NVIDIA makes Kumo Forecast available for time series forecasting using historical patterns
NVIDIA Kumo Forecast combines time series forecasts with retrieval of similar historical patterns and offers explanations for individual forecast horizons. Both the code and the weights are available under the Apache 2.0 license.
NVIDIA has made Kumo Forecast available for multivariate time series forecasting. It is based on MOMENT-1-large and includes a base encoder with approximately 340 million parameters and a trainable prediction layer. The code is available in the NVIDIA/Kumo-TS repository, and the weights are available under nvidia/Kumo-Forecast, both under the Apache 2.0 license.
The optional Domain-Aware Representation and Retrieval (DARR) mode combines a direct forecast with similar patterns retrieved from historical data. Historical context can be supplied as a separate pandas DataFrame, independently of the current input window. Developers can adjust the balance between the two components. Supported input lengths are 256, 512, 1024 and 2048 observations; the data must contain timestamps and numerical values.
The model also includes explanations of results for individual forecast horizons, including the contributions of variables and time lags. The library allows the forecast length to be set, and the repository includes an example of adapting the model to custom data in CSV format. The source recommends validating quality for a specific use case by comparing direct mode and DARR mode on the same test segments.
Why it matters
When working with time series, recent observations for forecasting and a broader history for finding similar patterns can be specified separately. Explanations for individual horizons allow users to explore which variables and time lags influenced the result. However, the benefit of historical retrieval needs to be validated on the specific data.
Release card
NVIDIA Kumo Forecast
NVIDIA
- Specifications
- The base encoder has approximately 340 million parameters. The model also includes a trainable prediction layer.
- Context
- Supported input sequence lengths are 256, 512, 1024 and 2048 observations. Separate historical context can be supplied to retrieve historical patterns.
- Inputs
- The input consists of time series with timestamps and one or more columns of numerical values. The output consists of forecasts of future values and, optionally, explanations for individual forecast horizons.
- Licence
- Apache 2.0
- Availability
- The code is available in the NVIDIA/Kumo-TS repository, and the weights are available under nvidia/Kumo-Forecast.
- Supports multivariate time series forecasting.
- Allows similar historical patterns to be incorporated into forecasts through DARR mode.
- Provides explanations of the contributions of variables and time lags for individual forecast horizons.
- Input data must contain timestamps and numerical values.
- In DARR mode, the historical context must have dimensions similar to those of the input data.
- Forecasts beyond the native model horizon require autoregressive extension.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
Two audiences, two different impacts
What this means
For individuals
During experimentation, an analyst or developer can adjust the relative weighting of the direct forecast and historical patterns and explore explanations of the results for different horizons.
For a business
When incorporating the model into business forecasting, quality needs to be validated for the required horizons and operating conditions; the availability of DARR mode alone does not demonstrate an improvement in a specific process.
ProcessesCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.