•
Zeng, Chen, Zhang and Xu (2022) ask whether Transformers help long-term time-series forecasting. One-layer LTSF-Linear (DLinear / NLinear) beats Informer, Autoformer, FEDformer on nine benchmarks, often by 20–50% MSE. Plain-language intro. Not investment advice.
Takeaway. Self-attention is permutation-invariant. On many long-horizon series the “order” is the signal, so a one-layer linear map from look-back (L) to horizon (T) can beat Informer / Autoformer / FEDformer. That is LTSF-Linear (vanilla Linear, DLinear, NLinear).
Are Transformers Effective for Time Series Forecasting? (Ailing Zeng, Muxi Chen, Lei Zhang, Qiang Xu, 2022; arXiv:2205.13504). CUHK / IDEA. AAAI 2023. Code: cure-lab/LTSF-Linear.
Fit for learners: Simple linear baselines that beat many Transformers — humility check before you overfit a trading net.
This Hub note is a plain-language intro. Not investment advice.
2020–2022 LTSF papers stacked sparse attention, frequency blocks, and seasonal-trend gadgets. Their non-Transformer baselines were often iterated multi-step (IMS) models that accumulate error. The authors switch the comparison to direct multi-step (DMS) and ask: does attention still win?
A baseline, not a claim that attention is useless everywhere. PatchTST and later models still exist — start here so you do not skip the dumb model.
Before you pip install a 14M-parameter TS Transformer for nightly bars, fit one nn.Linear(L, T) (or two, after a rolling mean). If it wins, your net was fitting noise.
Tiny evening exercise: with NumPy, W @ x for x of length 48 → 12-step forecast. Then subtract x[-1], predict, add back (NLinear). Compare MSE on a random-walk vs a sine. Learning only.
Landing: arXiv abs
Not investment advice. This Hub note is educational only. Linear beating Transformer on ETT/electricity is not a licence to trade FX or equities with nn.Linear.
Single course CTA: SoloQuant Academy’s Introduction to Python Programming—matrix multiply and broadcasting are exactly DLinear’s (W X).