NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability

NVIDIA has pushed Kumo Forecast, a time-series model that tries to do more than predict the next point and move on. The system, previously called NV-Tesseract Forecasting, combines a pretrained Transformer with a retrieval path that can pull related patterns from historical data and fold them into the forecast.

The core of the model is MOMENT-1-large, used as a frozen encoder while a trainable forecasting head learns the job-specific mapping to future values. NVIDIA says the base encoder is about 340 million parameters. That split matters: the representation model stays general, while the prediction layer adapts to the task at hand.

There are really two forecast modes here. The direct path uses the model’s own learned representations. The DARR mode, short for Domain-Aware Representation and Retrieval, adds a nearest-neighbor search over historical context and blends those retrieved patterns with the direct prediction. The library keeps the input window and the retrieval context separate, with the latter passed in through a context_df argument.

Kumo Forecast also ships with an interpretability framework that sits alongside the model rather than replacing it. When turned on, it produces lag and feature attributions for each forecast horizon, plus semantic-flow and trajectory-stability diagnostics, with optional embedding Integrated Gradients. NVIDIA says these explanations come with the forecast itself, which is a nicer idea than forcing people to squint at a black box and guess.

The package is aimed at pandas users, with a DataFrame-first API, configurable horizons, and support for direct forecasting plus autoregressive extension beyond the model’s native range. Inputs need to be numeric temporal data, with supported sequence lengths of 256, 512, 1024, and 2048 observations. Kumo Forecast is out under Apache 2.0, and the Kumo-TS repository includes code, weights, and an example for fine-tuning on a user-supplied CSV.