Inductive Learning and Theory Testing: Applications in Finance
MetadataShow full item record
CitationZimmermann, Tom. 2015. Inductive Learning and Theory Testing: Applications in Finance. Doctoral dissertation, Harvard University, Graduate School of Arts & Sciences.
AbstractThis thesis explores the opportunities for economic research that arise from importing empirical methods from the field of machine learning.
Chapter 1 applies inductive learning to cross-sectional asset pricing. Researchers have documented over three hundred variables that can explain differences in cross-sectional stock returns. But which ones contain independent information? Chapter 1 develops a framework, deep conditional portfolio sorts, that can be used to answer this question and that is based on ideas from the machine learning literature, tailored to an asset-pricing application. The method is applied to predicting future stock returns based on past stock returns at different horizons, and short-term returns (i.e. the past six months of returns) rather than medium- or long-term returns are recovered as the variables that convey almost all information about future returns.
Chapter 2 argues that machine learning techniques, although focusing on predictions, can be used to test theories. In most theory tests, researchers control for known theories. In contrast, chapter 2 develops a simple model that illustrates how machine learning can be used to conduct an inductive test that allows to control for some unknown theories, as long as they are covered in some way by the data. The method is applied to the theory that realization utility and nominal loss aversion lead to the disposition effect (the propensity to sell winners rather than losers). An inductive test finds that short-term price trends and other features of the price history are more important to predict selling decisions than returns relative to purchase price.
Chapter 3 provides another perspective on the disposition effect in the more traditional spirit of behavioral finance. It assesses the implications of different theories for an investor's probability to sell a stock as a function of the stock's return and then tests those implications empirically. Three different approaches that have been used in the literature are shown to lead to the, at first sight, contradictory findings that the probability to sell a stock is either V-shaped or inverted V-shaped in the stock's return. Since these approaches compute different conditional probabilities, they can be reconciled, however, when the conditioning set is taken into account.
Citable link to this pagehttp://nrs.harvard.edu/urn-3:HUL.InstRepos:17467320
- FAS Theses and Dissertations