Smart Modelling and Large Data Sets

In recent decades, semiparametric and nonparametric models have received increasing interest, which can be explained by the desire to get away from the strong restrictions of parametric models. Although their rate of convergence is slower, semiparametric and nonparametric models offer greater flexibility for estimation. This thesis proposes to use these models for respectively economic and econometric modelling in chapter one and two and to provide a solution to the distributed data problem in chapter three. In the first chapter, we use a general additive semiparametric model to estimate the long run efficiency of offshore wind farms. We rely on mainly well-established nonparametric methods that we had to modify appropriately to fit with the economic model we wanted to estimate. In the second chapter, we use a varying coefficient semiparametric model to estimate Regional Knowledge Production Function. We rely on general ideas of economic and econometric modelling of bilateral trade by gravity models and develop semiparametric estimators that could estimate such a sophisticated model structure. Finally in the third chapter, we start from the data distributed problem. We propose an operational way to get timely estimates, fits or predictions with huge but distributed data sets, including model and parameter selection. Our approach is to think fully locally using local linear nonparametric estimation with LASSO penalty for statistical analysis, may it be estimation, prediction, or attribution.