The forecasting of time series values plays an important role in many areas. The forecast values for time series are generally determined using statistical models, which require a long and consistent history. If this is not the case, machine learning approaches can be applied to create predictive model for time series. Regression tree and linear model are the most commonly used predictive models in machine learning community. Random forest offers competitive prediction performance as an ensemble of regression trees. This thesis aims to provide a comprehensive analysis of the applicability of random forest and linear model to create a predictive model for time series. Here, random forest is presented as a novel way to build a predictive model. Moreover, the improvements to increase the accuracy of this model, with the help of linear model, are covered from three different perspectives:

  • Selection of the relevant features
  • The bias correction
  • Ensemble of models

The accuracy of the random forest has improved by using the relevant features and assembling with linear model. The ensemble of linear model and random forest has a better accuracy than either the linear model, or the random forest alone.