From Temporal Foundation to World Models: Counterfactual Probing with Exact Ground Truth
Maria F Scarcela ⋅ Wellington Franco ⋅ Carlos Caminha ⋅ Dennis Thumm
Abstract
Time series foundation models (TSFMs) increasingly accept known future covariates, which makes it possible to hand them a planned intervention. Yet it remains unclear whether their forecasts respond to that intervention as a world model would, by representing the alternative future under the action. We probe ten public TSFMs to assess whether they can receive such interventions and, when they can, whether their forecasts respond in a causally meaningful way. Three models expose no usable intervention channel, while another cannot propagate the supplied action to the queried outcome under our protocol. Among the remaining models, responses vary widely, and responsiveness alone does not imply causal correctness.
Chat is not available.
Successful Page Load