There has recently been a confluence of advances which, in combination, have made the analysis of text an interesting proposition for systematic investors. These advances include the availability of greater computing resources, the evolution of mathematical methods, and the dramatic increase in both publication and collection of machine-readable text. Alongside the technological advances is a long-standing acknowledgement that valuable information resides within textual data. Non-systematic investors spend significant amounts of time reading news stories, conversing with other market participants and engaging with company management (which are activities of a different kind from analysing numerical data). While it is not clear that all these activities are value accretive, the breadth of possibility coming from being able to elicit information from text is a tantalising prospect. Textual analysis offers the ability to rigorously assess the informational content of a document in much the same way we do with numerical data. However, with vast swathes of data to examine, how do we find valuable information amidst the noise?

The technical challenges associated with textual analysis are exciting in themselves. However, just because we can examine millions of social media posts, that does not mean we should. Extracting valuable information from text is highly dependent on understanding the purpose of the text and aims of the analysis. Who is the author? What are their motivations? What information can we realistically learn? These challenging questions are answered by economics, not algorithms. The algorithms used need to be strongly informed by the economics. This is an important point which is easy to say but hard to achieve.

The project we undertook systematically utilises the information that company managements transfer to investors when engaging with analysts. It uses textual analysis to accomplish this. This not only reduces the information advantage fundamental investors have due to information obtained through this channel, but also creates our own information advantage via the breadth of companies we are able to analyse through systematising the process. Meeting management and attending meetings is clearly a time-consuming process which cannot feasibly be performed for all stocks in an index. Furthermore, systematising the process helps us to avoid behavioural biases to which fundamental investors may be prone when evaluating company management, such as anchoring current opinions to past outcomes.

There are multiple textual datasets available for the analysis of company management, and they have varying degrees of history and uniformity. Our focus on earnings call transcripts is primarily motivated by the information being collected directly from company management. Regulatory filings will have been gentrified by legal and compliance teams and are possibly not written, or even seen, by senior management. Conversely, transcripts from earnings calls, where management deliver the message in person, are harder for the company to gentrify and require some form of engagement from senior management. Earnings calls also contain two-way interaction between management and analysts, offering insight into how management respond to analyst challenges and limiting the extent to which they can either gentrify the message or avoid difficult but relevant topics. Finally, they are regular, scheduled calls discussing both a company’s recent performance and forward prospects, making the data broadly comparable across many companies.

When examining management communications, we explicitly focus on sentiment, believing it offers insight into management thinking, helping us forecast future company fundamentals and asset prices. It naturally follows that this model enhancement would lie within our sentiment factor. It is important to consider that company management are incentivised to portray their company in the best light. To incorporate this feature of our text, we include measures of Quality, designed to distinguish between realistic and embellishing sentiment. These design choices clearly demonstrate our “economics first” approach to textual analysis. This new source of information is exciting. Incorporating soft, non-factual information for the first time within our investment process is a significant enhancement to our existing processes. However, we are very aware of the inflated hype that can surround such approaches. It is important to highlight that this project is an augmentation of our existing information mosaic. It is not a panacea, but is rather an evolution of our existing processes, designed to enhance our view of the world and to help us make better investment decisions.