| Abstract: |
Empirical research on oil price dynamics for modeling and forecasting purposes
has brought forth several unsettled issues. Indeed, statistical support is
claimed for various models of price paths, yet many of the competing models
differ importantly with respect to their fundamental temporal properties. In
this paper, we study one such property that is still debated in the
literature, namely mean-reversion, with focus on forecast performance. Because
of their impact on mean-reversion, we account for non-constancies in the level
and in volatility. Three specifications are considered: (i) random-walk models
with GARCH and normal or student-t innovations, (ii) Poisson-based
jump-diffusion models with GARCH and normal or student-t innovations, and
(iii) mean-reverting models that allow for uncertainty in equilibrium price
and for time-varying convenience yields. We compare forecasts in real time,
for 1, 3 and 5 year horizons. For the jump-based models, we rely on numerical
methods to approximate forecast errors. Results based on future price data
ranging from 1986 to 2007 strongly suggest that imposing the random walk for
oil prices has pronounced costs for out-of-sample forecasting. Evidence in
favor of price reversion to a continuously evolving mean underscores the
importance of adequately modeling the connvenience yield. |