The abstract for Morpheus-2.0 frames a familiar problem: a universal forecaster must work across tasks that differ in schema, forecast horizon, and temporal resolution. It also notes that the number of targets and covariates varies from dataset to dataset, which makes a one-size-fits-all model difficult to build.
Morpheus-2.0 is presented as a response to this challenge, with a focus on probabilistic forecasting. The key architectural idea is dual-axis attention, which the title suggests is meant to handle multiscale dependencies in time-series data.
Because only the abstract is available, the details of the attention mechanism and training procedure are not described in the source. The paper's contribution, as stated, is a universal forecaster that can adapt to heterogeneous forecasting tasks without being retrained for each one.