TimeSSM-10M: A Small Forecaster That Chooses Its Sampling Rate

TimeSSM-10M is a zero-shot probabilistic forecaster with 9.98 million parameters. Its trunk reads patches of 16steps through four pairs of a diagonal state space layer (S4D) and a causal attention layer, run twice with the sameweights. Its first and last layer is RateIN: the context is resampled to a few sampling rates, the trunk forecasts ateach of them, and the forecasts are brought back to the native rate and mixed. RateIN has no learned weight; itsmixture is fitted at inference on the past of the series. A single scalar tells the trunk which rate it is reading, byscaling the time step of the state space layers and the positions of the attention. On GIFT-Eval the model scoresa CRPS of 0.4763 and a MASE of 0.6948, relative to seasonal naive. In the official results folder on 12 November2026 this places it 24th of 88 foundation models, ahead of Chronos-2 (120M parameters), TimesFM-2.5 (200M)and the zero-shot TiRex-2 (82.5M), and third of the models under 10 million parameters, behind PDM-Femto(6M, served by an API) and TimeH2O-v1-tiny (5.8M, non-commercial weights). It is at its best on series sampledevery 5 to 15 minutes and on economic and cloud-operations data, and at its weakest on daily and weekly series.The model was trained in 31 hours on one GPU, on 13.3 billion points from which every series found in GIFT-Evalwas removed, and selected on held-out data only.Code: https://github.com/IUseAMouse/TimeSSMWeights: https://huggingface.co/IUseAMouse/timessm-10m

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24202828
Primary Topic
Time Series Analysis and Forecasting
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

TimeSSM-10M: A Small Forecaster That Chooses Its Sampling Rate

Yvann Vincent
1 links
Zenodo (CERN European Organization for Nuclear Research)
Time Series Analysis and Forecasting
preprint

TimeSSM-10M: A Small Forecaster That Chooses Its Sampling Rate

Yvann Vincent
preprint en

Abstract

TimeSSM-10M is a zero-shot probabilistic forecaster with 9.98 million parameters. Its trunk reads patches of 16steps through four pairs of a diagonal state space layer (S4D) and a causal attention layer, run twice with the sameweights. Its first and last layer is RateIN: the context is resampled to a few sampling rates, the trunk forecasts ateach of them, and the forecasts are brought back to the native rate and mixed. RateIN has no learned weight; itsmixture is fitted at inference on the past of the series. A single scalar tells the trunk which rate it is reading, byscaling the time step of the state space layers and the positions of the attention. On GIFT-Eval the model scoresa CRPS of 0.4763 and a MASE of 0.6948, relative to seasonal naive. In the official results folder on 12 November2026 this places it 24th of 88 foundation models, ahead of Chronos-2 (120M parameters), TimesFM-2.5 (200M)and the zero-shot TiRex-2 (82.5M), and third of the models under 10 million parameters, behind PDM-Femto(6M, served by an API) and TimeH2O-v1-tiny (5.8M, non-commercial weights). It is at its best on series sampledevery 5 to 15 minutes and on economic and cloud-operations data, and at its weakest on daily and weekly series.The model was trained in 31 hours on one GPU, on 13.3 billion points from which every series found in GIFT-Evalwas removed, and selected on held-out data only.Code: https://github.com/IUseAMouse/TimeSSMWeights: https://huggingface.co/IUseAMouse/timessm-10m

Zenodo (CERN European Organization for Nuclear Research)
Time Series Analysis and Forecasting
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