Irregular observations, missing values, and high-frequency signals. One time-series platform for forecasting, anomaly detection, and imputation.
Trials are by invitation.
Meet yoft-o, yoft’s orchestration model. It combines time-series models to deliver accurate forecasts without task-specific fine-tuning.
The orchestration model behind yoft
The lowest normalized MASE among forecasting methods that use neither dataset-specific fine-tuning nor LLMs, based on the published GIFT-Eval results.
‡ Our ranking by normalized MASE uses the public results retrieved on 2026-09-14. “No fine-tuning” means no dataset-specific training of model weights. “No LLM” means no LLM in the forecasting pipeline, including retrieval selection. EXAONE-Forecast-Agent and STRIDE w/ Synapse, the two higher-scoring entries, are excluded because their primary sources document LLM use in forecasting. Under these conditions, yoft-o (listed as TIMEHEDGE) ranks first. This is not an official leaderboard category. The assistant for explaining results and conversation is a separate feature.
† Based on 129 GIFT-Eval entries retrieved on 2026-09-14, ranked by normalized MASE. This differs from the leaderboard’s default ordering. Lower MASE and CRPS are better.
These results evaluate the model used for regularly sampled forecasts. Performance on public datasets does not guarantee accuracy on your data.
yoft handles forecasting and, beyond it, the whole path from field data to a decision, on one foundation.
Break forecasts into trends, seasonal patterns and residuals to examine the reasoning behind a prediction. We can integrate DecompSSM through a custom implementation tailored to your data and workflows. Component breakdowns are not included in standard API responses.
Irregular intervals, missing values, and high-frequency signals. Designed to work with data before it has been cleaned and reshaped.
Forecast, detect anomalies, fill gaps and obtain embeddings for similarity search or classification. Classifiers and label decisions are implemented by the caller.
From forecasts to analysis and explanation, combine our standard APIs with custom implementations for your workflows.
| Item | Major public foundation modelsPrimarily focused on forecasting | yoftStandard APIs and custom implementations |
|---|---|---|
| Forecasting | supported | supported |
| Probabilistic forecasts (ranges) | supported | supported |
| Covariates | supported | supported |
| Irregular spacing and imputation | partly supported | supported |
| Signal waveforms | Outside core scope | supported |
| Anomaly detection | Outside core scope | supported |
| Embeddings for search / classification | Outside core scope | supported |
| Explainability | Outside core scope | Custom implementation |
Public-model coverage is a general tendency, not an exhaustive survey. The yoft API returns embeddings; callers implement their own classifiers. Component-based explanations are available through custom implementation, separately from the standard API.
Intervals vary, values go missing, and high-frequency waveforms are mixed in. yoft takes such field data as input without rounding it.
Pass the observation times as they are. There is no resampling to a grid; the spacing itself carries information.
Gaps stay gaps. When you need values, yoft returns estimates with a forecast range.
ECGs and vibration are the same kind of input. yoft reads the cycle without breaking it and finds the beat that does not fit.
Illustrative signals demonstrate the processing flow; they are not measured data or an accuracy evaluation.
Time-series forecasting and signal reconstruction research accepted at ICASSP 2026, the world's top international conference in signal processing. These technologies are available through custom implementations tailored to your data and workflows.
A forecasting study with separate state-space branches for trends, seasonal patterns and residuals. Shared multivariate context helps examine how each component behaves.
Learn more
Research into reconstructing vital signs from photoplethysmography (PPG), combining flow matching with a state-space model.
Learn moreOur team brings research in computer vision, AI and signal processing into practice. Members have also won a CVPR competition and worked on joint research and technical advisory projects with national university hospitals and major companies.
Send your past numerical data and how many future points you want to predict. Receive forecasts and prediction ranges to use in your service, with no model selection or fine-tuning setup.
curl -X POST "$YOFT_API_BASE/v1/forecast" \
-H "Authorization: Bearer $YOFT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"series": [{ "name": "load", "values": [412.0, 398.5, 401.2, "…", 455.8] }],
"horizon": 24,
"quantiles": [0.1, 0.5, 0.9]
}'{
"request_id": "req_a1b2c3d4e5f67890a1b2c3d4e5f67890",
"model": "yoft",
"model_version": "v1",
"forecast": [{
"name": "load",
"values": [458.1, 462.7, "…"],
"quantiles": {
"0.1": [441.3, 443.9, "…"],
"0.5": [458.1, 462.7, "…"],
"0.9": [474.8, 481.2, "…"]
}
}],
"usage": { "input_points": 168, "output_points": 24, "series_count": 1 }
}| POST/v1/forecast | Forecasting with multivariate, covariate, quantile and irregular-timestamp support |
| POST/v1/anomaly | Per-step anomaly scores (distributional surprise and forecast error) |
| POST/v1/impute | Estimates and quantiles at missing positions |
| POST/v1/embed | Fixed-length vectors for similarity search and classification |
| GET/v1/models | List the models available to your key |
up to 64 series per request16–2048 points per serieshorizon up to 512quantiles at no extra charge
Time-series data sent to the API is never used to train or fine-tune models. Raw time-series values are not stored or logged.
Read the API documentation — no sign-in required
Invited accounts can try 16 use cases and their own CSV in the Playground. Result aggregates and chat content are sent to an LLM only when using the AI assistant in the Playground. When the assistant is enabled, aggregates are temporarily stored with a 24-hour expiry. Sign in
Deployed across a wide range of industries: anywhere a number moves with time.
Time-series data sent to the API is never used to train or fine-tune models. Raw time-series values are not stored or logged and are discarded after analysis.
Yes. Attach timestamps and the series is forecast without resampling. Missing values themselves can be reconstructed with /v1/impute, as estimates with a forecast range.
Numerical forecasts use time-series models, not an LLM. Result aggregates and chat content are sent to an LLM only when you use the AI assistant in the Playground. Raw time-series arrays are not automatically attached from the results. Anything you enter in the chat is sent and may be processed outside Japan.
yoft-o, yoft’s orchestration model. It combines models developed in Japan with public time-series foundation models. The internal composition evolves while the API contract stays stable.
No. Use the public benchmark results and a check on your own data as the basis for a decision. We run such checks on request.
Ask us. The API is the default, but other forms of delivery are considered case by case.