Forecasting Inflation Using News Feeds: BERTopic vs. LDA

Forecasting inflation is an important applied problem. One way to improve inflation forecasting is to extract features from news, since, firstly, news may contain information about important events that directly or indirectly affect inflation, and secondly, news may affect inflation expectations. One of the most popular methods of extracting features from text in economic literature is topic modeling. This paper is devoted to a comparative analysis of two topic extraction methods - BERTopic, which is based on neural networks and used for the first time in this paper, and the more popular LDA, which is based on the analysis of the frequency distribution of words in texts - for forecasting inflation in Russia using data from popular Telegram channels. Using the example of forecasting CPI components, it was shown that using topics extracted with BERTopic gives better forecast metrics compared to topics extracted with LDA. In addition, BERTopic itself is more convenient to work with, since it is a more modern method of natural language processing. The main contribution of this paper is to demonstrate that a simple replacement of the topic modeling method can improve the quality of inflation analysis.