses unique challenges, including diverse frequencies, dimensions, and patterns across datasets. Recent endeavors have studied and evaluated various design choices aimed at enhancing LTSM training and generalization capabilities. However, these design choices are typically studied and evaluated in isolation and are not benchmarked collectively. In this work, we introduce LTSM-Bundle, a comprehensive toolbox, and benchmark for training LTSMs, spanning pre-processing techniques, model configurations, and dataset configuration. It modularized and benchmarked LTSMs from multiple dimensions, encompassing prompting strategies, tokenization approaches, training paradigms, base model selection, data quantity, and dataset diversity. Furthermore, we combine the most effective design choices identified in our study. Empirical results demonstrate that this combination achieves superior zero-shot and few-shot performances compared to state-of-the-art LTSMs and traditional TSF methods on benchmark datasets.
Paper Structure (30 sections, 3 figures, 21 tables)

This paper contains 30 sections, 3 figures, 21 tables.

Table of Contents

  1. Introduction
  2. Notations and Problem Formulation
  3. LTSM-Bundle Package
  4. Package Design
  5. Package Interface
  6. Benchmarking LTSM Training
  7. Pre-processing: Instruction Prompts
  8. Pre-processing: Tokenizations
  9. Model Configuration: Training Paradigm
  10. Model Configuration: Base Model Selection
  11. Dataset configuration: Quantity
  12. Dataset Configuration: Diversity
  13. Comparison with Baselines
  14. Experimental Settings
  15. Zero-shot and Few-shot Results
  16. ...and 15 more sections

Figures (3)

  • Figure 1: An overview of important design choices in training LTSM-Bundle framework.
  • Figure 2: An example of data point information visualization in LTSM-Bundle. The information feature provides detailed insights into specific data points on a chart, typically displayed as a side panel.
  • Figure 3: Time series prompt generation process based on the given training time series data.