TimePre: Bridging Accuracy, Efficiency, and Stability in Probabilistic Time-Series Forecasting
Lingyu Jiang, Lingyu Xu, Peiran Li, Dengzhe Hou, Qianwen Ge, Dingyi Zhuang, Shuo Xing, Wenjing Chen, Xiangbo Gao, Ting-Hsuan Chen, Xueying Zhan, Xin Zhang, Ziming Zhang, Zhengzhong Tu, Michael Zielewski, Kazunori Yamada, Fangzhou Lin
cs.LG, cs.CV
2025-11-24
TimePre uses 5% trimmed instance norm to stabilize MCL on a linear backbone, winning Distortion on all six GluonTS sets and cutting Electricity error 38.8% at 0.03s/batch.
Probabilistic time-series forecasting wants a distribution over future trajectories, not a point. TimeGrad, TempFlow, and TACTiS-2 get there by iterative sampling, which is accurate and slow. TimeMCL instead emits K discrete hypotheses in one forward pass and trains them with winner-takes-all (WTA): only the head with the smallest loss on a given sample gets the main gradient.
Linear and MLP backbones already beat many Transformers on deterministic long-horizon forecasting (DLinear, TiDE, TimeMixer). Wiring that efficiency into MCL collapses the hypothesis set: of 16 heads, one or two keep learning and the rest starve. Isolation experiments pin the failure on the interaction of channel-wise scale imbalance, naive normalization, and hard competition. A linear layer has no LSTM-style manifold to damp scale gaps, so a head that happens to sit closer to a high-magnitude channel keeps winning.
TimePre is three stages: Stabilized Instance Normalization (SIN), a per-channel linear encoder, and K parallel heads.
SIN is RevIN with trimmed statistics. For each instance and each channel it drops the top and bottom p=5% of the look-back window, estimates mean and variance on the remainder, standardizes, and inverts after the forecast. Electricity and Wiki mix channels whose units differ by orders of magnitude; a single spike can wreck ordinary mean and variance. Under WTA, that scale error decides the arg min, so gradients lock onto a few heads.
The encoder is one linear map from look-back length L to horizon H, channel by channel. Each decoder head outputs a trajectory plus a sigmoid confidence score. Training uses relaxed WTA with ε=0.3 (the winner takes 1-ε, the rest share ε) plus a β=0.5 binary cross-entropy that asks the confidence head to name the winner. Inference is one forward pass, K=16, no autoregression and no sampling. The same hyperparameter set is used on all six datasets: Adam at 10^{-3}, batch 200, 200 epochs, patience 10.
Six GluonTS datasets, K=16, five random seeds. Distortion is Euclidean distance from the target trajectory to the nearest hypothesis (lower is better).
| Method | Electricity | Exchange | Solar |
| TimeMCL (R.) | 12693±1772 | 0.0380±0.0025 | 292.15±11.68 |
| TimeMCL (A.) | 10335±767 | 0.0443±0.0051 | 308.16±14.87 |
| TimeGrad | 14255±1682 | 0.0576±0.0090 | 406.91±16.08 |
| TimePre | 7774±203 | 0.0275±0.0004 | 267.06±1.55 |
On Electricity, TimePre is 38.8% below TimeMCL (R.); on Exchange, 27.6% below. Traffic 0.68, Taxi 182.27, Wiki 263492: first on Distortion on all six. The Wiki margin over TACTiS-2 (263975) is tiny. Run-to-run spread shrinks with it: ±203 versus ±1772 on Electricity.
On CRPS-Sum (paper values ×100) TimePre leads on Electricity (3.15), Exchange (0.72), and Wiki (6.14); second on Solar (39.79) and Taxi (21.72); well behind TimeGrad on Traffic (11.81 vs 5.89). Sampling still wins full-distribution calibration. It also costs 64.2s per batch on Exchange against TimePre's 0.03s and 4.28×10^4 FLOPs (TimeGrad: 1.91×10^8). DeepAR uses even fewer FLOPs (2.90×10^4) and posts Distortion 184424.
Normalization on Electricity: BatchNorm 12447, LayerNorm 9453, RevIN 8106, SIN 7774. Head utilization is harsher: mean scaling, GroupNorm, and InstanceNorm keep 1 active head, RevIN 5, SIN 6. On Traffic, SIN lifts winner entropy by 1.0 bit over RevIN. A linear backbone with mean scaling already hits Distortion 68670 at K=1, before any WTA; SIN brings that to 9626. An LSTM with mean scaling survives K=1 at 18000. Heavier backbones do not help: a TiDE variant blows Solar Distortion to 478.26.
This is for people who need probabilistic forecasts and cannot wait on diffusion sampling. Load, traffic, and FX sit in that bucket. The recipe is close to DLinear plus trimmed RevIN plus multiple heads, one hyperparameter set for all six datasets.
Call it incremental. The authors do: SIN is not a new primitive, it is trimmed statistics inside reversible instance normalization. The transferable claim is narrower. Linear MCL collapses because of scale; fix the normalizer first.
Three author-stated limits. The SIN-over-RevIN gain is dataset-dependent; on Wiki, winner entropy is slightly lower. The confidence head is trained with winner-identification cross-entropy, not a proper scoring rule: Electricity AUROC is 0.410 and argmax accuracy 0.029, worse than the majority baseline 0.286. Evaluation covers six fixed-horizon GluonTS sets, and the channel-independent design drops cross-variable dependence.
The SIN-on-TimeMCL LSTM comparison is a single-seed diagnostic; it even beats five-seed TimePre on Taxi and Wiki, so that contrast is weak. Traffic CRPS-Sum remains far behind TimeGrad: K discrete trajectories do not cover a full conditional law. Solar far-horizon degradation hits 5.42×, attributed to the day-night cycle; other strongly diurnal series were not retested.