MSc Thesis Defense by Alberto La Mantia
Title: Optimization of Econometric Models using Evolutionary Algorithms
Econometric models are created by economic analysts in order to forecast the behaviour of the sales in the market. This thesis was pursued at the Insight Group Nordic and considers the optimization of real-world econometric regression models developed by the company. These regression models are fitted based on historic data using an hybrid method followed by ordinary least squares (OLS), where the hybrid method employs derivative-free optimizations.
So far, the company employs a Genetic Algorithms and a Numerical Gradient Descent algorithm for the derivative-free optimizations. In this thesis, we compare several alternatives for this task: A Genetic Algorithm, Numerical Gradient Descent, and Covariance Matrix Adaption Evolution Strategy (CMA-ES) in two variants – with and without restart with increasing populations (IPOP). The mean-squared-error (MSE) after OLS is used to assess the accuracy of the models.
Experiments have been conducted on two distinct econometric models, combining the algorithms in five configurations: CMA-ES, IPOP, Genetic Algorithm followed by Numerical Gradient Descent, Genetic algorithm followed by CMA-ES, and finally CMA-ES followed by Numerical Gradient Descent. The two models have been selected among real models generated by Insight Group Nordic.
Supervisor: Christian Igel
Censor: Carsten Witt, DTU