JointTimePFN: amortized joint prediction on partially observed multivariate time series

Multivariate time series rarely form a complete history followed by an unknown future. Missing measurements, channel-specific horizons, and known future covariates create partially observed time-channel grids, requiring models to use arbitrary observations and capture dependencies among missing values. We introduce JointTimePFN, a mask-conditional prior-data fitted network that unifies forecasting and imputation and returns an explicit joint predictive distribution in one forward pass. A mask-aware temporal encoder and low-rank Gaussian head are pretrained on a broadened linear model of coregionalization prior. We assess posterior fidelity against exact and MCMC-based multi-output Gaussian process references under matched and shifted conditions. Under in-prior forecasting and ragged masks, the joint head reduces median KL to a Bayesian reference by 75-83% relative to its diagonal approximation. On MIMIC-IV, the same checkpoint transfers zero-shot across various masking patterns and improves joint likelihood over its own independent marginals throughout. These results show that temporal PFNs can amortize useful joint dependence across flexible observation patterns.