Changing Focus of the FOMC Through the Financial Crisis John E. Miller Computer & Information Sciences University of Delaware Newark, DE 19711 jmiller@udel.edu Kathleen F. McCoy Computer & Information Sciences University of Delaware Newark, DE 19711 mccoy@udel.edu Abstract The financial crisis of 2007-8 with effects continuing to this day is a complex beast. We use Latent Dirichlet Allocation (LDA) topic analysis and time series analysis to examine behavioral changes in the Federal Open Market Committee (FOMC) during 2005-08. We are the home fix-it guy with a new pair of pliers (topic analysis) and we will try it out on anything that needs fixing such as “Who and what was the financial crisis.” It’s not the only tool for the task, but it is a fun and informative tool to use. 1 Introduction We investigate the financial crisis of 2007-8 us- ing natural language processing tools to focus on “who and what was the financial crisis” (PoliIn- formatics, 2014). We are not very good at color- ing within the lines and we have a new pair of pli- ers, LDA topic analysis, so we apply topic analysis to the Federal Open Market Committee (FOMC, 2014) transcripts of 2005-08 to see how FOMC behavior changed over time. Our analysis process consists of: • Preparation of FOMC transcripts, • LDA topic analysis, • Mean square successive differences to iden- tify changing topic mixtures, and • Sequence plots of topic mixture proportions. Our work products include: FOMC topics, se- quence plots of topic proportions, interpretation of findings, and critique of our procedure. 1.1 LDA Topic Analysis LDA topic analysis uses a per document bag of words approach to determine topic compositions of words and document mixtures of topics. Fig- ure 1 (Steyvers and Griffiths, 2007) shows a cor- pus explained as the product of topic compositions (Φ) and document mixtures (Θ). Analysis con- structs topics and mixtures of topics by assigning words to topics within documents. Topic compo- sitions are interpreted as topics or themes of doc- uments, conversations, or discussions. Document topic mixtures can be examined to see how doc- ument mixture proportions vary over time. Topic analysis reduces the dimensionality of a corpus by orders of magnitude from millions of words to fre- quency distributions of hundreds of topics. Corpus words documents words topics documents topics = x   Figure 1: Topic Model LDA topic analysis (Blei, 2012; Blei et al., 2003) is based on a generative probabilistic model where document mixtures and topic compositions are generated according to multinomial probabil- ity distributions Θ and Φ respectively. Analysis re- verses the generative model, calculating weighted document topic mixtures Θ and topic word com- positions Φ from the corpus. The topic analysis implementation used in this study borrows from UMass Mallet (McCallum, 2002). 1.2 Transcript Preparation and Analysis Document preparation of the FOMC transcripts included: • Clean the corpus and segment transcripts into ≈2000 word documents. • Transform currency and numeric amounts, dates and durations, infrequent words, stop- words, and punctuation into reserved words. • Form bigrams of word sequences with counts > 1 / 3 of the individual word counts; drop stop-word and punctuation reserve words.