SPEAKER: Professor Ben Van Roy

ABSTRACT:

When used to guide decisions, linear regression analysis typically  involves estimation of regression coefficients via ordinary least  squares and their subsequent use in an optimization problem. When  features are not chosen perfectly, it can be beneficial to account for  the decision objective when computing regression coefficients.  Empirical optimization does so but sacrifices performance when  features are well-chosen or training data are insufficient. We propose  directed regression, an efficient algorithm that combines merits of  ordinary least squares and empirical optimization. We demonstrate  through computational studies that directed regression generates  performance gains over either alternative. We also develop a theory  that motivates the algorithm.