Abstract
We propose a two-stage stochastic frontier model that can handle complex non-linear patterns. In the first stage, we apply a panel data neural network to predict the demeaned composed error term. In the second stage, we apply traditional Stochastic Frontier Analysis to the residuals to obtain efficiency estimates. To illustrate our methodology, we employ quarterly data to estimate the technical efficiencies of large US banks from the first quarter of 1984 to the second quarter of 2010. The mean efficiency of US banks during this time period is 93.97%. The second quarter of 2004 through the fourth quarter of 2008, the median efficiencies of these banks are significantly lower than the overall average, with an average of 87.86%. This is in line with the financial conditions experienced during this time period.
JEL Classification: C23, C45, D24, G21.