NLP for Detecting Deception

American Century scans earning calls for fibs

  • Company earning-call transcripts are a rich source of alternative data.
  • American Century Investments wanted to detect deception in management language.
  • Sentiment Analysis scores texts on emotional markers.
  • The asset manager used sentiment analysis to spot deception signals.

Gaining an edge with earning-calls transcripts

Before being made publicly accessible, relatively few had access to analyst-call discussions, where company management discusses its recent accounts and responds to analysts’ probing questions.

With this information no longer as exclusive, investment managers want new sources of insights, to maintain an edge in identifying mispriced assets. One approach is to use alternative data: anything from satellite images and social media postings, to credit card data and ecommerce transactions.

Earning calls transcripts are one such source of added insight. Investors can automatically scour comments made by a company’s management team. Potentially, it’s a way to capture signals that aren’t reflected in the company’s latest financial reports.

NLP Concept: Sentiment Analysis

Texts discussing the markets often carry a sentiment, by casting events that could impact a position in either in a negative or positive light.

Human readers can easily judge the sentiment polarity of a text (whether it’s positive or negative). They can also detect more subtle signals, like enthusiasm, humor and bias. But – relative to machines – are are slow readers.

Sentiment Analysis is the task of automating the decision of whether a text is positive or negative, while processing large amounts of content. Because the language cues for polarity are often subtle, this machine learning task often involves showing the model many examples of positive and negative texts, to help it tell them apart.

American Century developed a sentiment model

Asset management firm American Century complements its investment processes with a sentiment model that looks to detect deception in management language during quarterly earnings calls.

It checks for four elements of deception:

    • Omission (failure to disclose key details),
    • Spin (exaggeration from management and overly scripted language),
    • Obfuscation (overly complicated storytelling), and
    • Blame (deflection of responsibility).

Such signals are more subtle than merely polarity (see NLP Concept above). In an effort to avoid biases, researchers accounted for both the unique style of a given management team and the collective language of its industry peers.

To develop the model, the researchers first turned to financial journals and psychology texts to identify linguistic patterns associated with deception. This was later enhanced with feedback from the investment team.

How finance uses NLP