// the find
cure-lab/LTSF-Linear
[AAAI-23 Oral] Official implementation of the paper "Are Transformers Effective for Time Series Forecasting?"
A research repo from AAAI 2023 that makes the uncomfortable argument that a single linear layer beats Transformer-based time series forecasting models on standard benchmarks. The three models (Linear, NLinear, DLinear) are genuinely tiny — DLinear does trend/seasonality decomposition with two linear layers. Aimed at researchers and ML practitioners who work on time series and want a strong, embarrassingly simple baseline.
The core claim is backed by rigorous benchmarks across nine datasets and five Transformer variants, making it useful as a baseline sanity check before reaching for anything complex. DLinear's decomposition approach (moving average for trend, remainder for seasonality) is simple enough to understand in five minutes but meaningful enough to matter in practice. Weight visualization actually works as advertised — you can see periodicity patterns directly in the learned weights, which is rare in deep learning models. It's now included in NeuralForecast, so you can use it without touching this repo at all if you don't want research-grade shell scripts.
The repo requires Python 3.6.9, which hit end-of-life in late 2021 — if your environment runs anything modern you'll hit compatibility friction immediately. The code structure is a copy of the Autoformer codebase with minimal cleanup; FEDformer and Pyraformer live in separate subdirectories with their own requirements files, making the overall repo feel like three repos stapled together. Last meaningful commit was early 2024 and the issues tab is full of unanswered questions, so don't expect maintainer support. The benchmark datasets all require manual download from Google Drive and manual placement — no automated data pipeline.