Time series forecasting is an essential tool in many fields. In recent years, machine learning has gained popularity as an appropriate tool for time series forecasting. When employing machine learning algorithms, it is necessary to optimise a machine learning pipeline, which is a tedious manual effort and requires time series analysis and machine learning expertise. AutoML (automatic machine learning) is a sub-field of machine learning research that addresses this issue by providing integrated systems that automatically find machine learning pipelines. However, none of the available open-source tools is yet explicitly designed for time series forecasting. The proposed system TSPO (Time Series Pipeline Optimisation) aims at providing an autoML tool specifically designed to solve time series forecasting tasks to give non-experts the capability to employ machine learning strategies for time series forecasting. The system utilises a genetic algorithm to find an appropriate set of time series features, machine learning models and a set of suitable hyper-parameters. The optimisation objective is defined as minimising the obtained error, which is measured with a time series variant of k-fold cross-validation. TSPO outperformed the official machine learning benchmarks of the M-Competition in out of randomly selected time series. TSPO captured the characteristics of all analysed time series consistently better compared to the benchmarks. The results indicate that TSPO is capable of producing robust and accurate forecasts without any human input.