Production planners of customer-specific Engineering-to-Order products face the challenge of accurately forecasting assembly times in order to ensure efficient production planning. Due to individual customer requirements, high deviations between planned and actual assembly times observed. Traditional methods are often based on empirical values, which are subject to subjective judgement and are highly prone to error. An approach based on historical data using machine learning is shown in this paper and approach is validated at a Machine Tool manufacturing company. The approach consists of a supervised machine learning algorithm combined with historical order and assembly times. The data preparation includes various feature engineering methods for categorical data and dimension reduction methods for high-dimensional data sets in order to extract relevant influencing factors. For modelling, a three-stage process is carried out to find the best model with the best data preparation. This process includes a solvability check, a selection of the best feature engineering methods and a final model optimization. Various algorithms are evaluated for its suitability of this problem, such as random forest, gradient boosting, stacking models and artificial neural networks. The results show that gradient boosting achieves the highest forecast accuracy. The average plan-actual deviation and thus the assembly planning accuracy of 81.12% was achieved. Overall, the findings show machine learning processes are able to significantly improve the forecasting accuracy and therefore support data-driven decision-making in production planning.