A new preprint proposes a neural network that learns wavelet transforms instead of using fixed ones, and claims big gains on time series tasks.
Researchers describe the method, called m-WCN, in a paper posted September 25, 2026 on arXiv (2609.29317, https://arxiv.org/abs/2609.29317). The system replaces a classical multi-wavelet transform called GHM with trainable convolutional layers, aiming to capture both time-domain patterns and frequency-domain structure in one model instead of treating them separately. Built on that base, the team adds two task-specific variants: one for classification and one for forecasting. On 64 datasets from the UCR archive and seven public forecasting benchmarks, the authors report average improvements of 19.97% for classification and 19.92% for forecasting over baseline models.
Time series research has long split into frequency-focused methods, like Fourier or wavelet filters, and time-focused methods, like recurrent or attention models tracking sequences directly. m-WCN's pitch is that making the wavelet transform itself trainable, rather than fixed, lets one model do both jobs while staying interpretable at multiple resolutions. If those numbers hold up outside the paper's own test set, that's a meaningful architectural shift for anyone building forecasting or classification tools on financial, medical, or sensor data.
Those improvement percentages come from the authors comparing against their own choice of baselines, so treat them as a starting point for scrutiny, not a settled result.