Synthetic Causal Priors for In-Context Time-Series Classification

Time-series classification (TSC) tasks differ in sensors, sequence lengths, channel counts, and label semantics, making support-conditioned in-context prediction a natural interface. We study synthetic task priors and evaluation protocols for in-context TSC. Our first finding is that causal synthetic priors must control label visibility: labels should be generated by meaningful latent mechanisms, but their effects must remain recoverable from observed channels. A hidden-label SCM prior with an enforced label--observation path reaches 0.776 accuracy under full-support inference and 0.802 with retrieval-64 on the 42-task binary UCR subset, approaching an in-domain real-data mixup reference without using target benchmark series during pretraining. Our second finding is that protocol choices materially change reported results: moving from TiCT's re-split to the official UCR split reduces UCR-128 accuracy by 6.2 points under retrieval-64, and switching from retrieval-64 to full-support inference further lowers accuracy under the official split. Together, these results identify label visibility and protocol reporting as central issues in the design and evaluation of synthetic TSC priors.