Dynamic data

Dynamic data are generated by systems that evolve over time, such that both the observed state and the governing relationships may change. This includes time series, streaming data, and spatiotemporal systems where current behaviour depends on past evolution. The challenge is therefore not only temporal ordering, but the need to model ongoing change.

Static models trained on fixed datasets are often inadequate in such settings, as they cannot adapt to new information or shifting regimes. Common issues include outdated predictions, inability to capture rapid transitions, and accumulation of errors over time. Sequence and dynamics-based models provide a natural framework for handling temporal dependence and state evolution. Physics-informed methods are valuable when system dynamics are partially known, while graph-based approaches are useful for modelling interactions across networks or spatial structures. Uncertainty-aware methods are critical, as prediction confidence typically degrades over longer time horizons.

In many applications, data assimilation provides a powerful approach by continuously updating model states using incoming observations while accounting for uncertainty.