Towards Financial World Modeling: An Open Efficiency Question

We pose an open question about latent world models as an efficiency mechanism, not a solved one. On a Yahoo Finance benchmark of 319 equities (2018–2026),a compact latent world model matched or exceeded a Transformer’s directional accuracy on raw price data while training in a fraction of the time. No existing theory explains why imposing explicit latent-dynamics structure should be compute8 efficient here, or predicts when it should be preferred over unconstrained attention elsewhere. We offer this single-dataset only as a concrete anchor and ask the community to help identify the general principle, if one exists.