A test procedure based on ranks is suggested to test for nonlinear cointegration. For two (or more) time series it is assumed that there exist monotonic transformations such that the normalised series can asymptotically be represented by independent Brownian motions. Rank test procedures based on the difference between the sequences of ranks are suggested. If there is no cointegration between the time series, the sequences of ranks tend to diverge, whereas under cointegration the sequences of ranks evolve similarly. Monte Carlo simulations suggest that for a wide range of nonlinear models the rank tests perform better than their parametric competitors. To test for nonlinear cointegration a variable addition test based on ranks is suggested. As empirical illustrations we consider the term structure of interest rates. Only weak evidence for a nonlinear long run relationship between interest yields of bonds with different time to maturity is found.
unit roots nonlinear cointegration rank test arcsine distribution