Words, Not Numbers, Drive Cutting-Edge Economic Intelligence | Lord Abbett
Image alt tag

Error!

There was a problem contacting the server. Please try after sometime.

Sorry, we are unable to process your request.

Error!

We're sorry, but the Insights and Intelligence Tool is temporarily unavailable

If this problem persists, or if you need immediate assistance, please contact Customer Service at 1-888-522-2388.

Error!

We're sorry, but the Literature Center checkout function is temporarily unavailable.

If this problem persists, or if you need immediate assistance, please contact Customer Service at 1-888-522-2388.

Tracked Funds

You have 0 funds on your mutual fund watch list.

Begin by selecting funds to create a personalized watch list.

(as of 12/05/2015)

Pending Orders

You have 0 items in your cart.

Subscribe and order forms, fact sheets, presentations, and other documents that can help advisers grow their business.

Reset Your Password

Financial Professionals*

Your password must be a minimum of characters.

Confirmation Message

Your LordAbbett.com password was successully updated. This page will be refreshed after 3 seconds.

OK

 

Economic Insights

In addition to traditional data reports, investors can now access advanced economic indicators that rely on text for more real-time analysis. 

Read time: 5 minutes

The written word has been with us for some time, but with cheaper data storage, voice transcription, and internet communication, the world is now awash with text and machines can easily read the bulk of it. In this second edition of our series on alternative data (read the first part), we explore how machine analysis of text is changing the world of economics.

According to Nobel Laureate Robert Shiller, the origination and spread of popular narratives have notable impacts on financial and business cycles. Traditionally one would have to identify these narratives by scouring news sources and amalgamating the themes using one’s own intuition. But on a practical basis this is only feasible up to a certain point. There are limits to the amount one individual can read in a day. Lastly, we are still subject to behavioral biases – we may read only the good news and not the bad and we may only read from certain sources. In contrast, machines are presumably immune to these issues and, with careful programming, they may analyze these trends in economic narratives more effectively than humans.

How do the machines convert text into data? This is a sprawling topic that is hard to do justice in a short article, but we can summarize it into four buckets.

  1. Count Words: This is as simple as it sounds: within a block of text, count up the number of times a specific word appears. We hear this often when a politician gives a speech – how many times did she say “taxes”? One prominent example is the oft-cited Economic Policy Uncertainty Index, which is partly constructed from a count of specific words from news articles. Although more mentions of a word suggest many people are talking about it, we do not know the context or sentiment surrounding the word.
  2. Score the Words: After counting the words, it’s time to assess their potential implications. Researchers do that by assigning a score to each word in a block of text. For example, we can score words based on their negative or positive meaning in common usage. Two common lexicons used to translate words into scores are the Harvard General Inquirer and the Liu Sentiment Lexicon, whose research teams list words and their associated scores as developed for specific purposes. One drawback here is that a lexicon developed for restaurant reviews may not apply words in the way we need for investing. (Diners and stock chartists may not think the same way about candlesticks, for example.) Additionally, we still do not know the context surrounding the word.
  3. Determine Meaning from Sets of Words: Researchers have also designed algorithms to determine the meaning of sets of words. This is a very broad area of linguistics called natural language processing (NLP) that attempts to build models to output the meaning of language. The main idea in this approach is that once armed with a model for language we can better ascribe meaning and sentiment to text. Again, this is something humans can do but when applied at scale by machines may save time and avoid bias.
  4. Group Words into Clusters: We could get a sense of the prevalence of news topics from the NLP language model. But we can also determine topics trending in news by grouping words into clusters. For each document or block of text we attempt to ascribe groups of words to topics and measure their frequency or intensity. For example, we could say a certain set of words is associated with the topic of education (“school”, “teacher”, “student”) while a different set of words is associated with the legal system (“court”, “attorney”, “judge”). Alternatively we can try to allow an algorithm to determine the topics without our input by analyzing how words group together in text.

 

 

How do these four approaches apply to measuring the economy? They are generally used in three ways: to follow central bank policy, to determine sentiment about the economy, and to follow trending macroeconomic themes.

Analyzing Central Banks with Text Data

In decades past central banks did not talk a lot about what they were doing outside of announcing their interest rate decisions. This was the era of the central bank watcher, the former insider who could ascribe the true underlying meaning to the few words released by central bankers. However, over time central banks began publishing more statements, minutes and speeches. And more recently, as developed central banks globally have moved rates closer to zero, they are mostly using “unconventional” policy like quantitative easing and forward rate guidance that require considerable verbiage to explain fully to financial markets.

Many companies are trying to apply analytical techniques to this growing trove of documents.  One example we use is to take the Federal Reserve’s Beige Book report on economic conditions and apply sentiment scoring to the text along with a little customization to adapt the standard lexicon to central banking. We find that it results in a fairly accurate description of the sentiment in the Beige Book over time.

