Browsing by Author "Koskinen, Juhani"

Traditional empirical research in economics is focused on statistical inference. Therefore, economists rarely pay attention to model accuracy. Model accuracy is important when the research question is focused on prediction. Many questions in economics can be framed as a prediction problem. Thus, it is necessary to utilize techniques aimed at achieving prediction accuracy. Machine learning is a field that focuses on prediction accuracy. This thesis aims to explain the core concepts behind machine learning and how these can be applied in economics. The concepts introduced in this paper are used to predict house prices. The data are an open source dataset containing over 20000 records of houses sold in Kings County, USA. Three different models are trained on data that exclude variables describing the location of the house: linear regression, decision tree, and random forest. In addition, decision tree, and random forest are also trained on data that contain location variables. The models’ accuracies are examined by making predictions on a subset of data that was not used to train the models. Random forest performs the best as measured by the mean squared error. Using data without location variables, the mean squared error achieved by the random forest is 0.074. Linear regression and decision tree both achieve a mean squared error of 0.090. When location variables are added, the mean squared error of the random forest and decision tree drop to 0.028 and 0.045, respectively. The results suggest that the techniques utilized in machine learning can be applied successfully to economics. Machine learning excels at finding complex interactions between different variables in large amounts of data. As the datasets used in empirical economic research grow in size and complexity, machine learning can provide valuable tools to economists.