From Forecasting to World Modeling: Can Temporal Foundation Models Recover Causal Response?

To use a temporal foundation model as a world model, a forecast has to be distinguished from the response to an intervention. In retail, predicting sales at a price the retailer has not set is not the same as predicting what setting that price would do, and the two answers differ when the conditions that moved price lie outside what the model sees. We test the difference with both real retail transaction data and a synthetic retail panel whose interventional outcomes are known by construction. We find that additional context, the resource these models are designed to exploit, reduces forecast error while leaving the recovered price response no more accurate. A model pretrained for causal inference rather than for forecasting inherits the same bias. Moreover, what improves causal recovery is the form in which the identifying variable enters and not the information it carries. Identification is not a property pretraining can supply, and more observational data is no substitute for an identification strategy.