A comparative study of predictive algorithms for time series forecasting

Abstract

Forecasting is an important activity in economics, finance, marketing and various other domains like environmental and social sciences. There are several methods for making forecasts, but they all fall into two categories: causal methods and time series methods. In many cases, predictive algorithms implementing time series are good candidates for forecasting. In this paper we run a comparative study of three of these algorithms: Linear Regression, Support Vector Machines and Multilayer Perceptron in order to determine their performances in term of implementing times series for predictive systems. To assess the performance of these algorithms, we have conducted experiments over four representative datasets. The results exhibit that linear regression produced the best forecasts. The other two algorithms show a good behavior as well.