4 min read
Identifying inflection points in tone on earnings transcripts
June 11, 2019
4 min read
We are excited to launch our breakthrough sentiment model that understands the breadth and nuances of financial language. Leveraging Deep Learning AI algorithms and a massive, 10+ year trove of proprietary human-labeled training data, the result is the market’s first and only intelligent sentiment analysis capability for financial language.
Buy-side and sell-side analysts have always recognized the importance of the sentiment conveyed by company management on earnings calls, and its connection to stock price movements. But there’s been no way to systematically or accurately monitor and catch these sentiment shifts with technology.
Until now. With the launch of AlphaSense Sentiment, you can apply a sentiment lens to transcripts within moments of them appearing on the platform. You can easily screen documents to see:
- Companies with the largest Sentiment score shift from the previous quarter
- The most negative or positive documents, on a relative score from -100 to 100
- Highlighting of phrase-level positive or negative sentiment underlying score shifts, honing in on the specific tonal text that our model has identified with high confidence.
How our approach to sentiment leaps past all other prior “sentiment” tools or metrics based on financial language
AlphaSense Sentiment takes a radically improved approach to applying the sentiment to financial documents than any product on the market. Here’s how:
Our model leverages Deep Learning AI and understands context. Legacy sentiment scores in the market are either dictionary-based, aka, the “bag-of-words” approach (simply counting hits against positive and negative word lists) or rules-based (trying to improve upon simple dictionaries with hand-coded rules). Both of these legacy approaches are naive in failing to understand the context and hand-coded scenarios and fail to scale beyond a narrow set of scenarios present in business and financial language.
Our AI model is trained on 10 years of human-curated financial text. Like all of the most successful AI systems, our Sentiment model is trained on a massive data set. Our AI model surpasses all past sentiment models, leveraging our proprietary access to the vast human-labeled training data we have meticulously built over the last 10 years. Besides a huge investment of time and resources, achieving this performance requires a very high level of proficiency on the latest AI models.
We made the long-term investment because we understand that high accuracy is critical to informing investment decisions – and in fact, we have held the quality bar so high that our own models didn’t exceed it until 10 years of work by over 30 people just on this one capability.
Our model is not easily confused. Simple dictionary-based approaches fail with many common scenarios like negations, co-references, and others. Our AI model has learned these from the vast training data, and succeeds where past models have failed. Our confidence in the model is demonstrated in the fact that we show phrase-level highlighting of positive/negative sentiment. AlphaSense Sentiment shows you exactly how we’re arriving at our insights and that level of transparency stands alone among sentiment models.
Getting started with AlphaSense Sentiment in just three clicks:
Here’s how it works: When viewing a transcript within the AlphaSense platform, click the “Show Sentiment” button.
The Sentiment view reveals phrase-level sentiment while ignoring boilerplate language. The middle “hits” pane shows both the overall score of sentiment as well as score change.
Analysts are already using AlphaSense Sentiment to transform their earnings research
A couple of early beta users leveraging AlphaSense Sentiment for deeper earnings analysis have shared some of their learnings here:
- Sell-side analyst Jan Svenda digs into Sentiment
- Oil & Gas analyst Joseph Triepke takes a deep drive into O&G trends using Sentiment
If you’re an AlphaSense client, log in now to get started with Sentiment.
More like this
3 min read
3 min read