Learning about out -of -sample predictability and its impact on real -time investment decisions

Huseyin Gulen, Purdue University

Abstract

This paper develops an empirical framework that allows the degree of out-of-sample predictability in any specific dataset to serve as conditioning information in portfolio choice and market timing decisions. The forecasting model is developed from a real-time investment perspective by allowing investors to learn about out-of-sample predictability before making investment decisions. Learning is modeled in a recursive forecasting framework where the investment decisions are based on direct evidence of out-of-sample predictability prior to the decision period. In market timing problems, forecasting models are chosen based on their out-of-sample performance. In portfolio choice problems, at each decision period, a trading portfolio is formed using only the assets that are predictable out-of-sample. This simple adjustment leads to a forecasting methodology that has a minimum amount of look-ahead bias, reduces investors' judgment errors via learning, improves investment performance, and increases the size and power of the performance tests. To assess the usefulness of information about out-of-sample predictability, simulations are performed at different levels of predictability. When there is no predictability, active investment strategies based on conditioning information of out-of-sample predictability converge to a simple buy-and-hold strategy. Thus, the investors learn to buy-and-hold in the absence of predictability. When there is predictability, however, the active strategy outperforms the buy-and-hold strategy by identifying the sources of out-of-sample predictability. These results are robust to using different out-of-sample performance measures and model selection procedures. The value of using out-of-sample predictability as a basis of choice is shown to be useful especially at higher transaction cost levels. The findings of the paper have important implications for tests of conditional asset pricing models, portfolio performance measures, and real world investment practices.

Degree

Ph.D.

Advisors

McConnell, Purdue University.

Subject Area

Finance

COinS