license: other
license_name: yhat-labs-api-terms
license_link: https://yhatlabs.com/terms
pipeline_tag: time-series-forecasting
tags:
- time-series
- forecasting
- zero-shot
- probabilistic-forecasting
- foundation-model
- covariates
- hosted-api
model-index:
- name: ChakraTS
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: 0.67
- type: CRPS
name: CRPS (geometric mean, relative to seasonal naive)
value: 0.461
source:
name: GIFT-Eval leaderboard
url: https://huggingface.co/spaces/Salesforce/GIFT-Eval
- task:
type: time-series-forecasting
dataset:
name: fev-bench
type: autogluon/fev_datasets
metrics:
- type: win rate
name: Win rate (all 100 tasks)
value: 0.855
- type: SQL skill score
name: Scaled quantile loss skill score (all 100 tasks)
value: 48.1
source:
name: fev-bench leaderboard
url: https://huggingface.co/spaces/autogluon/fev-leaderboard
- task:
type: time-series-forecasting
dataset:
name: TIME
type: Real-TSF/TIME
metrics:
- type: MASE
name: MASE (geometric mean, relative to seasonal naive, 98 tasks)
value: 0.639
- type: CRPS
name: CRPS (geometric mean, relative to seasonal naive, 98 tasks)
value: 0.537
source:
name: TIME leaderboard
url: https://huggingface.co/spaces/Real-TSF/TIME-leaderboard
ChakraTS
ChakraTS is a zero-shot probabilistic time-series forecasting model from YHat Labs, served as a hosted API. This repository holds the model card; there are no weights to download.
Send a batch of series and a horizon, get back nine quantiles (10% to 90%) for every future step. ChakraTS forecasts zero-shot: nothing is trained or fitted on your data, and fine-tuning on customer data is not part of this release. It accepts known-future covariates, works at any regular frequency from minutes to years, and returns the per-step prediction intervals that scheduling, procurement, bidding and inventory decisions depend on.
- Website: https://yhatlabs.com
- Request API access: https://yhatlabs.com/models/access
- Code, benchmark files and notebooks: https://github.com/yhatlabs/yhatlabs
- Leaderboards: TIME · GIFT-Eval · fev-bench
- Research configuration (non-commercial): yhatlabs/ChakraTS-Lab
- Tabular model: yhatlabs/ChakraTab
Getting started
Access is by API key, issued on request. One call forecasts a batch of series:
curl -X POST https://api.yhatlabs.com/v1/forecast \
-H "Authorization: Bearer $YHAT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"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"
}'
Response (abridged):
{
"horizon": 7,
"quantile_levels": [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9],
"forecasts": [
{"id": "store_1", "quantiles": [[22.1, 23.4, 24.3, 25.0, 25.7, 26.4, 27.2, 28.2, 29.8], "..."]},
{"id": "store_2", "quantiles": ["..."]}
],
"weights": {"...": "contribution of each underlying forecaster to this request"}
}
A Python client (pip install yhatlabs) with a DataFrame interface, automatic frequency inference and the same argument names as other forecasting APIs is in preparation; it will be published in the GitHub repository.
Input and output contract
| Field | Meaning |
|---|---|
series[].id |
your identifier, echoed back |
series[].values |
past observations, oldest first, at a regular spacing; null marks a missing value |
series[].future_covariates (optional) |
known-future inputs aligned to the horizon, for example a demand forecast, a published day-ahead price, a promotion or holiday flag |
horizon |
number of steps to forecast |
freq |
spacing of the observations as a pandas offset alias (5min, 15min, h, D, W, MS, QS, YS) |
forecasts[].quantiles |
a [horizon, 9] array for the levels 0.1, 0.2, …, 0.9 |
weights |
how much each underlying forecaster contributed to this request |
Quantiles are monotone in the level. The median (level 0.5) is the point forecast.
Capabilities
| Property | Detail |
|---|---|
| Training on your data | none; zero-shot |
| Context | up to 16,384 past points per series |
| Horizon | up to 1,024 steps |
| Frequencies | any regular frequency, 5-minute to yearly |
| Covariates | known-future numeric or categorical inputs |
| Missing values | allowed, marked as null |
| Output | 9 quantiles per step (10% to 90%), monotone |
| Batching | many series per request; latency grows with the batch, not with history length |
Evaluation
The TIME, GIFT-Eval and fev-bench entries are all published on the official leaderboards.
All numbers come from the official benchmark protocols on the full benchmarks, scored against the official seasonal-naive baselines. No benchmark test data is used at any point, and no benchmark training split is used to fit anything.
