Using Machine Learning to find out When to use Box-Cox Transformation on Time Series Data

Abstract

The Box-Cox transformation improves the normality of the data. This improvement in normality does not guarantee better forecasting results using ARIMA as compared to when the data transformation is not used. So when should one use the transformation on the data to get good forecasting results? The current study tries to answer this by building a predictive model on a set of independent variables which are the characteristics of the data and the dependent variable which tells if use of Box-Cox transformation for forecasting by ARIMA is useful or not. This model gives 64% accuracy and needs improvement. Along with this, another prediction model is obtained which shows from the characteristics of the data whether the 95% prediction intervals obtained by using Box-Cox transformation are better than the intervals obtained by not using the transformation. This model gives 82% accuracy.