This paper endeavours to show how the specification of the regression testing the uncovered interest parity (UIP) condition can determine whether or not the hypothesized proportional relationship between international interest rate differences and exchange rate changes is rejected. Across major currencies, various terms to maturity, different data frequencies and the short as well as the long time horizon, single‐equation regressions partly reject the UIP condition. However, this ‘UIP puzzle’ tends to disappear when panel data regressions account, for example, for risk premiums by means of two‐way unobserved component specifications with random or fixed effects for both currencies and time periods. The closest concurrence with the UIP condition arises when specifying the time‐specific component as fixed effect, which provides a way to address the potential bias when unobserved exchange rate risk premiums correlate with interest rates. Copyright © 2016 John Wiley & Sons, Ltd.