Central bank announcements of interest rate trajectories and monetary policy are among the most impactful events to global financial markets. This project uses data science to analyze Federal Reserve policy statements and seeks insight on their sentiment and relationship to financial market variables. By conducting this study, we identify some tangible results of interest but also lay out a basic foundation for a programme of future inquiry.
Our approach to analyzing the FOMC statement is through the lense of textual analysis. Two approaches are used in this project: Sentiment analysis is applied to the FOMC text corpus and used to construct a sentiment time series. That time series is then used as an explanatory variable to compare to real-word financial time series. Time series charting and inear regression are the main tools of that approach.
The second approach is via text classification. The textual analysis of FOMC statements is not exclusive to data science. Financial market practitioners pore over every word, gesture and media interview of the Federal Reserve chairman or chairwoman, governors and District presidents. Some financial news vendors like Bloomberg even publish text comparisons - showing the redlined differences between two FOMC statements. Our approach via text classification attempts to automate the classification of each FOMC statement by several attributes. The attributes include the hawkishness or dovishness of the statement, the FOMC’s opinion of economic growth or labor market health. The method used here is to manually label each FOMC statement and then to train a support vector machine algorithm to predict each attribute.
This paper is organized as follows: Section 2 gives background on the Federal Reserve, the FOMC and past research. Section 3 describes the sources of data: the FOMC statements, the manual labelling of statements, the financial time series. Section 4 performs exploratory data analysis on these data sources. Section 5 conducts two analysis: the first half addresses the machine learning training work to classify statements by 5 attributes. the second half conducts sentiment analysis. Section 6 discusses the results and Section 7 concludes this project.
Our project is inspired and guided by past work in this field. There is a research literature and past work on FOMC statement analysis using data science methods. A related but distinct literature on the financial market impact of central bank communications is also relevant motivation. Moreover, the authors are aware of several financial institutions and companies engaged in the use of machine learning techniques to analyze central bank communications. We will touch on these in turn.
Cannon’s 2015 paper on FOMC sentiment analysis uses the transcripts instead of FOMC statements. He uses the financial dictionary of Loughran-McDonald to construct sentiment and defines it using a bag of words count method. The R package we use to derive sentiment calculates the same metric which is:
\[sentiment(Document)=\frac{PositiveWords - NegativeWords}{PositiveWords + NegativeWords}\] However, he compares the sentiment index against real economic activity proxied by the Chicago Fed National Activity Index – not against financial market variables. A 2011 paper by Lucca and Trebbi analyzes the FOMC statements directly but measure hawkish and dovish tone not sentiment. In addition, their measurement technique is to use search engines (Google or Factiva) to generate a correlation of the word count hits in the search engine’s corpus between the words “hawk” or “dove” and each relevant word or N-gram from the policy statement. They call their approach a semantic orientation method to analyzing the statements. This measurement technique is not reproducible and computationally impractical. Another paper by Schmeling and Wagner (2019) analyzes sentiment in European Central Bank (ECB) policy statements, which are in English, using the Loughran-McDonald dictionary. They find that tone does seem to affect the risk premia of equities through a risk-based channel. I.e. Higher beta stocks respond more to ECB tone than lower beta stocks. They also find that corporate bond credit spreads between BBB and AAA rated bonds tighten when tone is positive. Lastly, Fuksa and Sornette (2012) analyze FOMC Beige Book, minutes and policy statements for sentiment. They find predictive power in the Beige Book which is released 3 weeks before the policy statement. Thus, analyzing Beige Book sentiment could predict FOMC policy actions.
A separate literature on the central bank impact on asset prices finds the FOMC meetings and statements are important. Cieslak, et.al. (2018) find that US and global stock returns are driven by the FOMC meeting cycle. Since 1994 the equity premium has been earned on even numbered weeks of the FOMC cycle. However, Brusa, et. al. (2017) find evidence that no other central banks have the same equity market impact as the Fed.
Our observation period for the FOMC data is 2007-2019. We wanted a period to cover both hawkish and dovish periods. The period should have enough observations to include the 2008 financial crisis and a full business cycle and both useful in statistical estimation for regression analysis.
The selection of our time series data from FRED followed specific criteria. We choose them because of several considerations. First, we need a public source of financial time series data. FRED meets that requirement. Second, we need relevant time series. The choices of the 3 financial time series below meet those requirements.
The US Treasury 10 Year yield is associated with the bellwether fixed income asset. It is possibly the most followed bond yield in the world. The breadth of its historical data is more than sufficient to cover the observation period of our study.
