NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability
NVIDIA introduced Kumo Forecast, a time-series model combining pretrained Transformers with historical pattern retrieval and interpretability tools for improved forecasting accuracy.
NVIDIA Kumo Forecast, previously named NV-Tesseract Forecasting, integrates a pretrained Transformer backbone with Domain-Aware Representation and Retrieval (DARR) to enhance multivariate time-series predictions. The model combines direct forecasting with historical pattern retrieval, offering interpretability features without altering the underlying forecaster. Built on the MOMENT-1-large architecture, it uses a 340-million-parameter base encoder paired with a trainable forecasting head, available under an Apache 2.0 license.
The system operates through two pathways: a direct forecasting path using a Transformer and forecasting head, and a DARR mode that retrieves related historical patterns via nearest-neighbor search. Developers can configure the balance between these pathways, with the retrieval context supplied separately through a historical DataFrame. The library supports configurable forecast horizons and provides an interpretability framework that generates horizon-specific explanations, including lag attributions and trajectory diagnostics.
Kumo Forecast consists of four main components: a pretrained temporal encoder based on MOMENT-1-large, a trainable forecasting head, a historical-pattern retrieval mechanism, and an interpretability framework. The pretrained encoder remains frozen during training, while the forecasting head adapts to task-specific predictions. For longer horizons, autoregressive extension is supported, distinguishing direct forecasts from recursively extended predictions.
The model requires numeric temporal data with supported input sequence lengths of 256, 512, 1024, or 2048 observations, including a timestamp column and one or more numeric value columns. The Kumo-TS repository offers an example for fine-tuning on user-supplied CSV data, providing both a pretrained inference workflow and a documented adaptation path. NVIDIA’s Factory Operations Blueprint includes early access to a Predictive Maintenance module leveraging Kumo Forecast and Kumo Anomaly for real-time anomaly prediction.