Compared to daily rainfall data, observed subdaily rainfall times are rare and often very short. For hydrologic modeling, this problem is often addressed by generating synthetic hourly rainfall series, with rainfall generators calibrated on relevant rainfall statistics. The required subdaily rainfall statistics are traditionally derived from daily rainfall records by assuming some temporal scaling behavior of these statistics. However, as our analyzes of a large data set suggest, the mathematical form of this scaling behavior might be specific to individual gauges. This paper presents, therefore, a novel approach that bypasses the temporal scaling behavior assumption. The method uses multivariate adaptive regression splines; it is learning-based and seeks directly relationships between target subdaily statistics and available predictors (including (supra-) daily rainfall statistics and external information such as large-scale atmospheric variables). A large data set is used to investigate these relationships, including almost 340 hourly rainfall series coming from gauges spread over Switzerland, the USA and the UK. The predictive power of the new approach is assessed for several subdaily rainfall statistics and is shown to be superior to the one of temporal scaling laws. The study is completed with a detailed discussion of how such reconstructed statistics improve the accuracy of an hourly rainfall generator based on Poisson cluster models.