Nonparametric Forecasting of the Manufacturing Output Growth with Firm-level Survey Data
A large majority of summary indicators derived from the individual responses to qualitative Business Tendency Survey questions (which are mostly three-modality questions) result from standard aggregation and quantification methods. This is typically the case for the indicators called balances of opinion, which are the most currently used in short term analysis and considered by forecasters as explanatory variables in linear models. In the present paper, we discuss a new statistical approach to forecast the manufacturing growth from firm-survey responses. We base our predictions on nonparametric forecasting algorithms inspired by statistical pattern recognition, such as the k- nearest neighbors and random forest regression methods, which are known to enjoy good generalization properties. Our algorithms exploit the heterogeneity of the survey responses, work fast, and allow the treatment of missing values. Starting from a real application on a French data set related to the manufacturing sector, we argue that these procedures lead to significantly better results than more traditional competing methods.