In this paper, we introduce an enhanced schema-less database language that supports prediction queries natively-the Prediction Query Language (PQL). Data in the PQL representation can be naturally modeled as an exchangeable multi-dimensional array. The seminal result by Aldous and Hoover (1980s), generalizing the classical result of De Finetti (1937), provides a canonical latent variable model characterization for such an exchangeable multi-dimensional array. We present a three-layer neural-network-based architecture that encodes this latent variable model representation and realizes an atomic prediction query. Using PQL, learning problems of Regression, Classification, Time-Series, Matrix and Tensor Completion can be solved simply by defining “schema” in PQL and then running predictive query. With the help of various benchmark datasets for each of Classification, Regression, Time Series and Matrix/Tensor Completion, we find that this out-of-the-box performance of PQL is comparable with the state-of-the-art results obtained with solutions tailored specifically for the scenarios.