An Investigation of Metrics to Compare Significantly Different Approaches to Time Series Forecasting

This article briefly discusses two fundamentally different approaches to time series forecasting. The first method is based on universal coding. It has been studied in various real-life problems and shown to be effective. The second is based on a ten-headed finite state machine, which effectively predicts multilinear sequences. As a result of work the the considered methods of forecasting time series show the result that are incomparable with each other. This makes their direct comparison difficult. The authors propose to convert the results into one form and use widely used metrics for comparison, such as Mean Absolute Error (MAE), Mean Square Error (MSE) and others. This standardization will facilitate a comprehensive comparison of the two methods, revealing their relative strengths and weaknesses across diverse datasets. Also description of the application of metrics to forecasting methods is given. The result of the work is a conclusion on the possibility of using one or another metric.