While using online datasets for machine learning is commonplace today, the quality of these datasets impacts on the performance of prediction algorithms. One method for improving the semantics of new data sources is to map these sources to a common data model or ontology. While semantic and structural heterogeneities must still be resolved, this provides a well established approach to providing clean datasets, suitable for machine learning and analysis. However, when there is a requirement for a close to real time usage of online data, a method for dynamic Extract-Transform-Load of new sources data must be developed. In this work, we present a framework for integrating online and enterprise data sources, in close to real time, to provide datasets for machine learning and predictive algorithms. An exhaustive evaluation compares a human built data transformation process with our system’s machine generated ETL process, with very favourable results, illustrating the value and impact of an automated approach.