NVIDIA Kumo Forecast Combines Pretrained Time-Series Models with Historical Pattern Retrieval

NVIDIA Kumo Forecast Combines Pretrained Time-Series Models with Historical Pattern Retrieval

Published October 21, 2026

Highlights

NVIDIA Kumo Forecast, formerly called NV-Tesseract Forecasting, is a multivariate time-series forecasting model. It combines a pretrained Transformer backbone with Domain-Aware Representation and Retrieval (DARR), a context-enhanced mode that blends direct model predictions with patterns retrieved from historical data. The system also provides interpretability features for its results.

Kumo Forecast is built on the MOMENT-1-large backbone. Its base encoder contains approximately 340 million parameters and works with a trainable forecasting head.

The library provides a DataFrame-first API, configurable forecast horizons, optional historical context for retrieval, and horizon-specific explanations without requiring changes to the underlying forecasting model.

Kumo-Forecast is available under the Apache 2.0 license.

• Model code: NVIDIA/Kumo-TS
• Model weights: nvidia/Kumo-Forecast

Context-Enhanced Time-Series Forecasting

A forecasting system can use two sources of information: representations learned during pretraining and historical observations from the application domain. Kumo Forecast combines these sources through a pretrained forecasting model and an optional retrieval pathway over historical data.

In its direct forecasting path, Kumo Forecast uses a MOMENT-based Transformer and a trained forecasting head to predict future values. In DARR mode, it also retrieves related patterns from historical data using a nearest-neighbor method and combines those patterns with the direct forecast.

The model’s input window and retrieval context serve different purposes. The input window contains recent observations used to generate the forecast, while a separate historical DataFrame provides additional examples for pattern retrieval. Configuration options control the relative influence of the two pathways on the final prediction.

Additional information about DARR is available in the paper Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains.

How Kumo Forecast Works

Kumo Forecast consists of four primary components: a pretrained temporal encoder, a trainable forecasting head, a historical-pattern retrieval mechanism, and an interpretability framework.

Pretrained Temporal Representations

The model uses the MOMENT-1-large Transformer encoder as its backbone. During forecasting-head training, the encoder and embedder weights remain frozen. This separates the pretrained representation model from the task-specific prediction layer.

Direct Forecasting

The forecasting head converts the model’s representations into future values for the requested horizon. Kumo Forecast also supports autoregressive extension for horizons beyond the model’s native capability. This distinguishes a directly generated forecast from a longer forecast produced by recursively extending earlier predictions.

Domain-Aware Representation and Retrieval

DARR augments direct forecasting with k-nearest-neighbor retrieval from historical context data. The SDK accepts this historical information through a separate context_df argument and combines retrieved patterns with the direct forecast.

Interpretability

Kumo Forecast includes a model-agnostic interpretability framework. When enabled, it generates an explanation bundle containing:

• Lag and feature attributions for each forecast horizon
• Semantic-flow diagnostics
• Trajectory-stability diagnostics
• Optional embedding Integrated Gradients

These outputs accompany the forecast and do not require a separate forecasting architecture.

Further information about the interpretability approach is available in NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates.

Limitations

Kumo Forecast requires numeric temporal data. Supported input sequence lengths are 256, 512, 1024, and 2048 observations.

Inputs must include a timestamp column and one or more numeric value columns. Data can be supplied as a pandas DataFrame or through CSV and JSON input. When DARR mode is used, the context dimensions should be similar to those of the input data.

Predictive Maintenance Example

The Factory Operations (FOX) Blueprint provides Early Access to a Predictive Maintenance module that uses Kumo Anomaly and Kumo Forecast. The module analyzes real-time data to predict future anomalies.

Platform Integration

A platform integration session was presented at GTC San Jose 2026.

Evaluating Kumo Forecast

For an application-specific evaluation, direct and retrieval-enhanced configurations can be compared on the same held-out forecasting windows. Forecast quality should be validated across the intended horizons and operating conditions, with testing tailored to the specific use case.

The NVIDIA/Kumo-TS repository includes an example for fine-tuning Kumo Forecast on a user-supplied CSV. It supports both a pretrained inference workflow and a documented starting point for adapting the model to new data.