TimeMosaic

This is the official implementation of our paper: TimeMosaic: Information-Density Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding

This repository contains the official implementation of TimeMosaic, a novel framework for multivariate time series forecasting. It dynamically partitions input sequences into variable-length patches based on local temporal information density, and performs segment-wise forecasting through prompt-guided multi-task learning.

Usage

1. Install Python 3.10. For convenience, execute the following command.

pip install torch==2.4.1 torchvision==0.19.1  --index-url https://download.pytorch.org/whl/cu121

pip install transformers==4.40.1 accelerate==1.10.1 lightning==2.3.3 \
    gluonts==0.14.4 numpy==1.26.4 pandas==2.1.4 \
    -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install scikit-learn -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install reformer-pytorch -i https://pypi.tuna.tsinghua.edu.cn/simple
pip install uni2ts
pip install chronos-forecasting
pip install -U "ml_dtypes==0.4.0" "jax[cpu]==0.4.28" "jaxtyping==0.2.28"
pip install -U "accelerate==0.31.0"
pip install transformers==4.40.1

2. Prepare Data. You can obtain the well pre-processed datasets from Time-Series-Library. [Google Drive], Then place the downloaded data in the folder./dataset. Here is a summary of supported datasets.

Run the following commands from the BasicTS root directory to download BLAST from Hugging Face:

huggingface-cli download ZezhiShao/BLAST \
  --repo-type dataset \
  --local-dir ./dataset/BLAST

After the download finishes, the data will be under dataset/BLAST. The weights of TimeMosaic trained on large-scale datasets can be accessed here: Google Drive Link.

3. Train and evaluate model. We provide the experiment scripts for all benchmarks under the folder ./scripts/. You can reproduce the experiment results as the following examples:

# Please comment the code for saving weights.
bash scripts/fair.sh
bash scripts/fair_x.sh
bash scripts/search.sh

# bash scripts/TimeMosaic/ETTh1.sh

Result

Results under the hyperparameter search setting described in Paper Appendix Section I.

Zero-shot forecasting results on two datasets.

Included Models

Our benchmark integrates over 20+ state-of-the-art time series forecasting models spanning various design paradigms, including Transformer-based, MLP-based, and hybrid architectures. All models are implemented under a unified codebase with consistent settings to ensure fair and reproducible comparisons.

The following models have been included in our evaluation suite:

• ✅ TimeMosaic (Ours): A novel framework combining adaptive patch granularity and segment-wise prompt tuning.
• ✅ SimpleTM : SimpleTM: A Simple Baseline for Multivariate Time Series Forecasting [ICLR 2025].
• ✅ TimeFilter : TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting [ICML 2025].
• ✅ xPatch : xPatch: Dual-Stream Time Series Forecasting with Exponential Seasonal-Trend Decomposition [AAAI 2025].
• ✅ DUET : DUET: Dual Clustering Enhanced Multivariate Time Series Forecasting [KDD 2025].
• ✅ PathFormer : Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting [ICLR 2024].
• ✅ PatchMLP : Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting [AAAI 2025].
• ✅ iTransformer : iTransformer: Inverted Transformers Are Effective for Time Series Forecasting [ICLR 2024].
• ✅ PatchTST : A Time Series is Worth 64 Words: Long-term Forecasting with Transformers [ICLR 2023].
• ✅ TimesNet : TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis [ICLR 2023].
• ✅ DLinear : Are Transformers Effective for Time Series Forecasting? [AAAI 2023].
• ✅ TimeMixer : TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting [ICLR 2024].
• ✅ MICN : MICN: Multi-scale Local and Global Context Modeling for Long-term Series Forecasting [ICLR 2023].
• ✅ FreTS : Frequency-domain MLPs are More Effective Learners in Time Series Forecasting [NeurIPS 2023].
• ✅ Crossformer : Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting [ICLR 2023].
• ✅ TiDE : Long-term Forecasting with TiDE: Time-series Dense Encoder [arXiv 2023].
• ✅ LightTS : Less Is More: Fast Multivariate Time Series Forecasting with Light Sampling-oriented MLP Structures [arXiv 2022].
• ✅ Autoformer : Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting [NeurIPS 2021].
• ✅ FEDformer : FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting [ICML 2022].
• ✅ Informer : Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting [AAAI 2021].
• ✅ Pyraformer : Pyraformer: Low-Complexity Pyramidal Attention for Long-Range Time Series Modeling and Forecasting [ICLR 2022].
• ✅ Reformer : Reformer: The Efficient Transformer [ICLR 2020].
• ✅ ETSformer : ETSformer: Exponential Smoothing Transformers for Time-series Forecasting [arXiv 2022].

All models are trained under identical lookback lengths and evaluation metrics (MSE/MAE), and the same drop_last=False setting to ensure fairness and reproducibility.

Acknowledgements

We appreciate the following resources a lot for their valuable code and datasets:

• Time-Series-Library (https://github.com/thuml/Time-Series-Library)
• iTransformer (https://github.com/thuml/iTransformer)
• BasicTS (https://github.com/GestaltCogTeam/BasicTS)