The increasing availability of manufacturing data and advanced analysis tools are forcing the demand fordata-driven approaches to improve the quality of workpieces and the efficiency of manufacturing processes.The analysis of real manufacturing data is challenging due to frequent changes in production circumstances.In this work, machine learning methods based on the data along the value chain of hydraulic valves are usedto predict the leakage results during end-of-line testing. The leakage volume flow measurement results arevery sensitive to changes in gap geometry and temperature level in the measurement cross-section.Additional measurements and experiments are required to interpret the systematic influences of the inputdata on the target variable and to introduce the missing information into the dataset. The design of a metamodel using experiment data supports the identification of statistical patterns to be applied to the realproduction dataset as a feature. This paper presents a systematic approach to hand-crafted feature engineeringthat improves the quality prediction of end-of-line hydraulic leakage testing.