Skip to main content

Time-Series Foundation Model

To use a time-series foundation model, deploy its service first. For instructions, see Deploy a Time-Series Foundation Model.

Description

Time-series foundation models are specifically trained to handle advanced time-series analysis tasks such as forecasting, anomaly detection, and imputation. These models inherit the strong generalization capabilities of large foundation models and can perform predictive analysis based on input data without the need for complex parameter configuration.

NumberParameterDescription
1tdtsfm_1TDtsfm v1.0
2timemoe-fcTime-MoE SQL model name
3moiraiSalesforce time-series foundation model
4chronosAmazon time-series foundation model
5timesfmGoogle time-series foundation model
6momentCMU time-series foundation model

After a model service is deployed and configured in TDgpt, invoke it through SQL by specifying its model name and the number of output rows.

  • The following statement forecasts 10 rows of data using TDtsfm:
SELECT _frowts, FORECAST(i32, "algo=tdtsfm_1,rows=10") FROM foo;
  • The following statement forecasts 10 rows of data using Time-MoE:
SELECT _frowts, FORECAST(i32, "algo=timemoe-fc,rows=10") FROM foo;

References

  1. Time-MoE
  2. Moirai
  3. Chronos
  4. TimesFM
  5. MOMENT