Athens University Economics and Business (AUEB)
Accounting for model misspecification in financial decisions: a loss-based approach
Abstract
dc:descriptionThe need for construction of procedures robust to various sources of uncertainty is widely acknowledged by both academics and practitioners. Financial institutions, hedge funds and specialists on the financial profession base their decisions on statistical models that associate past information and prior views into forecasts. Despite the fact that statistical models often present a necessity for the description of a complex reality, one cannot neglect that in most cases researchers face restrictions in the form of working with finite samples and dealing with various types of misspecification.Asset allocation is a typical research area in financial economics where model specification and estimation error have serious implications on investor decisions. This is mainly attributable to the difficulty of specifying, estimating and forecasting the conditional distribution of asset returns that is employed in the construction of optimal portfolio rules.In this research study, we step back from the standard econometric treatment of portfolio choice problems and we investigate an alternative econometric procedure that is directly related with the preferences of the user of statistical models. In particular, instead of using likelihood-based or squared error loss functions to infer the parameters of a possibly misspecified model for asset returns we focus on criterion functions derived from a representative investor's decision problem. In that way, we are able to construct frequentist or Bayesian-type decision-based estimators accounting for inconsistencies in forecast evaluation and estimation stages of the modeling process. Our analysis progresses around three key objectives: We first conduct an extensive empirical investigation for a single period decision framework where an investor has to choose between an aggregate stock index and cash, accounting at the same time for different utility preferences and investment constraints. Second, extend our methodology to incorporate horizon effects and multiple assets in a dynamic intertemporal portfolio choice setting. Finally, we revisit the Meese and Rogoff puzzle concerning exchange rates forecasting accuracy from the perspective of an investor that allocates her wealth between foreign and domestic risk free bonds.We opt to work in such settings for several reasons. The most important one is that the implications of model misspecification have severe consequences on the optimal portfolio choice of investors and hence on their utility loss and wealth. In addition, there are vast time series data and increased availability of records of bias corrected data for a variety of already established asset classes. This enables us to test our estimator efficiently and compare its properties with the standard estimators proposed in the literature. Nevertheless, decision-based estimators can extend naturally to various policy-making frameworks where the decision maker's objectives differ from the statistical loss criteria used to estimate decision rules. It is even possible to apply the methodology to non-financial decision settings, e.g. central banks’ reaction functions, as long as certain appropriate conditions about the loss function are satisfied.Our main findings are summarized as follows: First, decision-based parameter estimates account for misspecification in time varying higher order moments of portfolio returns according to their significance in the loss function of the investor. Second, in the presence of specification error, a utility-based criterion function penalizes asymmetrically the parameters that imply association of equity returns and exchange rates with economic state variables. Finally, the decision-based estimator outperforms commonly used estimators with gains increasing with the investment horizon and the number of assets in the investment universe. Overall, our research provides further empirical evidence that the standard practice of grounding asset allocation decisions on possibly misspecified models estimated with statistical accuracy criteria entails high utility costs. Consequently, our findings are consistent with the argument in the forecasting literature that the appropriate evaluation criteria should be an integral part of the estimation method for the parameters of a possibly misspecified model. With that in mind, likelihood-based estimators are by construction not capable of capturing specification error in portfolio rules thus leading to increasingly sub-optimal portfolio policy functions.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2018
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Giannakidis, Charilaos
- Γιαννακίδης, Χαρίλαος
Subjects
dc:subject × 12Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/46390