Mechanism diagram · CNN / TCN
CNN / TCN (DILATED) Input series lookback × N Conv 1×k dilation 1 Conv 1×k dilation 2 Conv 1×k dilation 4 Residual + LayerNorm stacked depth-wise blocks Pool + head Forecast

How it works

TimesNet computes the FFT of the lookback series and selects the top-k dominant frequencies.

For each dominant period p, the 1-D series is reshaped into a 2-D tensor of shape (n_periods, p), where rows are repeats of the same intra-period pattern.

Inception-style 2-D convolutions are applied; period-wise outputs are softmax-weighted by FFT amplitude.

Pros and cons on this universe

Pros

  • Native handling of multi-period structure — useful when daily and weekly cycles co-exist.
  • Strong rank #12 IC on the audited screen.
  • ICLR 2023 paper with solid replication.

Cons / failure modes

  • Fails on non-stationary periods — TwinS critique notes that crypto periods drift.
  • Cluster-redundant with TimeMixer.
  • RMSE-focused literature claim does not transfer 1:1 to directional Sharpe.

References