A new generative model skips raw time steps and builds synthetic time series out of wavelets instead.
The approach, called wavelet flow matching, decomposes a time series into multilevel wavelet coefficients before generating anything. Coarse structure and fine detail get represented at separate scales, and because those scales naturally have different variances, the model gets a coarse-to-fine generation process for free, without needing a hand-built schedule for it. Since wavelet transforms work channel by channel, the researchers paired the method with a channel-token transformer whose attention handles the cross-channel dependencies the wavelet step can't see. They tested it across seven benchmark datasets and four sequence lengths.
Synthetic time series matter because real time series, sensor logs, financial data, health records, are often too sensitive or too scarce to share or train on directly. The method's biggest, most consistent gains showed up in Context-FID and discriminative score, both measures of how easily a classifier can tell fake data from real. That is the metric pair that actually matters for privacy-preserving data sharing and augmentation.
But read the results paragraph closely: the model was "best or tied" on a majority of dataset-metric combinations, not all of them. That is a real result, not a clean sweep, and worth noting given how often papers in this space lean on vaguer language to describe a mixed scoreboard.