Pretraining for Forecasting in Relational Databases

Real-world forecasting increasingly relies on asynchronous transaction logs in relational databases rather than uniformly sampled arrays. While relational deep learning (RDL) models can forecast from these complex histories, they are typically trained separately for each specific target and root entity type. We introduce Query-Conditioned Relational pretraining (QCRP), a self-supervised framework that enables a standard RDL backbone to act as a universal forecaster across a database. QCRP generates diverse forecasting objectives using a structured temporal query grammar and interleaves them through a shared conditional predictor. This produces a robust, frozen encoder reusable for downstream tasks via lightweight predictors or a zero-shot interface. Evaluated on large-scale RelBench databases, QCRP learns frozen representations that support competitive forecasting across tasks and entity types. Furthermore, a single unified QCRP model checkpoint is competitive with isolated, entity-specific specialists. We demonstrate promising results for reusing the learned representations from standard RDL backbones for forecasting tasks in relational databases.