In the last decades, machine learning has proven its ability to outperform human experts in a various fields. Its prominent performance is influenced by many if not all of it’s components. Among others, data that used for training and prediction is one of the most influencing component of a machine learning framework. Furthermore, the number of data source is sometimes too much and it is not practical to use all data source for training, since it will slow down the training process and demands large store space. The main goal of this thesis is to find a way to automatically choose good data source from a large of available data sources for a given regression task. In particular the data is in form of time series, which are gathered from sensors distributed in various places. At the end, the framework with automatically chosen data sources under certain budget constraints should be comparable against the same framework with unrestricted access of data sources in terms of prediction performance, e.g., MSE and MAE.