NVIDIA Kumo Forecast: Pretrained Time-Series Backbone Plus Historical Pattern Retrieval and Interpretability

NVIDIA Kumo Forecast (formerly NV-Tesseract Forecasting) is a multi-variate time-series forecasting model that pairs a pretrained Transformer backbone with Domain-Aware Representation and Retrieval (DARR), a mode that blends direct model predictions with patterns retrieved from historical data and returns interpretable explanations alongside the forecast. It matters because it splits forecasting into two information sources — pretrained representations and the historical observations of the current application — and lets developers tune how much the final prediction leans on each, while adding an interpretability layer that does not require changing the underlying forecaster.

What Kumo Forecast Is

Kumo Forecast is released under the Apache 2.0 license and ships as both model code and model weights.

The model is built on the MOMENT-1-large backbone and uses an approximately 340-million-parameter base encoder plus a trainable forecasting head. The library provides a DataFrame-first API, configurable forecast horizons, optional historical context for retrieval, and an interpretability framework that produces horizon-specific explanations without modifying the underlying forecaster.

Context-Enhanced Forecasting: Direct Prediction Plus Retrieval

Kumo Forecast is designed around two distinct sources of information: representations learned during pretraining, and historical observations available for the current application.

In its direct forecasting path, Kumo Forecast uses a MOMENT-based Transformer and a trained forecasting head to predict future values. In Domain-Aware Representation and Retrieval (DARR) mode, it additionally retrieves related patterns from historical data using a nearest-neighbor approach and combines that information with the direct forecast.

This creates a distinction between the model's input window and its retrieval context. The input window supplies recent observations for the forecast, while a separate historical DataFrame can supply additional context for pattern retrieval. Developers can adjust how strongly the final prediction relies on each pathway through the library's configuration options.

The DARR mode is described in the published paper Memory-Augmented Forecasting: Scalability and Generalization Across Temporal Domains.

The Four Architectural Components

Kumo Forecast combines four main 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. The architecture keeps encoder and embedder weights frozen while training the forecasting head, separating the pretrained representation model from the task-specific prediction layer.

Direct forecasting

The forecasting head maps the model's representations to future values for the requested horizon. For horizons beyond the model's native capability, Kumo Forecast supports autoregressive extension. This distinguishes a directly generated forecast from a longer forecast that extends predictions recursively.

Domain-Aware Representation and Retrieval (DARR)

DARR augments direct forecasting with k-nearest-neighbor retrieval from historical context data. The process combines direct forecasts with historical-pattern retrieval, and the SDK accepts the historical context through a separate context_df argument.

Interpretability

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

  • lag and feature attributions per forecast horizon,
  • semantic-flow diagnostics,
  • trajectory-stability diagnostics,
  • optional embedding Integrated Gradients.

These outputs accompany the forecast rather than requiring a different forecasting architecture. Further details are documented in the paper NVExplain: Explaining Time Series Forecasting with Latent Trajectory Analysis and Structure-Preserving Surrogates.

Supported Inputs and Limitations

Kumo Forecast is restricted to numeric, temporal data. Supported input sequence lengths are 256, 512, 1024, and 2048 observations. Inputs consist of a timestamp column plus one or more numeric value columns, represented as a pandas DataFrame or supplied as CSV or JSON input. When using DARR mode, the dimensions of the context must be similar to the context of the input.

Demonstrated Applications

Two integrations are shown in the release material.

Predictive maintenance. The Factory Operations (FOX) Blueprint includes Early Access to the Predictive Maintenance module, which leverages Kumo Anomaly and Kumo Forecast. It analyzes real-time data to predict future anomalies. A demo video is published alongside the model.

Platform integration. A GTC San Jose 2026 session demonstrates platform integration; the release links to the on-demand recording hosted by NVIDIA.

How to Evaluate and Adapt Kumo Forecast

NVIDIA recommends application-specific evaluation: compare the direct and retrieval-enhanced configurations on the same held-out forecasting windows, and validate forecast quality across the intended horizons and operating conditions, consistent with NVIDIA's requirement for use-case-specific testing.

The Kumo-TS repository also provides an example for fine-tuning Kumo Forecast on a user-supplied CSV, giving developers both a pretrained inference workflow and a documented starting point for adapting the model to their own data.

Sources