Abstract—In finance, the flow of news is constantly updated in a way that changes the investors’ understanding, influences their sentiment, and hence, shapes the financial markets (Asset returns, Volatility, Interest rates, etc.). Thus, one of the most important challenges that an investor has to overcome is to extract useful information from text data (News) in order to be used in decision making after reviewing all the literature dealing with the analysis of financial news. In this paper, we introduce an automated methodology to extract relevant words and topics from the FOMC (The Federal Open Market Committee) Minutes reports. Using the Latent Dirichlet Allocation (LDA) algorithm to track the main topics in the FOMC reports especially the “inflation” topic, we made a dimension reduction of our corpus by considering only the first 250 words belonging to this topic. Using the reduced corpus, we applied filters to clean non discriminative words and we employ deep learning in detecting only sentiment-charged words. By the end, we obtained a list of the most relevant words (dictionary) in the FOMC minutes and a prediction (with an accuracy of 98.76% and a correlation of 77%) of the Fed Funds rates.
 

Index Terms—FOMC, classification, LDA, CNN, prediction, news, finance, market, interest rates, neural networks