Variable Selection Based on Statistical Learning Approaches to Improve PM10 Concentration Forecasting

Abstract

In this work, the problem of variable selection for regression is investigated in order to improve the forecasting accuracy. To that effect, the Support Vector Regression (SVR) and the Random Forests (RF) are used to assess the variable importance. Then, a stepwise algorithm is built to select the best subset of predictors. An intensive comparative study is conducted on simulated and real datasets. The real datasets expose the problem of particulate matter concentration forecasting in two monitoring stations from Tunisia. We have proposed a combined approach using SVR and RF for variable importance assessment and for variable selection. We have achieved a significant improvement in forecasts accuracy for the two stations when using only a reduced number of selected predictors.

Keywords: support vector regression, random forests, variable selection, stepwise algorithm, selection bias, particulate matter forecasting