Sentiment Analysis in IR & Mexico’s 3Q21 Earnings Calls

The audience for earnings calls and reports is no longer only human. Natural language processing (NLP) programs are also listening and making important judgments about companies’ perceived performance, and even triggering real-time buying and selling. In response, many companies have changed the way they communicate in an attempt to get the desired results from NLP. In this week’s post we’ll examine this technology, look at current tools, and discuss results from GBM’s text analysis of Mexico’s 3Q21 Earnings Calls. 

About AI Text/Voice Analysis in IR

First let’s define some of the terminology that is used in this space. Artificial intelligence (AI) refers broadly to the discipline of intelligent machines, while machine learning (ML) refers more specifically to systems that can learn from experience, and natural language processing (NLP), a subset of AI that uses ML, is a set of systems that can understand language and improve. NLP uses a combination of linguistics, neuroscience, mathematics and computer science. Recent advances in ML, along with increased computational potential, have made NLP more reliable and scalable. 

An example of NLP in a more common context is Sonix, the company that converts audio to text, it first processes the audio with automated speech recognition (a type of machine learning) and then once it is in text format uses NLP to understand the meaning of the text (which can be quite complicated given abbreviations and colloquialisms). In finance, NLP is used for sentiment analysis, question-answering (chatbots), financial audits, document classification, and in general reducing the amount of manual labor. This blog will focus on sentiment analysis as it is most relevant to IR. 

NLP financial sentiment analysis is being used to examine all kinds of written or audio inputs that can give company insight. This includes basic disclosure information like earnings calls, pitch presentations, and quarterly reports, but also goes beyond that with things like social media and news NLP sentiment analysis being used for real time trading. NLP programs are generally looking for frequency of certain terms and positive, neutral, or negative sentiments. However, it can even be capable of detecting the tone of voice of the speaker, for example to get a reading on positivity and excitement in an earnings call. 

Traditionally corporate disclosures have tried to convey the desired sentiment and tones for humans listening or reading. However, a recent report by Stevens found that many large companies have changed the way they write reports and scripts; executives are even getting professional help from voice coaches to hit the right tone of voice for NLP. Companies are also using algorithmic programs to test drafts of their disclosures to gauge their expected sentiment score. 

Programs/Services that examine financial language

In the sentiment analysis space, and in general, the top 4 financial tools are Bloomberg, S&P Capital IQ, FactSet, Refinitiv Eikon. We’ve also included some information on Dataminr, FinBERT, and Wordstat.

Bloomberg 

Bloomberg (BBG) is able to identify a news story or tweet as being about a specific company and then assign it a sentiment score. BBG’s news sentiment analysis tool {TREN <GO>} allows investors or analysts to use news fragments to build statistical forecasting models that dynamically adjust price targets.

Besides being a leader in the financial language processing space, BBG is considered the top financial data tool with the highest market share in the space (33%). It is best used for fixed income (most comprehensive data sets), buy side, sales, trading, and asset management. A Bloomberg terminal costs approximately $25,000 USD per year.

Refinitiv

Refinitiv (20% market share, approx $22k per year for full version), is the most direct competitor of BBG. It’s focus is on sell side research. But in general is considered to be a less robust version of BBG, according to Wall Street Prep.

Refinitiv News Analytics, Refinitiv’s natural language processing engine, uses news sentiment machine learning. See this paper on its use in intraday forecasting of a major stock index and on daily trading of the 100 most liquid S&P 500 stocks. 

S&P Global Market Intelligence (Capital IQ + SNL)

S&P (6.2% market share, approx $13,000 per user per year, minimum of 3 users)  has a fleet of data scrubbers, searching footnotes and disclosures of earning reports for the numbers investment bankers need. It’s excel plug-in is considered high quality, but not as robust as FactSet’s. A nice touch is they send you $50 if you catch a mistake in their database, according to Wall Street Prep.

FactSet

FactSet, another leading financial data tool (4.5% of market share, and approximate cost of $12,000 per year), provides sentiment analysis for all of the earnings calls that they cover, giving scores by speaker, per section, overall, and word frequency analysis. They also have a document search tool which allows you to search through multiple companies. For example, in this study, FactSet looked at mentions of inflation in all of the S&P 500 earnings calls and found 2Q21 had the highest number of mentions since 2010, and the largest YoY increase. 

