摘要
When dealing with textual data, previous studies mainly used a keyword-based matching method to construct indices. The economic policy uncertainty (EPU) index proposed by Baker et al. (2016) is an example. In this paper, we argue that due to its neglect of semantics, such keyword matching generates excessive noise, which affects the index quality and further leads to incorrect inferences. We investigated several neural network models and selected the best-performing classifier to remove the noise caused by keyword matching. Our empirical results revealed that the de-noised EPU index is useful in predicting economic variables and generating superior out-of-sample forecasts. Furthermore, the effects of policy uncertainty shocks on core macro variables of interest are consistent with the predictions of macroeconomic theory. Because the proposed approach is a general framework, in the future all keyword matching-based indexes can be improved under the same approach.