Introducing TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis for after-work quants

Takeaway. Markets are multi-periodic (intraday + weekly). A 1D CNN/Transformer sees only neighbors. TimesNet FFT-picks periods, folds the series into 2D maps (column = inside one period, row = same clock across periods), and runs 2D Inception kernels. One backbone for forecast / impute / classify / anomaly.

TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, Mingsheng Long, 2022; ICLR 2023; arXiv:2210.02186). Tsinghua School of Software / BNRist.

Fit for learners: Reshape 1D markets into 2D period maps — nice idea for overnight/intraday seasonality projects.

This Hub note is a plain-language intro. Not investment advice.

The problem

RNN forgets long range. TCN is local. Attention on scattered timestamps struggles when several cycles overlap. Autoformer/FEDformer already lean on periodicity in frequency. This paper makes the 2D structure explicit so vision kernels can eat it.

The idea / core intuition

• Period(): FFT amplitudes, average over channels, keep top-k frequencies in 1..floor(T/2). Period length p_i = ceil(T / f_i). k=5 for long forecast, k=3 for impute/classify/anomaly.
• Reshape: pad, then map 1D length-T into a 2D tensor p_i by f_i by C. Columns = intraperiod (short pattern inside a day). Rows = interperiod (how today-9:30 relates to yesterday-9:30). No clear period => frequency 1, intraperiod is just the raw series.
• TimesBlock: residual. Shared Inception (multi-scale 2D conv) on each map, reshape+truncate back to 1D, mix the k branches with softmax of those FFT amplitudes. Parameter count does not grow with k.
• Stack 2 (forecast) or 3 (anomaly) blocks. They also swap Inception for ResNeXt / ConvNeXt / Swin — vision backbones become TS backbones.
• Tasks: long LTSF (ETT, ECL, Traffic, Weather, Exchange, ILI), short M4, imputation (mask 12.5–50%), UEA classification (10 sets), anomaly (SMD/MSL/SMAP/SWaT/PSM reconstruction F1).
• Headline averages: long-term ETTm2 MSE 0.291 / MAE 0.333; Electricity 0.192 / 0.295; M4 weighted SMAPE 11.829 vs N-HiTS 11.927 / N-BEATS 11.851; UEA accuracy 73.6% vs Rocket 72.5%; anomaly avg F1 Inception , ResNeXt .