ChakraTS
ChakraTS is a probabilistic time-series forecasting model from YHat Labs, served through an API. This repository holds the model card; there are no weights to download. Updated November 2026.
Send your series and a horizon, and get quantiles for every future step. ChakraTS forecasts new series without being trained on them, reads past and known-future covariates (numeric or categorical), forecasts several targets of a series together, and works at any regular frequency from seconds to years.
- Website: https://yhatlabs.com
- API access: https://yhatlabs.com/models/access
- Python library:
pip install yhatlabs(GitHub) - Recipes and benchmark files: yhatlabs-cookbook
- Tabular model: yhatlabs/ChakraTab
Benchmark standing (as of November 2026)
ChakraTS is the best model across forecasting benchmarks.
| Benchmark | Standing | Result | Leaderboard |
|---|---|---|---|
| fev-bench (Amazon) | 1st by win rate | Win rate 90.1%, skill score 51.4 | fev-bench |
| GIFT-Eval (Salesforce) | 2nd among non-agentic models | MASE 0.640 | GIFT-Eval |
| TIME | 1st by MASE rank | MASE 0.623, CRPS 0.522 | TIME |
GIFT-Eval and TIME scores are normalized by Seasonal Naive (lower is better). The results use the pro preset. Result
files and the code to reproduce them are in the cookbook.
Getting started
from yhatlabs import Client
yh = Client() # reads YHAT_API_KEY
fc = yh.forecast(df, h=24, preset="pro") # df: one row per series and timestamp
print(fc.df.head())
Or call the API directly:
curl -X POST https://api.yhatlabs.com/v1/series/forecast \
-H "Authorization: Bearer $YHAT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "chakra-ts",
"preset": "pro",
"series": [
{"id": "store_1", "values": [12, 15, 11, 18, 20, 17, 22, 19, 24, 21, 25, 23, 27, 26]},
{"id": "store_2", "values": [40, 38, 45, 47, 44, 50, 52, 49, 55, 53, 58, 57, 60, 62]}
],
"horizon": 7,
"freq": "D"
}'
Each forecast comes back as a [horizon, 9] array of quantiles (levels 0.1 to 0.9 by default). The median is the
point forecast.
Capabilities
| Property | Detail |
|---|---|
| Training on your data | not needed; optional tuning on your own history |
| Presets | fast (default) and pro (most accurate, slower) |
| Context | up to 32,768 past points per series |
| Horizon | up to 2,048 steps |
| Targets | one or several per series, forecast together |
| Covariates | past and known-future, numeric or categorical |
| Missing values | allowed, marked as null |
| Output | quantiles at any levels (the API returns 0.1 to 0.9 by default), sample paths |
| Batching | up to 256 series per request; the Python library splits larger inputs |
Intended use
- Energy: load, renewable generation and electricity price forecasting, with published forecasts as covariates.
- Retail and supply chain: demand per product and location, with promotions, prices and calendar effects.
- Operations and infrastructure: traffic, capacity, call volumes, cloud and sensor metrics.
- Finance and planning: revenue, cash-flow and KPI projections with prediction intervals.
- Scenario analysis and backtesting: the same series under alternative covariate paths, and rolling-origin evaluation on your own history.
Not intended use
- Series without a regular time index.
- Very short histories: forecasts are returned, but the intervals are wide.
- Classification or anomaly labelling as such; see ChakraTab for prediction on tables.
Terms
Use of the API is governed by the YHat Labs terms, which also describe how request data is handled. ChakraTS is available for commercial use under those terms; enterprise and on-premises arrangements on request.
Citation
@misc{chakrats2026,
title = {ChakraTS: probabilistic time-series forecasting as a service},
author = {YHat Labs},
year = {2026},
url = {https://huggingface.co/yhatlabs/ChakraTS}
}
Contact
Evaluation results
- Win rate (all 100 tasks) on fev-benchfev-bench leaderboard0.901
- Scaled quantile loss skill score (all 100 tasks) on fev-benchfev-bench leaderboard51.400
- MASE (geometric mean, relative to seasonal naive) on GIFT-EvalGIFT-Eval leaderboard0.640
- CRPS (geometric mean, relative to seasonal naive) on GIFT-EvalGIFT-Eval leaderboard0.436
- MASE (geometric mean, relative to seasonal naive, 98 tasks) on TIMETIME leaderboard0.623
- CRPS (geometric mean, relative to seasonal naive, 98 tasks) on TIMETIME leaderboard0.522