TIME (98 tasks, 50 datasets) — published
TIME (ICML 2026) evaluates zero-shot forecasting on fresh datasets. Geometric mean over tasks, relative to seasonal naive; lower is better. The ChakraTS entry is its TIME configuration (chakra-ts-time). Comparison rows are the official leaderboard entries at submission time (October 2026).
| Model | MASE | CRPS |
|---|---|---|
| QiYao-M | 0.638 | 0.536 |
| ChakraTS | 0.639 | 0.537 |
| TimesFM-3 (non-commercial) | 0.640 | 0.536 |
| Toto-2.0-2.5B | 0.642 | 0.539 |
| Chronos-2 | 0.662 | 0.556 |
| T0-beta | 0.665 | 0.555 |
| TimesFM-2.5 | 0.669 | 0.567 |
GIFT-Eval (97 dataset configurations, 23 datasets) — published
Geometric mean over configurations, relative to seasonal naive; lower is better. The leaderboard sorts by mean rank across configurations; ChakraTS is 3rd among non-agentic models (pretrained, zero-shot and fine-tuned entries) at merge time (October 2026, PR #225). Rows below are the official non-agentic board in its order.
| # | Model | MASE | CRPS |
|---|---|---|---|
| 1 | TimesFM-3 (non-commercial) | 0.667 | 0.456 |
| 2 | EXAONE-Forecast | 0.673 | 0.460 |
| 3 | ChakraTS | 0.670 | 0.461 |
| 4 | Toto-2.0-2.5B-FT | 0.679 | 0.463 |
| 5 | Lingjiang 2.0 | 0.666 | 0.457 |
fev-bench (100 tasks) — published
fev-bench ranks models by win rate (how often a model beats each other model across the 100 tasks) and reports the scaled quantile loss skill score vs seasonal naive; higher is better for both. The ChakraTS entry is its fev-bench evaluation configuration (chakra-ts-fev), which uses no component pretrained on fev-bench data, so no task is flagged for leakage. Rows are the official leaderboard at merge time (October 2026, PR #192).
| Model | Win rate | Skill score |
|---|---|---|
| ChakraTS | 0.855 | 48.1 |
| TimesFM-3 (non-commercial) | 0.849 | 49.0 |
| Lingjiang 2.0 | 0.816 | 47.0 |
| Chronos-2 | 0.778 | 47.3 |
| T0-beta | 0.752 | 46.7 |
| Toto-2.0-2.5B | 0.745 | 44.4 |
| TiRex-2 | 0.741 | 45.5 |
| TimesFM-2.5 | 0.678 | 46.7 (leaderboard figure, after its leakage replacement) |
Per-task result files and the evaluation notebooks are in the GitHub repository.
Approach
ChakraTS is a mixture of expert forecasters. Several pretrained forecasting models each produce a full predictive distribution for every series. Nothing is learned across customers or across requests: the same rule is applied to every series, there is no trained gate that could memorise a benchmark.
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 expert in the served pool.
Intended use
- Energy: load, renewable generation and electricity price forecasting at 5-minute to daily resolution, with grid operators' published forecasts as covariates.
- Retail and supply chain: demand forecasting per product and location, with promotions, prices and calendar effects as covariates; safety-stock and replenishment from the quantiles.
- Operations and infrastructure: traffic, capacity, call volumes, cloud and sensor metrics for planning and alerting.
- Finance and planning: revenue, cash-flow and KPI projections with prediction intervals.
- Scenario analysis: the same series under alternative covariate paths (a different price, a different weather forecast).
- Backtesting: rolling-origin evaluation on your own history before adopting a forecast, using the same endpoint.
Not intended use
- Series without a regular time index.
- Histories shorter than a few seasonal cycles: forecasts are returned, but the intervals will be wide and should be treated accordingly.
- Classification of events or anomaly detection as such; ChakraTS returns forecast distributions, from which such signals can be derived but which it does not label. See ChakraTab for tabular classification and regression.
Data handling
Series sent to the API are used only to produce the response. They are not stored beyond the request, not used for training and not shared with third parties. Request metadata (timestamps, sizes, latency) is logged for operations and billing.
Terms
Use of the API is governed by the YHat Labs API terms. ChakraTS is available for commercial use under those terms; enterprise and on-premises arrangements on request.
Citation
@misc{chakrats2026,
title = {ChakraTS: zero-shot probabilistic time-series forecasting as a service},
author = {YHat Labs},
year = {2026},
url = {https://huggingface.co/yhatlabs/ChakraTS}
}