Lightweight Mixture-of-Experts Model Shows Strong Time-Series Forecasting Performance

Global: Lightweight Mixture-of-Experts Model Shows Strong Time-Series Forecasting Performance

A study released on arXiv on 18 September 2025 and updated on 27 January 2026 introduces a new model for time‑series forecasting that aims to combine high accuracy with low computational demand. The paper, authored by Liran Nochumsohn, Raz Marshanski, Hedi Zisling, and Omri Azencot, describes the Super‑Linear architecture, a pretrained mixture‑of‑experts system designed for zero‑shot forecasting across diverse domains.

Motivation and Context

Time‑series forecasting underpins operations in energy grids, financial markets, healthcare logistics, and other sectors that rely on accurate predictions. Recent large‑scale pretrained models such as Chronos and Time‑MoE have demonstrated strong zero‑shot capabilities but require substantial hardware resources, limiting their deployment in edge or resource‑constrained environments.

Model Design

Super‑Linear replaces deep neural stacks with a collection of linear experts, each specialized for a particular frequency regime. The experts are trained on resampled data that spans multiple temporal frequencies, and a lightweight spectral gating mechanism selects the most relevant experts for a given input sequence. This design enables the system to remain lightweight while preserving the ability to capture complex temporal patterns.

Benchmark Performance

According to the authors, the model achieves competitive results on standard forecasting benchmarks, matching or surpassing the accuracy of larger MoE models on several datasets. The reported experiments highlight improvements in mean absolute error and root‑mean‑square error while using a fraction of the parameters and inference time of comparable baselines.

Efficiency, Robustness, and Interpretability

The authors emphasize three practical advantages: (1) reduced computational cost, measured by lower FLOPs and memory footprint; (2) robustness to varying sampling rates, owing to the frequency‑aware expert training; and (3) enhanced interpretability, as the spectral gate provides insight into which frequency components drive each prediction.

Availability and Future Work

The implementation, including model weights and training scripts, has been released publicly on GitHub (link provided in the paper). The authors suggest that future extensions could explore adaptive expert addition and integration with domain‑specific priors to further improve performance on specialized forecasting tasks.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.