Specification tests of calibrated option pricing models

In spite of the popularity of model calibration in finance, empirical researchers have put more emphasis on model estimation than on the equally important goodness-of-fit problem. This is due partly to the ignorance of modelers, and more to the ability of existing statistical tests to detect specification errors. In practice, models are often calibrated by minimizing a loss function of the differences between the modeled and actual observations. Under this approach, it is challenging to disentangle model error from estimation error in the residual series. To circumvent the difficulty, we study an alternative way of estimating the model by exact calibration. Unlike the error minimization approach, all information about dynamic misspecifications is channeled to the parameter estimation residuals under exact calibration. In the context of option pricing, we illustrate that standard time series tests are powerful in detecting various kinds of dynamic misspecifications. Compared to the error minimization approach, exact calibration yields more reasonable model comparison result, and delivers more accurate hedging performance that is robust to both gradual and abrupt structural shifts of state variables.