The increasing availability of data and computing capacity drives optimization potential. In the industrial context, predictive maintenance is particularly promising and various algorithms are available for implementation. For the evaluation and selection of predictive maintenance algorithms, hitherto, statistical measures such as absolute and relative prediction errors are considered. However, algorithm selection from a purely statistical perspective may not necessarily lead to the optimal economic outcome as the two types of prediction errors (i.e., alpha error ignoring system failures versus beta error falsely indicating system failures) are negatively correlated, thus, cannot be jointly optimized and are associated with different costs. Therefore, we compare the prediction performance of three types of algorithms from an economic perspective, namely Artificial Neural Networks, Support Vector Machines, and Hotelling T² Control Charts. We show that the translation of statistical measures into a single cost-based objective function allows optimizing the individual algorithm parametrization as well as the un-ambiguous comparison among algorithms. In a real-life scenario of an industrial full-service provider we derive cost advantages of more than 17% compared to an algorithm selection based on purely statistical measures. This work contributes to the theoretical and practical knowledge on predictive maintenance algorithms and supports predictive maintenance investment decisions.