This article estimates a dynamic stochastic cost frontier, in which firm-specific technical inefficiency levels are assumed to follow an autoregressive (AR) process due to the existence of quasi-fixed inputs. The generalized method of moments method has to be applied for obtaining consistent parameter estimates under the framework of dynamic panel data. Evidence is found that an AR (2) specification of technical inefficiency is necessary to whiten the random disturbance of the cost frontier and in turn to solve the problem of endogenous regressors. Private banks adjust their technical inefficiency toward the efficiency frontier faster than public banks. Sample banks exhibit scale and scope economies on the basis of the dynamic cost frontier, while constant returns to scale and lack of product mix economies are derived using the static cost frontier.