2026-10-21 — views
NVIDIA Kumo Forecast Unveils DARR Mode for Interpretable Time-Series Prediction
Read this because Kumo Forecast introduces Domain-Aware Representation and Retrieval to blend direct predictions with historical context, offering developers a transparent, Apache 2.0 licensed solution for multivariate forecasting.
NVIDIA releases Kumo Forecast, a model combining pretrained Transformers with historical pattern retrieval and built-in interpretability for time-series tasks.
A New Approach to Multivariate Forecasting
NVIDIA has introduced Kumo Forecast, a multivariate time-series forecasting model designed to bridge the gap between learned representations and historical observations. Previously known as NV-Tesseract Forecasting, the system combines a pretrained Transformer backbone with a novel context-enhanced mode called Domain-Aware Representation and Retrieval (DARR). This architecture aims to improve prediction accuracy while providing interpretability for the results.
The model is built on the MOMENT-1-large backbone, utilizing an approximately 340-million-parameter base encoder. The architecture keeps the encoder and embedder weights frozen while training a separate forecasting head. This separation allows the model to leverage robust pretrained temporal representations while adapting to specific forecasting tasks. The library provides a DataFrame-first API, allowing developers to configure forecast horizons and optionally supply historical context for pattern retrieval.
How DARR Enhances Predictions
Kumo Forecast operates through two primary pathways. In its direct forecasting path, the model uses the MOMENT-based Transformer and the trained forecasting head to predict future values. For horizons extending beyond the model’s native capability, the system supports autoregressive extension, which recursively extends predictions.
The DARR mode augments this direct forecasting by retrieving related patterns from historical data using a nearest-neighbor approach. This process distinguishes between the model’s input window, which supplies recent observations, and a separate historical DataFrame that provides additional context for pattern retrieval. Developers can adjust the weight of each pathway through configuration options, blending direct model predictions with retrieved historical patterns. This distinction allows the system to draw on two different sources of information: representations learned during pretraining and historical observations available for the current application.
Built-In Interpretability Framework
A key feature of Kumo Forecast is its model-agnostic interpretability framework. When enabled, this framework produces an explanation bundle that accompanies the forecast without requiring a different forecasting architecture. The bundle includes lag and feature attributions per forecast horizon, semantic-flow diagnostics, trajectory-stability diagnostics, and optional embedding Integrated Gradients. This allows users to understand the reasoning behind specific predictions, addressing the black-box nature often associated with deep learning models.
Practical Implementation Details
Kumo Forecast is available under the Apache 2.0 license, with model code hosted at NVIDIA/Kumo-TS and weights at nvidia/Kumo-Forecast. The system requires numeric, temporal data with supported input sequence lengths of 256, 512, 1024, and 2048 observations. Inputs must consist of a timestamp column and one or more numeric value columns, represented as a pandas DataFrame or CSV/JSON file. When using DARR mode, the dimensions of the context data must be similar to the context of the input.
The model is already being integrated into broader applications, such as the Factory Operations (FOX) Blueprint. This blueprint includes an Early Access module for Predictive Maintenance that leverages both Kumo Anomaly and Kumo Forecast to analyze real-time data and predict future anomalies. Developers can also find examples for fine-tuning the model on user-supplied CSV data, providing a starting point for adapting the pretrained workflow to specific use cases.
Practitioner note
Developers should validate forecast quality across intended horizons and operating conditions, as NVIDIA requires use-case-specific testing. Ensure your historical context data dimensions match the input context when leveraging DARR mode to avoid retrieval errors.