The Economic Value of Words: an Evaluation of News for Economic Analysis

In this study, we use various extraction techniques to analyze the polarity of economic topics sourced from Mexican news outlets. For this, we constructed a comprehensive daily dataset spanning eighteen years, covering a diverse range of topics. Following a standard data cleaning process, we utilized Latent Dirichlet Allocation (LDA) to identify main economic-related concept. Subsequently, we applied a combination of rule-based, lexicon-based, and pre-trained language models to determine the polarity and sentiment tone from sentences or document level. To assess the effectiveness of our sentiment approach, we explored traditional word representation, including Word2vec, GloVe, and FastText integrated into a Multilayer Perceptron classifier (MLP). Finally, we examine the relationship between the derived sentiment index and the Consumer Price Index (CPI). We observe that the lexicon achieved higher precision (88.39%), outperforming the other word embedding techniques.