14 May
9:30
Symbolic Regression - Model Recovery and Applications to Economics/Finance
Symbolic regression (SR) is an estimation method aimed at finding simple, interpretable equations in a fully data-driven manner. We investigate under which settings SR algorithms are expected to recover a "true" model. SR is formulated as a solution to a model selection problem, and we formalize the discussion on whether true model recovery is achievable. For the case of estimating an unknown nonlinear conditional expectation function, exhaustive symbolic regression (ESR) is conjectured to select the true model in large samples, provided the true model is contained in the search space and a consistent model selection criterion is used. Within the PAC-learning framework, ESR can be framed as a structural risk minimization (SRM) problem, and we argue that ESR is non-uniformly PAC-learnable under mild assumptions. Various applications of SR to economics and finance are proposed.