Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TimeFound: A Foundation Model for Time Series Forecasting

Congxi Xiao, Jingbo Zhou, Yixiong Xiao, Xinjiang Lu, Le Zhang, Hui Xiong

TL;DR

TimeFound addresses the challenge of zero-shot time series forecasting across diverse domains by introducing a transformer-based foundation model with multi-resolution patching. The encoder-decoder framework captures multi-scale temporal patterns, while patch-wise projections and a patch-aware fusion enable flexible cross-domain generalization. Trained on a large heterogeneous corpus with two model sizes, TimeFound achieves competitive or superior zero-shot performance relative to state-of-the-art foundation models and shows strong long-horizon robustness, highlighting the value of patch-based tokenization for time series tasks. The approach offers a scalable, domain-agnostic solution for forecasting in settings with limited or no task-specific labeled data.

Abstract

We present TimeFound, an encoder-decoder transformer-based time series foundation model for out-of-the-box zero-shot forecasting. To handle time series data from various domains, TimeFound employs a multi-resolution patching strategy to capture complex temporal patterns at multiple scales. We pre-train our model with two sizes (200M and 710M parameters) on a large time-series corpus comprising both real-world and synthetic datasets. Over a collection of unseen datasets across diverse domains and forecasting horizons, our empirical evaluations suggest that TimeFound can achieve superior or competitive zero-shot forecasting performance, compared to state-of-the-art time series foundation models.

TimeFound: A Foundation Model for Time Series Forecasting

TL;DR

TimeFound addresses the challenge of zero-shot time series forecasting across diverse domains by introducing a transformer-based foundation model with multi-resolution patching. The encoder-decoder framework captures multi-scale temporal patterns, while patch-wise projections and a patch-aware fusion enable flexible cross-domain generalization. Trained on a large heterogeneous corpus with two model sizes, TimeFound achieves competitive or superior zero-shot performance relative to state-of-the-art foundation models and shows strong long-horizon robustness, highlighting the value of patch-based tokenization for time series tasks. The approach offers a scalable, domain-agnostic solution for forecasting in settings with limited or no task-specific labeled data.

Abstract

We present TimeFound, an encoder-decoder transformer-based time series foundation model for out-of-the-box zero-shot forecasting. To handle time series data from various domains, TimeFound employs a multi-resolution patching strategy to capture complex temporal patterns at multiple scales. We pre-train our model with two sizes (200M and 710M parameters) on a large time-series corpus comprising both real-world and synthetic datasets. Over a collection of unseen datasets across diverse domains and forecasting horizons, our empirical evaluations suggest that TimeFound can achieve superior or competitive zero-shot forecasting performance, compared to state-of-the-art time series foundation models.

Paper Structure

This paper contains 26 sections, 7 equations, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Time Series Forecasting
  4. Time Series Foundation Models
  5. Problem Definition
  6. Method
  7. Input Module
  8. Normalization
  9. Multi-resolution Patching
  10. Projection
  11. Transformer Blocks
  12. Output Module
  13. Training Objective
  14. Forecasting
  15. Experiments
  16. ...and 11 more sections

Figures (1)

  • Figure 1: Illustration of TimeFound model. In (a), for simple illustration, we assume $K=2$ in multi-resolution patching method and it divides normalized time series using twp patch sizes $P_i$ and $P_2$. (b) and (c) show the detailed implementation of the projector in Input Module and the prediction head in Output Module respectively. (d) presents the model's different behaviors during pre-training and forecasting.