TimePre: Bridging Accuracy, Efficiency, and Stability in Probabilistic Time-Series Forecasting

TL;DR

TimePre tackles the longstanding trade-off between accuracy, efficiency, and stability in probabilistic time-series forecasting by unifying MCL with fast, MLP-backed backbones. The key innovation, Stabilized Instance Normalization (SIN), per-instance channel-wise pre-conditioning, stabilizes training and preserves hypothesis diversity under the WTA objective. Coupled with a linear temporal encoder and a multi-hypothesis decoder, TimePre achieves state-of-the-art probabilistic metrics while offering rapid, single-pass inference across six real-world datasets. The approach not only outperforms diffusion- and autoregressive-based methods in speed but also mitigates hypothesis collapse, providing a robust, scalable framework for uncertainty-aware forecasting.

Abstract

Probabilistic Time-Series Forecasting (PTSF) is critical for uncertainty-aware decision making, but existing generative models, such as diffusion-based approaches, are computationally prohibitive due to expensive iterative sampling. Non-sampling frameworks like Multiple Choice Learning (MCL) offer an efficient alternative, but suffer from severe training instability and hypothesis collapse, which has historically hindered their performance. This problem is dramatically exacerbated when attempting to combine them with modern, efficient MLP-based backbones. To resolve this fundamental incompatibility, we propose TimePre, a novel framework that successfully unifies the efficiency of MLP-based models with the distributional flexibility of the MCL paradigm. The core of our solution is Stabilized Instance Normalization (SIN), a novel normalization layer that explicitly remedies this incompatibility. SIN stabilizes the hybrid architecture by correcting channel-wise statistical shifts, definitively resolving the catastrophic hypothesis collapse. Extensive experiments on six benchmark datasets demonstrate that TimePre achieves new state-of-the-art accuracy on key probabilistic metrics. Critically, TimePre achieves inference speeds orders of magnitude faster than sampling-based models and, unlike prior MCL work, demonstrates stable performance scaling. It thus bridges the long-standing gap between accuracy, efficiency, and stability in probabilistic forecasting.

Figures (5)

• Figure 1: Top: Model performance comparison on the Distortion metric across six real-world benchmark datasets. Bottom: Visualization on the Electricity dataset comparing MeanScaler and SIN within the TimePre framework, illustrating SIN’s ability to place trajectories on the correct scale and support more accurate predictions.
• Figure 2: Computation–performance trade-off on the Exchange dataset under 16 hypotheses. The x-axis shows inference time (log scale) and the y-axis shows Distortion (lower is better). Circle size indicates FLOPs, and annotations denote each model’s FLOPs and inference time.
• Figure 3: Qualitative forecasting results on five public datasets, comparing three models that adopt the multi-hypothesis paradigm: TimeMCL (R.), TimeMCL (A.), and TimePre, under 16 hypotheses.
• Figure 4: Forecasting comparison across normalization layers on the Electricity dataset. LayerNorm and GroupNorm are shown with a wider y-axis range $[-200, 400]$ for visibility, while our method and BatchNorm use $[0, 200]$. Our normalization yields stable trajectories, whereas the others suffer from noticeable scale distortion.
• Figure 5: A case visualization on the Electricity dataset comparing the covariance structures of the original data, TimeMCL with mean scaling, and TimePre with Stabilized Instance Normalization.