We evaluate the asymptotic and finite-sample properties of direct multi-step
estimation (DMS) for forecasting at several horizons. For forecast
accuracy gains from DMS in finite samples, mis-specification and
non-stationarity of the DGP are necessary, but when a model is
well-specified, iterating the one-step ahead forecasts may not be
asymptotically preferable. If a model is mis-specified for a non-stationary
DGP, in particular omitting either negative residual serial correlation or regime
shifts, DMS can forecast more accurately. Monte Carlo simulations
clarify the non-linear dependence of the estimation and forecast biases on
the parameters of the DGP, and explain existing results.
CHEVILLON, G. et HENDRY, D. (2005). Non-parametric Direct Multi-Step Estimation for Forecasting Economic Processes. International Journal of Forecasting, 21, pp. 201-218.