Predictive regressions are often applied to test the predictive power of economic variable for future asset returns. Inference in such regression is complicated by the fact that predictors may display persistent, near-unit root behaviour.
In such cases, the null distribution of t-ratios is poorly approximated by the standard, normal and depends strongly on nuisance parameters.
One approach to resolve this problem was developed by Jansson and Moreira (Econometrica, 2006), who analyse a likelihood conditional on the observed information, leading to a uniformly most powerful conditionally unbiased test.
This paper provides an explanation of the relatively low power of this test in Monte Carlo simulations. We show that the type of conditioning by Jansson and Moreira involves a loss of statistical information on the predictive parameter; we try to quantify this information loss and its dependence on the parameters.
This analysis favours unconditional, (approximately) similar tests, based on the concept of approximate least-favourable distributions (Elliott, Mueller and Watson, Econometrica, 2015).
This is free event. Bring your colleagues, friends and classmate along.
Peter is Professor of Financial Econometrics in the Faculty of Economics and and Business at the University of Amsterdam and Fellow of the Tinbergen Institution.
His research focuses on;
- Unit roots and cointegration
- asymptotic theory
- semi(non) parametric and adaptive estimation and testing
- continuous record asymptotics
- multivariate GARCH
- stochastic volatility models
- econometrics of option pricing
- high-frequency data
- bounded rationality