Notes on time series foundation models, recurrent architectures, and the ideas behind our work, with interactive figures where they help.
A model can only learn to use covariates if its training data rewards it. This post covers the synthetic coupling pipeline behind TiRex-2: augmentation, six ways to tie series together, and the post-processing that makes it look real.
Read postHow TiRex-2 reads future-known covariates in both directions without letting the target see its own future. Covers the reverse-pass time mixer, the asymmetric attention mask, and the two-hop leak the mask prevents.
Read post