Parameter Estimation Error in Tests of Predictive Performance under Discrete Loss Functions

We analyze the effect of parameter estimation error on the size of unconditional population level tests of predictive ability when they are implemented under a class of loss functions we refer to as ‘discrete functions’. The analysis is restricted to linear models in stationary variables. We obtain analytical results for no nested models guaranteeing asymptotic irrelevance of parameter estimation error under a plausible predictive environment and three subsets of discrete loss functions that seem quite appropriate for many economic applications. For nested models, we provide some Monte Carlo evidence suggesting that the asymptotic distribution of the Diebold and Mariano (1995) test is relatively robust to parameter estimation error in many cases if it is implemented under discrete loss functions, unlike what happens under the squared forecast error or the absolute value error loss functions.