COMP8240 LTSF-Linear Reproduction and Generalisation

This repository supports a reproduction and extension of:

Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu.
"Are Transformers Effective for Time Series Forecasting?"
AAAI 2023.

Project Objective

This project investigates:

1. whether selected LTSF-Linear results can be reproduced using the authors' released code and original data;
2. whether the reported findings generalise to new public time-series datasets; and
3. under which controlled data conditions the relative performance of Linear, NLinear, DLinear, and selected Transformer baselines changes.

Current Status

• Official LTSF-Linear repository obtained and executed
• Canonical ETTh1 dataset obtained
• DLinear executed on ETTh1 for forecast horizons 96, 192, 336, and 720
• Linear reproduced under matched ETTh1 settings
• NLinear reproduced under matched ETTh1 settings
• Primary external dataset prepared
• Controlled synthetic dataset generated
• New-data experiments completed

Repository Structure

• src/ - preprocessing, generation, evaluation, and experiment scripts
• data/ - data documentation and generated datasets
• experiments/ - experiment-specific logs and outputs
• results/ - numerical comparisons and figures
• metadata/ - metadata describing constructed datasets
• environment/ - software and reproducibility information

Original Implementation

Official LTSF-Linear repository:
https://github.com/cure-lab/LTSF-Linear

Reproducibility

Experiment settings, software versions, preprocessing decisions, generated-data parameters, and random seeds will be recorded so that the reported results can be reconstructed.