At a Glance: Prattle Central Bank Analytics DataFeed
The Prattle Central Bank DataFeed uses a combination of natural language processing and machine learning to deliver unbiased, quantitative analysis of official communications from 15 central banks around the world.
The Prattle Central Bank DataFeed uses a combination of natural language processing and machine learning to deliver unbiased, quantitative analysis of official communications from 15 central banks around the world. Language is parsed and converted to quantitative data that helps investment professionals forecast market outcomes.
The product information and content statistics contained in this document are as of October 2018.
Prattle’s sentiment scores predict the market’s perception of central bank communications, leading to long-term insights related to central bank monetary policy decisions and more.
Prattle central bank analytics provide valuable inputs to identify buy, sell, or hold signals based on how assets respond to central bank language.
The Prattle Central Bank DataFeed provides analysis of policy statements, press releases, policymaker speeches, and more from 15 of the world’s most influential central banks. These official communications equate to roughly 1,500 events a year (Figure 1). For coverage by bank broken out by event type, see Figure 2.
Prattle’s scores are a reflection of how the market is likely to react to central bank language. For each bank, Prattle has algorithmically associated the words, phrases, sentences, and whole paragraphs used in central bank communications to their corresponding market impact, generating a lexicon of scored expressions unique to each institution. These lexicons allow Prattle to quantitatively assess the consequences of a given bank’s language on the market. The result of this analysis is an actionable numerical sentiment score for every speech, interview, press release, and set of meeting minutes by a central bank’s policymakers and committees.
Using a proprietary system involving natural language processing and machine learning, Prattle Central Bank Analytics fully accounts for the complexity of language in primary source material and evaluates all words within every central bank communication to produce the event’s sentiment score. For example, if the Fed issues the phrase “less hawkish” in its minutes or statements, Prattle’s algorithm would account for how markets will perceive the qualifying language and adjust the score according to previous market reactions to that language.
Figure 3 depicts the distribution of sentiment scores by bank for the past 10 years. Looking at the Federal Reserve, the notable negative outlier was the sentiment for the September 2, 2009 release of the Federal Open Market Committee meeting minutes. At this time, the U.S. unemployment rate hit 9.7%, the highest since 1983. In the minutes, Fed policymakers noted that they expect many companies will be “cautious in hiring,” further dampening the market’s sentiment on the unemployment rate.
Example Use Case
The Prattle Central Bank Analytics DataFeed identifies tradable opportunities, such as when treasury yields will rise or fall, using two scores that are provided for each communication; a raw score and residual score.
The raw score is a measure of how the Prattle algorithm measures the sentiment of a particular communication in relation to all other communications from the bank as a whole. The scores are normalized around zero, with positive numbers indicating hawkish sentiment and negative numbers indicating dovish sentiment. The scores can be thought of as a Z score: roughly 2/3 of scores fall between +1 and -1, and roughly 95 percent fall between +2 and -2. The residual score then provides historical context to the raw score by taking the score for a communication less that speaker's average score calculated on a rolling 12-month basis.
To test Prattle’s ability to predict Treasury yield movement, we will focus specifically on the Federal Reserve’s press releases after January 2000. Press releases can be seen as a reliable indicator of future policy. Prattle Central Bank Analytics makes it easy to identify strongly positive (hawkish) or strongly negative (dovish) events. For this study, only events with an average raw and residual score above or below an absolute value of .75 were utilized.
With those events identified, the next step is to overlay the average daily percent change for U.S. Treasuries (2,3,5,10, and 20 year) 25 days out from the press release. As seen in Figure 4, a prominent deviation occurs between the hawkish and dovish events. Those strongly hawkish events saw Treasury rates rise almost 10 basis points while dovish events led to an average drop of 2 basis points.
In recent research by the Federal Reserve Bank of San Francisco, How Important Is Information from FOMC Minutes?, the authors measure the difference in Prattle’s residual sentiment scores between official FOMC Statements and the published meeting minutes, to uncover effects on the Treasury yield curve when statements and minutes differ in tone. The FOMC meeting minutes are delivered three weeks after the corresponding FOMC Statements.
Adopting a similar approach, the difference between the Prattle residual scores of the FOMC Statements and their corresponding minutes was calculated for events between June 2012 to July 2018.
Identifying the spread between the scores of the two events helps capture human behaviour and market reaction to quantifiable differences in language between an official meeting statement and the written record of what occurred during the meeting.
The resulting values are then bucketed into quartiles based on the magnitude of difference. Quartile 1 contains those cases where the meeting minutes’ sentiment was the most positive in relation to the FOMC Statement and quartile 4 contains cases where the difference was the most negative. Said differently, Quartile 1 represents cases where the full minutes contained language that had a higher correlation to a positive market reaction than the initial press release. Quartile 4 represents cases where the full minutes’ language correlated to a more negative market reaction than the initial press release. Figure 5 charts the average return of the S&P 500 Index for each quartile, 15 days following the release of the minutes.
Cases where the language of the minutes was more correlated to positive market reactions (quartiles 1 and 2) show on average a .5% return on ‘1-Day’, the day of the minutes’ release. Quartiles 3 and 4, cases where the language was more negative, illustrate a slight market dip. On a longer time horizon, quartiles 1 and 2 continue to climb, seeing a roughly 1.3% return on average, while those events in quartile 4 saw a more neutral market reaction.
By providing sentiment scores on official communications from 15 influential central banks, Prattle’s Central Bank Analytics delivers comprehensive, unbiased, quantitative analytics that are tradable across multiple asset classes and time horizons.
For more information on Prattle’s data, including case studies, white papers, and additional research, please visit its provider page on the Open:FactSet Marketplace.