If investors are detectives seeking clues for outperformance in the US large-cap equity market, natural language processing is a team of tireless assistants.
Scouring company financial statements, listening to management calls, visiting facilities, analyzing competitors—an analyst’s daily life is filled from morning to night. And a river of incoming information, much of it unstructured, flows faster than ever. That means a lot of reading—and possibly less time for synthesizing data.
As fundamental analysts dig deep into companies and their business operations, conversations with management teams are a key source of information. Earnings calls are particularly important—a direct communication channel to company leaders, offering insights into a firm’s financial health, strategic direction and market trends.
Earnings-call transcripts may be fertile ground for insights—from forward-looking statements to key performance metrics, and from operational updates to market sentiment. Subtle clues matter too: tone, language and level of detail may help analysts gauge management's confidence and potential future performance. And our research suggests that incorporating these factors may lead to more informed investment decisions.
Sentiment: The “Mood” of a Message Matters
The sentiment metric is an intuitive but powerful indicator that captures a document’s tone: is it relatively positive, negative or neutral?
Sentiment assessed from earnings-call transcripts may impact future stock prices. Our analysis suggests that the stocks of companies whose communications score higher on sentiment have tended to outperform following the calls. But given the sheer volume of earnings calls, manually evaluating sentiment can be daunting.
This is where, in our view, natural language processing (NLP) sentiment analysis may be a useful tool for analysts.
“Bag of Words” and “Context-Aware”: NLP to the Rescue
NLP, a branch of artificial intelligence (AI), can be a powerful resource to help analysts extract insights from mountains of documentation. Instead of hiring a team of interns to comb for clues, NLP can do the job faster and more efficiently.
We think it makes sense to capture sentiment using two NLP approaches: “bag of words” and “context-aware.” Bag of words provides a baseline for assessing text by counting the number of positive and negative words in a document. This is highly intuitive, but also simplistic and subject to manipulation.
The context-aware approach, leveraging models such as BERT, GPT and LLaMA, analyzes the construction of sentences and their context within the overall document. These models measure aspects such as sentiment by analyzing overall language patterns, making their analysis more accurate and contextually relevant than simple counts of positive or negative words.
Applying both the bag-of-words and context-aware approaches to the earnings transcripts of US large-cap companies (Display) from 2010 to 2023 reveals a large sentiment dip in mid-2020. This bout of pessimism was related to the impact of the COVID-19 pandemic and its global economic aftershocks on company financials.
That downturn aside, both metrics indicate improving sentiment since 2010—more so using the bag-of-words approach. We haven’t yet analyzed this trend in detail, but other studies suggest that company executives are adapting to the age of AI by communicating with more positive words and phrases. And the sharper increase in the average bag-of-words sentiment suggests that firms are incorporating positive words more often.
A few examples of earnings-transcript text analysis (Display) suggest that the context-aware approach is more discerning. The bag-of-words approach interpreted the word “questions” within a phrase as negative, assigning a score of –0.20 on a scale of –1 to 1. However, the broader phrase is a fairly standard analyst comment during earnings calls—the context-aware approach captured this benign meaning, assigning a nearly neutral sentiment score.
“Nice quarter” is typically viewed as a positive statement by investors, but the absence of any sentiment-charged words leads to a 0 score from bag of words; context-aware captures the overall positive tone. In the third sentence, the positive “surpass” and other neutral words yielded slightly positive sentiment from bag of words, while context-aware captured the full positive tone. In the fourth sentence, context-aware assigned a much more confident negative score, befitting the overall tone.
“Context-Aware” Sentiment May Have More Potential
To see how sentiment might be effective more broadly, we can compare two equally weighted strategies that take long positions in stocks with top-20% sentiment scores (the most positive language) and short positions in bottom-20% sentiment stocks (the most negative language).
A hypothetical dollar invested at the start of 2010 in such a strategy based on the bag-of-words metric and rebalanced monthly would have grown to $1.16 (Display); the same strategy using context-aware sentiment would have produced $1.43. The more effective showing of context-aware sentiment, in our view, is because the bag-of-words approach is more easily manipulated by management teams, which can inoculate earnings calls with positive words.
While context-aware was more effective, the signal has weakened in the last few years—we think because management teams are incorporating more positive phrasing into earnings calls. This is where continued innovation and strong fundamental research come in. Analysts’ insights can help refine and fine-tune NLP tools to keep them effective, while human insight and knowledge interpret the full breadth of inputs that tell a company’s story.
The Big Picture: An Efficient, Tireless Research Assistant
Deploying NLP models on text-based data—from transcripts to news or even to patents—may foster insights that better equip analysts to assess performance potential in stocks.
The speed and efficiency of these models—governed by strong infrastructure—can save analysts from scouring websites for new information and digesting huge documents to understand high-level financial performance. They may also enhance risk monitoring, signaling changes in sentiment that might call for further investigation—and possibly portfolio action.
We think research analysts would do well to harness the power of NLP’s ability to rapidly process text-based information more efficiently. But data scientists must continue their efforts to sharpen and adapt these tools as markets—and company behaviors—change, because innovation can’t stand still.
The views expressed herein do not constitute research, investment advice or trade recommendations and do not necessarily represent the views of all AB portfolio-management teams. Views are subject to revision over time.
A message from Advisor Perspectives and VettaFi: To learn more about this and other topics, check out our full schedule of upcoming CE-approved virtual events.
© AllianceBernstein
Read more commentaries by AllianceBernstein