Aside from language analysis, the FactSet tends to be more popular with investment bankers, because of specific capabilities such as ease of seeing data in source documents, excel plugins, pitchbook macros with PowerPoint, and transaction screening tools. However, it’s not considered to have the most robust equity research and data scrubbing capabilities are good but not quite at the level of S&P, according to Wall Street Prep.

DataMinr

DataMinr is a real time AI platform that uses public data to detect signals of high-impact events. For example, DataMinr provides real-time Twitter data that can be analyzed for trade support, market awareness, client advisory, and thesis generation. 

FinBERT

FinBERT is based on a language representation model that was originally developed for biomedical text mining, and has been developed for the financial services sector. FinBERT operates on a dataset that contains financial news from Reuters and assigns sentiment using a phrase bank. Other than the cost to use Reuters it is free, but requires programming language knowledge to use. See more information at this Github page. 

Wordstat

Provalis Research has a text mining tool called Wordstat. It’s not necessarily for financial data but can be used for that. It has python integrations to modify even more specifically, including dictionary building, mapping etc. 

GBM Mexico Results

GBM used data science tools (unspecified) to analyze text as data in their recent report Top of Mind: Text Analysis from 3Q21 Earning Calls in Mexico and the World. They looked at Mexican quarterly earnings calls (400+) as well as global companies (22,700+) to identify common topics by word or phrase frequency.  

In Mexico, there was a positive trend in corporates’ top-of-mind vocabulary. “Supply chain” mentions, like “raw materials”, had the biggest increases and were dominant in the earning calls, which was also observed with global companies. Growth was also widely mentioned, which is a positive sign across Mexican companies. 

In terms of sentiment in Mexico, the proportion of negative pairs of words have continued to decline since 1Q20, while the list of positive words has remained stable. In 3Q21, 2.7% of monitored words were positive, and 1.6% were negative, compared to 1Q20 (1.7% positive, 7.0% negative). Arca Continental (AC) had the most positive mentions in its 3Q21 earnings call. 

Other interesting findings from the report were: 

• ESG and digital transformation topics continue to be on the rise in the Mexican corporate narrative, a trend that has been consistent over the past three years. 
• Compared to global reports, Mexico had less mentions of “labor” as it is less of an issue than in the US. 
• Pandemic – Mentions of the pandemic had a QoQ decrease of 2p.p. 
• Growth – AC, KOF, WALMEX, CEMEX, and GCC had the most growth mentions.
• Supply Chain – The following companies had the highest number of mentions CUERVO (37), NEMAK (21), ORBIA (20), and KIMBER (18). Supply related words in 3Q21 were up to 7.9% in 3Q21, compared to 4.5% in 2Q21, and 2.3% in 3Q20.
• Pricing – Pricing/inflation was relevant this quarter, up by 2p.p. On a 3 year trend, but stable QoQ. CEMEX had the highest exposure, followed by KIMBER, AC, and ALFA.
• Cost – CEMEX had greatest exposure, mostly for “input costs”, followed by KIMBER, ORBIA (“cost increases”), and FIHO (“cost control” and “cost-cutting”).
• ESG- CEMEX’s mentions of “alternative fuels” gave it the most exposure, followed by KOF (sustainability), and GCC.
• Digital – GFNORTE had the most mentions of digital growth, followed by TLEVISA and GENTERA.

Miranda IR is happy to help your company with sentiment analysis prep for earning calls, we use several of the tools mentioned in this blog.

Sources

• https://gbmenlinea.gbm.com.mx/Documentosanalisis/dss_top_of_mind_3q21.pdf
• https://www.stevens.edu/sites/stevens_edu/files/Machine%20Reading%2020210311.pdf 
• https://insight.factset.com/highest-number-of-sp-500-companies-citing-inflation-on-q2-earnings-calls-in-over-10-years 
• https://provalisresearch.com/products/content-analysis-software/ 
• https://www.wallstreetprep.com/knowledge/bloomberg-vs-capital-iq-vs-factset-vs-thomson-reuters-eikon/
• https://gbmenlinea.gbm.com.mx/Documentosanalisis/dss_top_of_mind_3q21.pdf 
• https://sonix.ai/articles/difference-between-artificial-intelligence-machine-learning-and-natural-language-processing 
• https://www.avenga.com/magazine/nlp-finance-applications/ 
• https://www.bloomberg.com/professional/sentiment-analysis-white-papers/ 
• https://www.bloomberg.com/professional/blog/trading-news-use-machine-readable-data-find-alpha/