Realization of Regression Methods of Demand Forecasting Using the R Language

Abstract

Regression analysis is considered as a key method for forecasting the magnitude of demand of goods. The list of methods that are the most effective for calculating the forecast estimate is presented: linear regression with regularization, regression on the basis of support vectors, random forest method. Necessary calculations are implemented in the programming language R, using both the basic functional and additional packages, which make it possible to use the methods in question. As input data, the store performance and product characteristics are used. The metric of the quality of the result of the operation of the algorithms is determined, i.e., the mean squared error. The sample of data is divided into training and test data, the results for each of the above algorithms are calculated in sequence. Conclusions are drawn that for the sample under consideration, the random forest algorithm yields the best result. The degree of correlation between forecasts for different algorithms is derived, on the basis of which an assumption is made about possible joint use of forecasts. Proceeding from this, the simplest combination of algorithms is constructed, i.e., the arithmetic mean. This ensemble of algorithms turned out to be better than all the considered methods of forecasting, taken separately. A plan for further research on the creation of a committee of algorithms based on methods of bugging, boosting or stacking is determined.