The Russell 1000 Index which we will used below is a liquid and large cross section of the most well-known and large capitalization US stocks. The particular time series is a total return index (dividends are assumed to be reinvested.)
The US Federal Reserve Funds target rates are the policy rates of the FOMC. At each FOMC meeting, these rates are changed. Prior to 2014, the FOMC published a single point estimate of that rate. After 2014, the FOMC decided to published a tight range with an upper and lower bound on the federal funds rate. This gives some latitude for the Open market operations desk to buy and sell Treasuries within this range.
Text classification of FOMC statements is not generally a research objective but we think it is worthwhile. Classification addresses a potential need: Can a machine correctly infer the opinion or direction of forward guidance or policy decisions in a structured text by the FOMC? In this regard, the classification problem for the FOMC is isomorphic to the Ham-Spam classification of incoming emails by an email program. However, the reader may object that FOMC statements are not so voluminous to require automated processing. Our response is that FOMC statements are merely the first baby step in a much larger classification problem: the public communications of all FOMC and Federal Reserve system members. As previously explained, the FOMC members give speeches, publish articles, appear on TV interviews. Moreover, FOMC meeting minutes are released several weeks after the policy statement is released. These are much longer and required more effort to read and digest. Also, the FOMC transcripts released several years after the meeting may run to over 100 pages each. They contain word for word replay of the entire meeting (excluding private discussions). Lastly, there are at least 16 relevant central banks around the world of interest. Although the Fed is the world’s more important central bank, the ECB, Bank of England, Bank of Japan, Bank of China, Bank of Australia, Bank of New Zealand, all produce communications. In summary, no single person can read all central bank communications. The ability to extract key messages from plain texts remains a valuable capability.
Our machine learning prediction backtest suggests that automated classification is feasible to detect limited features of a central bank communication. Our algorithm succeeds at detecting medium term rate outlooks, employment growth and policy rate changes. At these tasks, we have attained accuracy rates between 90-100 percent. The most challenging attribute to understand is inflation (64.5%). This is consistent with financial practitioner opinion. Inflation is the most complex of these areas to quantify, control and manage. That is because inflation has 4 distinct aspects: realized inflation (price changes from past surveys), market based real yields of TIPS bonds and inflation swaps, and long term expectations of inflation, inflation measured with or without volatile sectors: food and energy. Because the statements may treat some or all of these aspects we believe accuracy in understanding FOMC inflation views is hard.
We have discussed the results of our two analyses previously. This section focuses on broader considerations of the of FOMC statement analysis. How useful were the results in practical terms? First, the training and classification of FOMC statement attributes is clearly feasible. Our efforts combining laborious human review of the statements for labelling purposes with traditional supervised learning methods can produce useful prediction for some attributes. Other attributes such as inflation are not easily predicted using our approach. This does not mean that the problem cannot be solved with this line of attack.
Second, sentiment analysis is likely to be the more impactful tool in forecasting for investment management purposes. Our preliminary results, effectively a conventional approach to sentiment analysis, yielded a realistic looking financial sentiment indicator. We demonstrate some moderate level of explanatory power of sentiment and equity market levels using a linear regression. Much more extension regression analyses covering first differences of sentiment and price series is required. We believe the results are insufficient to be useful in practice but sufficient to justify further refinement and investigation to unlock value.
The project analyzed the FOMC statements using text based methods. The results are encouraging but not definitive in their utility. A much longer programme of research would be needed to explore the implications of this work. Others have gone done this path. Some companies such as JPMorgan Chase, BlackRock have dedicated teams to analyze interest rate markets and central bank communications with machine learning tools. One company called Prattle (www.prattle.co) has even commercialized this idea and provides central bank sentiment analysis for 16 central banks including the Fed. Therefore, we are on the right path – one led by earlier pioneers.
Brusa, Francesca, Pavel Savor, and Mungo Wilson, “One Central Bank to Rule Them All”, working paper, 2015, Oxford University [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2658953]
Cieslak, Anna and Morse, Adair and Annette Vissing-Jorgensen, “Stock Returns Over the FOMC Cycle”, Journal of Finance, forthcoming, [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2687614]
Cannon, San, “Sentiment of the FOMC: Unscripted”, 2015, [https://www.kansascityfed.org/~/media/files/publicat/econrev/econrevarchive/2015/4q15cannon.pdf]
Fuka, Michel and Didier Sornette “The Sentiment of the Fed”, 2013, [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2191937]
Lucca, David and Francesco Trebbia, “Measuring Central Bank Communication: An Automated Approach with Application to FOMC Statements”, 2011, [http://faculty.arts.ubc.ca/ftrebbi/research/lt.pdf]