Timestral

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.

Authors

Publication Details

Journal
arXiv (Cornell University)
Published
2026-10-27
DOI
https://doi.org/10.48550/arxiv.2610.37585
Primary Topic
Time Series Analysis and Forecasting
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

Timestral

Christophe Renaudin, André Jonasson, Gauthier Guinet, Arjun Majumdar et al.
arXiv (Cornell University)
Time Series Analysis and Forecasting
preprint

Timestral

Christophe Renaudin, André Jonasson, Gauthier Guinet, Arjun Majumdar, Fabian Schlager, Georgii Novikov, Graeme Nail, Andrew Zhao, Chris Bamford, Boris van Breugel, Amélie Héliou, Filippo Vicentini, Błażej Osiński, Étienne Goffinet, Gaurav Mahajan, Barna Pásztor, Bhargavi Bhaskar, Fabian Paischer, Etienne Duchesne, Florian Wiesner, Emmanuel Gottlob, Alan Jeffares, Arthur Fournier, Benjamin Tibi, Adrian Valente, Corentin Sautier, Baptiste Rozière, Elizaveta Demyanenko, Dylan Rubini, Richard Kurle, Benjamin Holzschuh, Filip Szatkowski, Federico Baldassarre, Ayşenur Karaduman, Andrew Bai, Alexandre Sablayrolles, Elliot Chane-Sane, Diego de Las Casas, Colin Bonatti, Binbing Hou, Daniel Marczak, Yassir Bendou, Abhijeet Somani, Artjom Joosen, Bernhard Stadlbauer, Giada Pistilli, Erik Aas, Augustin Lombard, Abdelaziz Bounhar, Adrien Petralia, Gonzague de Carpentier, Adrien Sadé, Aditi Kabra, Amos You, Cyprien Courtot, Chen-Yo Sun, Cristiana-Diana Diaconu, Anmol Agarwal, Guillaume Kunsch, Aleksandr Timashov, Baudouin De Monicault, Christian Wallenwein, Emilien Fugier, Corentin Barreau, Gaëtan Ecrepont, Alexandre Gavaudan, Clémence Lanfranchi, Alexandre Cahill, Albert Jiang, Benno Stein, Andreas Fürst, Andy Lo, Dominik Nuß, Darius Dabert, Antonia Calvi, Baptiste Bout, Gabrielle Berrada, Genevieve Hayes, Angele Lenglemetz, Charlotte Cronjäger, Alexandros Sanchez Vassopoulos, Gaspard Blanchet, Arata Suzuki, Dzmitry Soupel, Gaspard Donada-Vidal, Andrew J. Ehrenberg, Aylin Guliz Akkus, Chaoran Yu, Avinash Sooriyarachchi, Camille Le Scao, Alexandre Laval, Faruk Ahmed, Guillaume Breton, Georgios Christoglou, Ashwani Kumar, Yassine El Ouahidi, Daniel Lopes, Alberto Bietti, Félix Rosseeuw, Ahmed Djermani
preprint en

Abstract

We introduce Timestral, a family of four time series foundation models for probabilistic forecasting, with 5M, 28M, 123M, and 209M parameters. Timestral sets a new state of the art on GIFT-Eval, TIME, and fev-bench, where Timestral-209M and Timestral-123M take the top two places on every metric. We achieve this with three ingredients: Decoupled Horizon Forecasting (DHF), which separates prediction tokens from the causally encoded context, so that longer horizons are trained without masking context and each observed value is both a forecast target and context for later forecasts; a covariate framework that natively supports known-future covariates within a causal architecture; and a two-stage pipeline that pretrains exclusively on a diverse, large-scale synthetic corpus before post-training on a mix of real and synthetic data with extended context. This report describes the model design and its benchmark results. Weights for all four models are available on Hugging Face and inference code on GitHub.

arXiv (Cornell University)
Time Series Analysis and Forecasting
DOIOpen Access PDFOpenAlex
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon