NEUROGICA
Contact
← Back to newsNEWS
News

Neurogica Launches General API Availability of yoft, a Time-Series Foundation Model for Operational Data

Neurogica Inc. has developed yoft, a time-series foundation model for operational data, and opened general API access on the service site yoft.ai.

Key visual of yoft: a time-series foundation model for operational data, now generally available as an API

yoft is a foundation model for the kinds of data that operations produce: electricity demand, sales, equipment sensor readings and more. It handles irregular observation intervals, missing values and high-frequency signals, and performs forecasting, anomaly detection and imputation on one model. The API lets companies build forecasting into day-to-day work without developing a model for each use case.

Based on our survey of public information as of October 14, 2026, yoft is the first generally available API from a Japanese company that combines a time-series foundation model handling irregular intervals and missing values with all three of these functions.

Three problems in using operational data

  • Building a model per use case. Training only on the target's own data means redesigning and retuning every time a product or a machine is added. yoft is pre-trained on time series from many domains and works on unseen data without additional training.
  • Data that is not clean. Sensors sample at different times, and outages and maintenance leave gaps. yoft takes irregular intervals and missing values as they are.
  • Explaining a forecast. Supply planning and maintenance need the reasoning behind a number, not just the number. An explainability feature that shows which variables and time points mattered most is available on request.

First among methods with no additional training on GIFT-Eval

GIFT-Eval is a public time-series forecasting benchmark from Salesforce AI Research covering seven domains (energy, transport, economics and finance, healthcare, nature, sales, and web and cloud operations) across 97 evaluation settings.

Our orchestration model yoft-o (listed as TIMEHEDGE) ranked first in normalized MASE among methods that do not train on the evaluated data, and third among all 143 entries including those that do. Against the seasonal-naive baseline it reduced error by about 35% on the aggregate over the 97 settings (our reading of the public results as confirmed on October 14, 2026).

Availability

The forecasting, anomaly detection and imputation APIs are generally available at yoft.ai, where you can also try your own data in the browser. Inference runs on our servers, so no GPUs or model infrastructure are needed on the user side. Evaluations with your data, integration into business systems, and deployment on your own servers or edge devices are available by consultation.

The full press release (in Japanese) is on PR TIMES. Demos of yoft are also on show at our booth at CEATEC 2026, Makuhari Messe, October 13–16.