An AI Approach To Fed Watching written by Jeffrey N. Saret and Subhadeep Mitra of Two Sigma and reflects their views, not necessarily the firm’s – please see important legal disclaimer at bottom of post
Natural language processing techniques can translate Federal Open Market Committee (FOMC) meeting minutes into data. The results appear both intuitive and informative. For example, following the 2007-2009 financial crisis, the Fed increased the amount of time it devoted to discussing financial markets from 10 percent in 2007 to nearly 40 percent in late 2008. At more recent meetings, the Fed spent approximately equal time (~20 percent) discussing inflation, growth, financial markets, and policy. Topics like employment and trade have commanded only five percent of the Fed’s mindshare of late. Knowing what concerns the Fed might help allocators sharpen their focus on the longterm issues that matter for both monetary policy and the broader economy.
Mangrove Partners Narrowly Avoids “Extinction-Level Event”
An AI Approach To Fed Watching
In Shakespeare’s Hamlet, chief counselor to the king Lord Polonius asked the protagonist what he was reading. Hamlet accurately, if imprecisely, replied, “words, words, words.” Counselors to asset allocators might ask the same thing of so-called “Fed watchers,” who dutifully interpret the US Federal Reserve’s Federal Open Market Committee’s (FOMC) meeting minutes. The FOMC publishes these minutes three weeks after each of its eight scheduled annual meetings, likely to help asset allocators and others understand the slightly delayed but more detailed views held by the Fed’s monetary policy decision makers.
Historically, interpretations of those minutes required art, so Fed watchers pontificated and critiqued. Now natural language processing techniques can translate those minutes into relatively objective data. The results appear both intuitive and informative. For example, during the 2007-2009 financial crisis, the Fed increased the amount of time it devoted to discussing financial markets from 10 percent in 2007 to nearly 40 percent during late 2008. At more recent meetings, the Fed spent approximately equal time (~20 percent) discussing inflation, growth, financial markets, and policy. Topics like employment and trade have commanded less than five percent of the Fed’s mindshare during 2016. Knowing what matters concern the Fed might help allocators sharpen their focus on the long-term issues that matter for both monetary policy and the broader economy.
The FOMC’s Meeting Minutes Lend Themselves Well To Natural Language Processing
Analyses based on natural language processing (NLP) techniques usually require two main ingredients – a large number of texts worth analyzing and consistency across those texts. The FOMC’s meeting minutes offer both ingredients. Since 1993, the FOMC has published meeting minutes eight times per year using a relatively consistent structure.
Natural language processing (NLP) techniques can consistently and objectively identify and calculate timevariation in topics discussed in these FOMC minutes at a granular level. Blei, Ng, and Jordan (2003) describe one such technique – Latent Dirichlet Allocation (LDA). LDA algorithms try to classify the text into “unobserved topics” or groups, and then map each word of a text into those unobserved topics. Principal component analyses (PCA) of empirical data might offer a clarifying analogy, in that both LDA and PCA ignore any specific context or pre-conceived categories in order to statistically and unbiasedly identify relationships to reduce dimensionality or complexity. Similar to a PCA, a nearly infinite number of “unobserved topics” could perfectly describe a given text. For practical reasons, researchers typically select only a manageable few. The technical appendix offers more details.
Three Trends Jump Out Of The NLP Analysis Of The FOMC’s Meeting Minutes
The analysis highlights three trends in the FOMC meeting minutes (Figure 1). First, financial markets have captured and sustained a significant share of the FOMC’s discussions since the 2007-2009 financial crisis, seemingly at the expense of growth. Prior to 2007, the Fed typically devoted less than 10 percent of FOMC meetings to discussing financial markets. A notable exception occurred from 1997 to 1999 during the rolling Asian, Latin American, and Russian financial crises. Those events reaffirmed the stereotype that the Fed offered a “Greenspan put,” or an FOMC backstop to financial stress.
