Answers are grounded in this paper and cite their sources.
Prediction over structured data such as forecasting future values and classifying rows has long been a central concern for the database and data-mining communities. Our earlier tutorials presented in the community (VLDB, SIGMOD, KDD) traced the evolution from classical models and tensor methods to neural forecasting, probabilistic forecasting, and practical systems. Recently, foundation models have transformed machine learning, with large-scale pretraining followed by zero-shot or in-context inference now defining the state of the art. This paradigm has arrived for structured data, with foundation models emerging for both time series (Chronos, TimesFM, Moirai, Toto, etc.) and tabular tasks (TabPFN, TabICL, Mitra, Limix, etc.). This proposal continues our tutorial line and adds the modern layer of foundation models for structured data. The objective of this tutorial is to provide a comprehensive survey of foundation models for structured data for a database audience. We review the state of the art in two related areas: (1) time-series foundation models, covering tokenization, zero-shot forecasting, covariates, and multivariate prediction, and (2) tabular foundation models, covering synthetic priors, in-context learning, and relational pre-training. We also discuss how these two fields are converging with shared pretraining strategies and unified architectures. As for their implication for data management systems, we provide case studies from AWS how forecasting helps improving the efficiency of the system with predictive auto-scaling and capacity planning. Further, we argue that pretrained models open the door to tighter native integration of some predictive tasks into OLAP workflows, because many workloads can now shift from task-specific training pipelines to context construction plus inference inside analytical plans.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: