Density forecasting and momentum

Abstract

We fit 800 time series models to daily momentum returns in an out-of-sample exercise. We apply the Akaike Information Criterion for model selection and forecast the one day-ahead probability density function. Our findings show that a skewed and heavy-tailed density performs best, while a simple GARCH(1,1) specification for the conditional variance is picked most often. We also report the usefulness of low order ARMA models for the conditional mean. Our trading algorithms demonstrate that targeting downside risk substantially outperforms volatility scaling. Furthermore, a strategy designed to have an exposure to momentum which is linear in the one day-ahead Sortino ratio forecast generates an annualized Carhart four-factor alpha of 107.38%, in the absence of transaction costs.

Date of Award	15 Jan 2017
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	José Faias (Supervisor)