This paper argues that inferring long-horizon asset return predictability f
rom the properties of vector autoregressive (VAR) models on relatively shor
t spans of data is potentially unreliable. We illustrate the problems that
can arise by reexamining the findings of Bekaert and Hodrick (1992), who de
tected evidence of in-sample predictability in international equity and for
eign exchange markets using VAR methodology for a variety of countries from
1981-1989. The VAR predictions are significantly biased in most out-of-sam
ple forecasts and are conclusively outperformed by a simple benchmark model
at horizons of up to six months. This remains true even after corrections
for small sample bias and the introduction of Bayesian parameter restrictio
ns. A Monte Carlo analysis indicates that the data are unlikely to have bee
n generated by a stable VAR. This conclusion is supported by an examination
of structural break statistics. We show that implied long-horizon statisti
cs calculated from the VAR parameter estimates are very unreliable.