Beginning in mid-2007, stress in the financial markets again commanded the Fed’s attention, and the market vernacular shifted from the “Greenspan put” to the “Bernanke put.” At its peak in late 2008, the FOMC devoted nearly 37 percent of its meeting minutes to discussing financial markets, using words such as “securities,” “credit,” “dollar,” “rates,” and “mortgage.” Since then, the share of the FOMC’s attention has trended lower, though it has remained persistently above 20 percent during both the Bernanke and Yellen eras at the Fed. In contrast, the FOMC has reduced the relative fraction of its meetings dedicated to discussing growth. In 2002, for example, 50 percent of the FOMC minutes covered topics related to economic growth. That share has fallen by half in the minutes released so far during 2016 (January and March meetings).
inflation has become an increasingly important topic since 2014. During the 2007-2009 financial crisis, the FOMC devoted only 15 percent of its meeting minutes to discussing inflation. Since mid-2014, that fraction increased to more than 20 percent and, during the past few meetings, close to 30 percent. The other element of the Fed’s dual mandate, employment, has persistently commanded less attention (~5 percent).
Third, while inflation appears to have captured the greatest share of the FOMC’s mind for the past few months, the FOMC seems relatively unfocused. For the first time in the data set, four topics individually command more than 20 percent of the minutes: inflation, growth, policy, and financial markets. This might imply that these topics have become more interrelated today than in the past. It might also reflect a societal trend towards multitasking. Either way, the job of Fed watchers appears to have become more complex.
As one of the most informed and influential economic actors in the global economy, the FOMC’s views matter. At its eight regularly scheduled meetings per year, the FOMC reviews economic and financial conditions, determines its stance of monetary policy, and assesses the risks to its long-run goals of price stability and sustainable economic growth. Since the FOMC consists of a time-varying group of twelve, not always like-minded individuals who communicate in relatively qualitative and imprecise terms, minutes from the meeting represent subjective views.
Market observers trying to glean insights from these meeting minutes once needed to rely on the subjective interpretation of so-called expert “Fed watchers” (e.g., Romer and Romer, 1989) or their own interpretation. Now, asset allocators can apply natural language processing techniques to extract insights from the FOMC’s published meeting minutes, turning qualitative inputs into more easily analyzed, quantitative data. This paper’s analysis focused on topic identification and relative weights within FOMC discussions, but that represents only one potential application. Numerous other potential applications exist. For example, one could use this data to evaluate the tone or sentiment of a meeting. With the advancement of NLP, market observers might enjoy the option of limiting the amount of subjectivity they need to layer on to understanding the subjective views of the FOMC. Paraphrasing Polonius, market observers might discern more method in the Fed’s alleged madness.
Brief Technical Appendix
The mechanics of applying an LDA algorithm to FOMC minutes are straightforward. First, a custom text parser converts the raw text from the minutes into digestible input for the LDA algorithm. The parser drops the first section of the FOMC minutes describing previous open market operations, based on assumptions that those do not contain salient information. Next, the parser breaks each set of minutes into their constituent paragraphs. The parser removes “stop words” including pronouns, articles, prepositions, conjunctions, numbers, weeks, months, and other commonly occurring non-topic specific words, such as ”federal,“ ”reserve” or ”session.“
The number of unobserved topics is an important input into the LDA algorithm. According to Blei (2012), interpretability offers a legitimate reason for choosing the appropriate number of groups in the algorithm, as opposed to standard model selection techniques in machine learning, such as out-of-sample prediction accuracy. After iterating over LDA models with the number of topics varying from five to 12, an eight-topic model seemed to offer the most interpretable output for FOMC minutes, at least based on the words associated with each topic. Since these eight topics appeared to overlap somewhat (e.g., Topic 2 included words like “securities,” “credit,” and “market,” while Topic 4 included words like “financial,” “market,” and “equity”), Figure 1 aggregates these machine-selected topics into six more sensible groups: growth, inflation, financial markets, employment and trade.1 The figure reports the fraction of FOMC meeting minutes that discusses each of those six groups.
