A new arXiv paper introduces TimeThink, a technique intended to help time-series multimodal large language models reason more compositionally. The authors note that such models often struggle to capture dynamic temporal patterns when answering questions. The work focuses on eliciting stronger reasoning behavior from these models rather than treating forecasting as pure pattern matching.
Researchers present a new benchmark for time series forecasting that moves beyond purely numeric data to include contextual signals such as text and other modalities. They argue existing multimodal benchmarks are limited in both the volume of data and the range of context they cover. The work is posted as an arXiv preprint in cs.AI and cs.LG.
A new arXiv paper investigates how pretrained language models can be adapted for univariate time-series forecasting using parameter-efficient transfer learning. The authors focus on identifying which design decisions matter most for effective transfer between text and numerical sequences. The work is a cross-listed submission to arXiv's machine learning category.
A new arXiv survey examines how probabilistic forecasting methods have developed across time series and spatiotemporal research. The authors argue the field has grown fragmented, with statistical modeling, machine learning, and deep generative approaches advancing largely in parallel. The paper aims to organize this landscape and connect the differing methodological traditions used for forecasting under uncertainty.
A revised arXiv paper introduces WaveHiTS, a wavelet-enhanced hierarchical time series model designed for short-term wind direction forecasting in eastern Inner Mongolia. The approach targets common difficulties in directional data, including circular values, multi-step error accumulation, and complex meteorological interactions. It was cross-listed on arXiv's cs.LG and cs.AI categories.
Google Research has introduced TimesFM-3, a model designed to forecast time series by combining historical data with related signals and known upcoming events such as promotions or weather. It generates all future time points at once rather than predicting one step at a time, and has 330 million parameters.
A new arXiv paper examines how to measure whether text annotations actually improve a forecasting model's predictions when paired with time-series data. The authors propose benchmarking information-theoretic metrics intended to quantify how much a given text input contributes to forecast accuracy. The study targets multimodal forecasting pipelines that blend numerical series with textual context.
A new preprint introduces a decomposition method that extends feature-attribution explanations from single-number predictions to functional or multivariate outputs, such as demand forecasts that vary over time. The approach works in a Hilbert space so that the influence of each input feature can be separated across the whole output trajectory rather than summarized by one score. The authors position it as a general framework for settings where model predictions are curves or vectors instead of scalars.
Researchers introduce KairosAgent, an agentic approach to forecasting time series that span multiple domains and data types. The framework is designed to pair careful numerical analysis with semantic interpretation and to fuse information from different modalities. The paper is a revised version (v2) recently updated on arXiv.
A new study investigates how far into the future autoregressive models should be trained when forecasting dynamical systems. The authors identify a trade-off between predictive performance and learnability as the training horizon grows, suggesting an optimal horizon exists. The findings offer practical guidance for selecting prediction horizons during model training.
Researchers introduce SurF, a generative model designed for multivariate event streams that are sampled at irregular intervals. The work argues that tokenization-based approaches struggle when the gaps between events span orders of magnitude, and proposes an alternative suited to such data. The paper is a revised arXiv submission in machine learning.
IBM has published a revised edition of its Granite time series foundation model, PatchTST-FM-r2, on Hugging Face. Built on the Patch Time Series Transformer architecture, the model targets forecasting workloads and is offered under licensing terms that permit commercial use.
WHY IT MATTERS ↘A commercially licensed, openly available forecasting foundation model lowers the cost and legal friction of adopting time-series AI in production, an area where enterprises have mostly faced proprietary or research-restricted options. The rapid revision also signals that vendors are competing on maintained, enterprise-ready time-series models rather than one-off releases.