 

Figure 1. The Beige Gauge: Parsing This Fed Report May Give an Accurate Read on Economic Sentiment

Source:  lenkiefer.com. The Beige Book is a summary and analysis of economic activity and conditions prepared with the aid of reports from the district U.S. Federal Reserve Banks for policymakers before a U.S. Federal Reserve policy meeting. The chart shows a sentiment score for each Beige Book report from March 2008 to May 2020. The score was computed by assigning each word in each report to positive, negative, or neutral sentiment based on a 100,000-word sentiment dictionary. The sentiment index then counts number of positive minus the number of negative words and divides by the number of negatives. A score of 0 would have exactly half positive and half negative words.

 

This presents a thorny problem for central bankers. If everyone is using machines to read their pronouncements and using different methodologies, how can policymakers understand how market participants are internalizing their communications? And do central banks end up avoiding certain phrases to throw off the machines?

Using Text Data to Measure Sentiment

Sentiment in investing is hard to measure, but we are making some inroads using text as data. Researchers from the Federal Reserve Bank of San Francisco have created a large database of news articles, called a corpus, and selected articles where their news provider has tagged the article as having an economic topic. The authors then score the words in the articles using their own customized lexicon. They regard the index as highly correlated with consumer sentiment (as represented by the University of Michigan Consumer Sentiment Index), as we see in Figure 2. In contrast to the highly automated San Francisco Fed index, the Michigan index is calculated each month based on a household survey conducted by phone, a polling method that has not changed much over the decades. Researchers at the Reserve Bank of Australia also developed a sentiment index along similar lines, but extended the analysis to include a housing sentiment index, a useful gauge of real estate sector.

 

Figure 2. Old School and New Style: Two Approaches to Reading the Mood of the U.S. Consumer

U.S. Federal Reserve Bank of San Francisco Daily Economic News Sentiment Index (left axis) and University of Michigan Consumer Sentiment Index (right axis), January 2, 2019–July 26, 2020

Source: U.S. Federal Reserve Bank of San Francisco (San Francisco Fed) and Bloomberg. The San Francisco Fed’s Daily News Sentiment Index is a high frequency measure of economic sentiment based on lexical analysis of economics-related news articles. The University of Michigan Consumer Sentiment Index is a consumer confidence index published monthly by the University of Michigan. Consumer opinions on economic conditions are gathered in a phone survey. Opinions on current conditions make up 40% of the index, with expectations of future conditions comprising the remaining 60%.

 

Following Themes

An emerging area of research in economics is extracting themes from the news. These methods measure the state of the economy by parsing financial media reports. The topics extracted from the news itself are interesting for investment purposes as they may reflect the most important concerns of the day. The trending topics also tell us the nature of a particular shock hitting the economy. It is often the case that the economy is jolted by factors that differ from decade to decade and textual analysis allows us to understand these shocks better. Two prominent examples of this research are work by Kelly et al and Larsen and Thorsud.

A Final Word on Words

Using text as data is an emerging and important part of economic analysis in financial markets. We believe this trend may offer a more unbiased and more nuanced analysis of the economy on a large scale. We may also be able to measure things that traditional macroeconomic data cannot capture adequately, such as sentiment and the nature of economic shocks. In future editions of our alternative data briefs we will describe how this textual analysis trend can inform security selection and portfolio management.

 

Forecasts and projections are based on current market conditions and are subject to change without notice. Projections should not be considered a guarantee.

This article may contain assumptions that are “forward-looking statements,” which are based on certain assumptions of future events. Actual events are difficult to predict and may differ from those assumed. There can be no assurance that forward-looking statements will materialize or that actual returns or results will not be materially different from those described here.

The information provided herein is not directed at any investor or category of investors and is provided solely as general information about our products and services and to otherwise provide general investment education. No information contained herein should be regarded as a suggestion to engage in or refrain from any investment-related course of action as Lord, Abbett & Co LLC (and its affiliates, “Lord Abbett”) is not undertaking to provide impartial investment advice, act as an impartial adviser, or give advice in a fiduciary capacity with respect to the materials presented herein. If you are an individual retirement investor, contact your financial advisor or other non-Lord Abbett fiduciary about whether any given investment idea, strategy, product, or service described herein may be appropriate for your circumstances.

The opinions in the preceding commentary are as of the date of publication and are subject to change. Additionally, the opinions may not represent the opinions of the firm as a whole. The document is not intended for use as forecast, research or investment advice concerning any particular investment or the markets in general, and it is not intended to be legal advice or tax advice. This document is prepared based on information Lord Abbett deems reliable; however, Lord Abbett does not warrant the accuracy and completeness of the information.

ABOUT THE AUTHOR

image

Please confirm your literature shipping address

Please review the address information below and make any necessary changes.

All literature orders will be shipped to the address that you enter below. This information can be edited at any time.

Current Literature Shipping Address

* Required field