NVIDIA Releases Kumo Forecast for Time Series Prediction

NVIDIA releases Kumo Forecast, a multivariate time series forecasting model, as detailed in a company blog post. The model combines predictions from a pretrained Transformer with patterns retrieved from historical data.

Kumo Forecast uses the MOMENT-1-large backbone, with an approximately 340 million parameter encoder and a trainable forecasting head. Its optional Domain Aware Representation and Retrieval mode uses nearest neighbor retrieval to find relevant patterns in a separate historical data set, then combines them with the direct forecast.

The library accepts numeric temporal data through pandas DataFrames, CSV files, or JSON. It supports input lengths of 256, 512, 1024, and 2048 observations, configurable forecast horizons, and explanations that show lag and feature attribution for each horizon. NVIDIA provides the code and model weights under the Apache 2.0 license.