On the robustness of cointegration tests when assessing market efficiency

The finding that spot and lagged forward exchange rates are cointegrated with the vector (1,−1) is often given as evidence for long-run market efficiency. This paper examines the conjecture that a small unit root or fractionally integrated component in the relationship is dominated in finite samples by a large stationary component. Monte Carlo techniques demonstrate that with typical sample sizes and variable magnitudes, the Engle-Granger test overwhelmingly finds spurious cointegration. Conversely, under certain conditions, Johansen tests are shown to be relatively robust to differences in variable magnitudes. Although developed from recent empirical work in the forward currency market this result clearly has relevance for the use of predictive regressions in any asset market.