Econometrics
Financial Economics

Econometrics and Financial Economics (oral) 2000
Microeconomics and Macroeconomics (written) 1999

Genetic Programming Techniques for Identifying Structural Models Applied to the Earnings-Returns Relation

May 2004

Ph.D. (in progress), Economics, Yale University
B.A., Economics and Mathematics, Amherst College, 1998

Economics Department Fellowship, Yale University, summer 2002
Yale University Dissertation Fellowship, fall 2002
Yale University Graduate Student Fellowship, 1998-2002
Dwight Hitchcock Memorial Fellowship, Amherst College, 1998/99, 2000/01, 2003/04
Bernstein Prize for Best Senior Thesis in Economics, Amherst College, 1998
Phi Beta Kappa, Amherst College, 1998
Nelson Summer Research Fellowship in Economics, Amherst College, summer 1997

Teaching Assistant: Microeconomics (2000, 2001), Econometrics (2002)
Tutored a blind undergraduate student in Economics and Mathematics courses (1998-2002)

Summer Associate, National Economics Research Associates, New York, summers 2001, 2002

Search for an Empirical Specification of the Returns-Earnings Relation, 2003 [job market paper]

Short-Term Employment Dynamics, Adjustment Costs, and Firm Exit, 2001

Institutional Parameters and Workforce Dynamics: A Simulation and Empirical Approach, 2000 (with Walter Nicholson, Amherst College)

Capital Creation and Destruction in an Embodied Growth Model with Heterogeneous Firm Size, 1999

"Nonlinear Model Estimation using Genetic Programming" 11th Symposium of the Society for Nonlinear Dynamics and Econometrics, Florence, Italy, March 2003

The task of model identification is present in most applied econometric exercises. In the simplest case, researchers are faced with empirical data for two economic variables and would like to formulate a structural model characterizing their relationship. When the true data generating process is nonlinear, the inadequacy of linear regression models is apparent as they suffer from the usual symptoms of low explanatory power and biased coefficients. Yet economic theory provides few directions for using a particular functional form when linear regression fails. The selection of appropriate functional form rests with the researcher and has profound implications for the consistency and significance of estimated model parameters and the predictions obtained from the model.

I advocate the use of a flexible parametric estimation approach based on genetic programming (GP). Genetic programming is a stochastically driven optimization algorithm. In contrast to traditional hill climbing algorithms that proceed from one point in the search space to another, GP operates simultaneously on a set of points in the parameter space. Each point in this set is equivalent to a function-a possible data generating process for the data. The algorithm calculates the degree of fit of each function to the data and uses it to rank this function with respect to other functions already explored by it. It then constructs a new set of points in the search space using the expressions contained in the current set of functions. Theory shows that the GP algorithm is intrinsically parallel, i.e., by calculating the fit of every function it implicitly accumulates information about the fit of every element comprising this function. The algorithm explores the search space in an optimal way. More precisely, as the GP algorithm constructs new sets of functions, the elements with high degree of fit occur at an exponentially increasing rate in the newly generated functions while the occurrence of elements with low fit to the data decays exponentially. This allows the algorithm to expend optimal computational effort by spending its calculation time increasingly on functions containing valuable elements.

In the first part of the dissertation I develop and code a genetic programming algorithm specially suited to analyzing time series data. It is specifically designed to incorporate lagged variables and errors in variables explicitly. This is an advance over GP work since errors in variables have been treated as an obstacle in past research, causing a reduction in its performance. This is the first attempt to model errors explicitly within the GP framework.

In the second part of the dissertation the performance of the GP algorithm is evaluated over several test cases and its out-of-sample forecasting ability is evaluated with respect to other nonlinear modeling frameworks. I show that the increased flexibility of GP over traditional parametric methods comes at a relatively small cost. The rate of convergence of the GP estimator is better than that of several nonparametric estimators. A simulation exercise is conducted in order to derive the distributions of model parameters for several particular classes of data generating processes under the null hypothesis that genetic programming uncovers the true functional form up to nuisance parameters.

The third part of the dissertation uses the algorithm to model how stock prices react to unanticipated accounting earnings. I suggest several nonlinear parametric specifications of the reaction of excess current-period stock price returns to the unanticipated component of quarterly earnings. My motivation is the well-known misspecification problem in studies that ignore nonlinearity, as it presents a source of systematic error in earnings response regressions (see Freeman and Tse 1992 and Beneish and Harvey 1998, for a discussion of nonlinearity in this empirical context). My major findings are that: 1) using the described model identification procedure I am able to confirm the existence of a nonlinear model that has superior in-sample and out-of-sample predictive power over the linear model; 2) even if transitory items from earnings are excluded, the unexpected component still has a nonlinear effect on price, but the dominance over the linear model diminishes out-of-sample; 3) the approach presents an advantage over alternative nonparametric techniques as it can be used to derive a parametric model suitable for further empirical work on the earnings-returns relation.