A Bayesian estimation procedure is developed for estimating multiple-regime
(multiple-threshold) error-correction models appropriate for deviations fr
om financial arbitrage relationships. This approach has clear advantages ov
er classical stepwise threshold autoregressive analysis. Unlike many other
applications of threshold models, the knowledge of some costs involved in s
etting up arbitrage positions allows us to specify an informative prior. To
illustrate the Bayesian procedure, we estimate a no-arbitrage band within
which index futures arbitrage is not profitable despite (persistent) deviat
ions from parity.