See full PDF and legal disclaimer below.
IMPORTANT DISCLAIMER AND DISCLOSURE INFORMATION
This report is prepared and circulated for informational and educational purposes only and is not an offer to sell or the solicitation of an offer to buy any
securities or other instruments. The information contained herein is not intended to provide, and should not be relied upon for investment, accounting,
legal or tax advice. This document does not purport to advise you personally concerning the nature, potential, value or suitability of any particular sector,
geographic region, security, portfolio of securities, transaction, investment strategy or other matter. No consideration has been given to the specific
investment needs or risk-tolerances of any recipient. The recipient is reminded that an investment in any security is subject to a number of risks including
the risk of a total loss of capital, and that discussion herein does not contain a list or description of relevant risk factors. As always, past performance is no
guarantee of future results. The recipient hereof should make an independent investigation of the information described herein, including consulting its
own tax, legal, accounting and other advisors about the matters discussed herein. This report does not constitute any form of invitation or inducement
by Two Sigma to engage in investment activity.
The views expressed herein are not necessarily the views of Two Sigma Investments, LP or any of its affiliates (collectively, “Two Sigma”) but are derived
from the Two Sigma Alpha Capture system (the “Alpha Capture System”), which gathers inputs from sell-side contributors (not analysts) to the Alpha
Capture System who receive compensation for their participation, as further described in the section titled “Brief Explanation of the Data” (page 1
hereof) and the document titled “Overview of the Two Sigma Alpha Capture System”. Such views (i) may be historic or forward-looking in nature, (ii)
reflect significant assumptions and subjective judgments of the contributors to the Alpha Capture System as well as, in some instances, the authors
of this report, and (iii) are subject to change without notice. Two Sigma may have market views or opinions that materially differ from those discussed,
and may have a significant financial interest in (or against) one or more of such positions or theses. In some circumstances, this report may employ data
derived from third-party sources. No representation is made as to the accuracy of such information and the use of such information in no way implies an
endorsement of the source of such information or its validity.
This report may include certain statements and projections regarding the anticipated future performance of various securities, sectors, geographic
regions or of the Alpha Capture System generally. These forward-looking statements are inherently subject to significant business, economic and
competitive uncertainties and contingencies, many of which are beyond our control. In addition, these forward-looking statements are subject to
assumptions with respect to future business strategies and decisions that are subject to change. Factors which could cause actual results to differ
materially from those anticipated include, but are not limited to: competitive and general business, economic, market and political conditions in the
United States and abroad from those expected; changes in the legal, regulatory and legislative environments in the markets in which Two Sigma
operates; and the ability of management to effectively implement certain strategies. Words like “believe,” “expect,” “anticipate,” “promise,” “plan,” and
other expressions or words of similar meanings, as well as future or conditional verbs such as “will,” “would,” “should,” “could,” or “may” are generally
intended to identify forward-looking statements.
Two Sigma makes no representations, express or implied, regarding the accuracy or completeness of this information, and the recipient accepts all
risks in relying on this report for any purpose whatsoever. This report is being furnished to the recipient on a confidential basis and is not intended for
public use or distribution. By accepting this report, the recipient agrees to keep confidential the existence of this report and the information contained
herein. The recipient should not disclose, reproduce, distribute or otherwise make available the existence of and/or all or any portion of the information
contained herein to any other person (other than its employees, officers and advisors on a need-to-know basis, whom the recipient will cause to keep
the information confidential) without Two Sigma’s prior written consent. This report shall remain the property of Two Sigma and Two Sigma reserves the
right to require the return of this report at any time.
© 2016 Two Sigma Investments, LP | ALL RIGHTS RESERVED | “Two Sigma” and “2?” are trademarks of Two Sigma Investments, LP.