Tag: #time-series23 results
Nature 2026 paper introduces GluFormer, a Transformer-based foundation model for CGM data using self-supervised learning on 10M+ readings from 10k subjects. It outperforms HbA1c for 11-year diabetes/CVD risk prediction and generates realistic glucose curves across populations. Enables personalized diet/treatment forecasting from short-term data.
This research introduces a framework that combines LLM planners with Time-Series Foundation Models to secure industrial IoT. It uses a Counterfactual Physics Injection mechanism to validate LLM actions against physical constraints before execution.
This position paper argues that time series modeling should shift from transformer-based architectures to a dynamical systems perspective. It emphasizes that capturing underlying physical rules and chaotic structures is essential for true out-of-domain generalization.
TimeClaw is a new agentic framework designed to bridge the gap between generalist LLMs and structured time series data. It provides native runtime support, including executable tools and multimodal memory, to enable more accurate temporal reasoning across diverse domains.
GITCO is a novel framework that improves Time Series Foundation Models (TSFMs) by identifying and suppressing 'poisonous' input patches at inference time. It achieves a 1.95% MASE reduction on TimesFM 2.5 without requiring any model weight updates.
Peking University’s Lin Zhouchen team proposes DistDF, a novel loss function using optimal transport to align joint distributions in time series forecasting, addressing MSE's independence bias. It converts conditional distribution matching to joint via Wasserstein distance, enabling reliable estimation. Accepted to ICLR 2026.
A new GPU-accelerated, memory-efficient SoftDTW implementation for PyTorch addresses speed and memory limits in time series alignment. It offers 67x speedup, 98% less GPU memory, and supports long sequences beyond 1024. Includes barycenters for DTW-space averaging with full autograd support.
A developer improved a custom spiking neural network's memory depth on the NARMA-10 benchmark by implementing heterogeneous wire delays. This approach achieved a 15x reduction in arithmetic operations compared to standard continuous neural networks.
TSAuditor is an open-source, lightweight Python tool designed to detect common time-series data issues like chronological breaks, data leakage, and sequential spikes. It provides descriptive feedback and suggested fixes to streamline the EDA process for time-series projects.
A developer built an end-to-end air quality forecasting pipeline using OpenAQ and NASA data. By decoupling prediction horizons and using lag vectors, the model successfully overcame variance issues in chaotic environments.