A new forecasting model skips the random noise most generative predictors start from, and plugs in patterns it already learned instead.
The approach, called ProtoFlow, targets multivariate time series forecasting, the kind of multi-signal data used in finance, energy grids, or sensor networks. Existing fast forecasters compress series into a discrete vocabulary of patterns using vector quantization (VQ), then generate predictions token by token in autoregressive fashion. That step-by-step generation causes exposure bias: the model trains on real past tokens but at inference has to feed its own, sometimes wrong, predictions back into itself, and errors compound. Flow matching offered a non-autoregressive fix, generating a full forecast in one pass, but it typically starts from generic Gaussian noise, same as diffusion models. ProtoFlow's twist is to start from the VQ codebook's own learned prototypes instead of noise, then use a DiT-based rectified flow to transport those prototypes to the actual future values, conditioned on history.
That matters because exposure bias is one of the stubborn reasons autoregressive forecasters degrade over time, and generic noise priors in flow matching throw away information the model already has about what plausible outputs look like. Swapping in a learned prior is a cheap way to give the model a head start, and the paper reports faster training convergence alongside better accuracy on benchmark datasets.
Worth noting: this is an arXiv preprint, not yet peer-reviewed, and the gains are shown only on standard research benchmarks, not live production data. "Sensible idea that works on benchmarks" and "better than everything already running in a trading or grid-ops pipeline" are different claims.