Answers are grounded in this paper and cite their sources.
Multivariate time-series forecasting requires models to represent state transitions, temporal dependence, and distributional drift without conflating their roles. We propose VTTF-Net, a role-separated multiresolution framework organized by resolution before modality. At each shared temporal resolution, a transition-field branch encodes regularized variable-wise Markov transition fields, a local–global branch combines dilated convolution with self-attention, and a trend-aware branch couples adaptive smoothing with an autoregressive anchor. Each modality first learns its own scale allocation; an input-conditioned gate then routes the three aggregated representations to a factorized horizon–variable decoder. On Traffic, ETTh1, and Exchange-Rate, VTTF-Net achieved the lowest MSE and MAE among the protocol-aligned evaluated models in all 12 dataset–horizon settings. The unweighted mean of the 12 five-seed setting means was 0.383/0.357 for MSE/MAE, compared with 0.392/0.368 for a fixed reference using PatchTST on Traffic and ETTh1 and DLinear on Exchange-Rate. Controlled fusion-order and component analyses support the proposed organization under the evaluated protocol. Synthetic perturbations, modality-routing coefficients, and scale-allocation coefficients are reported only as diagnostic sensitivity and descriptive internal allocation statistics, not as evidence of causal importance or robustness to naturally occurring distribution shifts.
No takes yet. Share an insight, caveat, or question.
Lei et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: