license: other
license_name: non-commercial-research-evaluation
license_link: https://yhatlabs.com
pipeline_tag: time-series-forecasting
tags:
- time-series
- forecasting
- zero-shot
- probabilistic-forecasting
- foundation-model
- research
model-index:
- name: ChakraTS-Lab
results:
- task:
type: time-series-forecasting
dataset:
name: GIFT-Eval
type: Salesforce/GiftEval
metrics:
- type: MASE
name: MASE (geometric mean, relative to seasonal naive)
value: 0.66
- type: CRPS
name: CRPS (geometric mean, relative to seasonal naive)
value: 0.452
source:
name: GIFT-Eval leaderboard
url: https://huggingface.co/spaces/Salesforce/GIFT-Eval
ChakraTS-Lab
ChakraTS-Lab is the research configuration of ChakraTS, YHat Labs' zero-shot probabilistic forecasting model. It is the same mixture of expert forecasters with models that are not part of YHat-Labs model family. It exists to show what the approach reaches when every strong open model is admitted, and it is available for research and benchmark evaluation only. This repository holds the model card; there are no weights to download.
- Commercial product: yhatlabs/ChakraTS
- Website: https://yhatlabs.com
- Code, benchmark files and notebooks: https://github.com/yhatlabs/yhatlabs
- Leaderboards(coming soon): GIFT-Eval · fev-bench
Access
ChakraTS-Lab is not sold. Benchmark maintainers and researchers who need to reproduce the leaderboard entries can request evaluation access at https://yhatlabs.com. The request and response format is identical to ChakraTS (see its model card); the Lab configuration is served from a separate endpoint.
Evaluation
Leaderboard submissions are not yet published; the numbers below are from our runs of the official protocols.
Official protocols, full benchmarks, official seasonal-naive baselines. ChakraTS-Lab is zero-shot: no benchmark training split, GIFT-Eval's included, is used to fit, tune or select anything, and no benchmark test data is seen at any point. The per-series weights are computed from each series' own past windows at request time.
GIFT-Eval (97 dataset configurations) — not yet published
| Model | MASE | CRPS | Licence |
|---|---|---|---|
| ChakraTS-Lab | 0.660 | 0.452 | research only |
| TimesFM-3 | 0.667 | 0.456 | non-commercial |
| ChakraTS | 0.671 | 0.462 | commercial API |
| Toto-2.0-FnF | 0.676 | 0.463 | |
| T0-beta | 0.687 | 0.474 | Apache-2.0 |
| TiRex-2 | 0.697 | 0.478 | Apache-2.0 |
| Chronos-2 | 0.698 | 0.485 | Apache-2.0 |
| TimesFM-2.5 | 0.705 | 0.490 | Apache-2.0 |
fev-bench (100 tasks) — not yet published
| Model | Skill score |
|---|---|
| ChakraTS-Lab | not evaluated on fev-bench |
ChakraTS (chakra-ts-fev configuration) |
48.1 |
| TimesFM-3 | 48.7 |
| Chronos-2 | 47.3 |
| TimesFM-2.5 | 42.2 (leaderboard figure, after its leakage replacement) |
| T0-beta | 46.7 |
| TiRex-2 | 45.5 |
Per-task result files and the evaluation notebooks are in the GitHub repository.
Approach
ChakraTS-Lab is a mixture of expert forecasters. Each expert produces a full predictive distribution for every series and the distributions are pooled. There is no trained gate and nothing is learned across datasets, which is why the configuration can be evaluated on any benchmark without a leakage question. The difference from ChakraTS is only the expert pool.
YHat Labs has trained its own forecasting models. The current one that powers ChakraTS is a 30M-parameter patch-based transformer trained on a curated corpus that is disjoint from the public benchmarks. It is one of the experts in the pool.
Intended use
Research comparison and benchmark evaluation of zero-shot forecasting systems.
Not intended use
This model is not for any commercial or production use. For production use, see ChakraTS.
Citation
@misc{chakratslab2026,
title = {ChakraTS-Lab: research configuration of ChakraTS},
author = {YHat Labs},
year = {2026},
url = {https://huggingface.co/yhatlabs/ChakraTS-Lab}
}