AI helps investors to uncover the truth hidden behind executive’s soothing words

Francis deSouza, chief executive officer of Illumina and its gene sequencing company, did his best to remain positive on his last earnings call.

The $8bn acquisition of Grail, a cancer screening company, sparked a Carl Icahn campaign, fought with competition authorities both on the Atlantic and in Europe, as well as criticism by Grail’s founder directors.

DeSouza said analysts that the drama only affected “a very small portion of the company”.

Speech Craft Analytics analysed audio recordings using artificial intelligence and found that each time he spoke about Grail his pitch, volume, and rate of speech changed. The use of filler words such as “um” or “ah”, and an audible gulp were also increased.

David Pope, chief data scientist at Speech Craft Analytics, said that the combination “reveals signs of anxiety and stress when dealing with this sensitive issue”.

DeSouza has resigned after less than two months.

Some of the largest investors in the world have been interested by the idea that audio recordings can provide insight into executives’ true emotions.

Several funds use algorithms that comb through transcripts from earnings calls and presentations of companies to find signals in the words used by executives. This field is called “Natural Language Processing”, or NLP. They are now trying to discover additional messages within the words spoken.

Mike Chen, Robeco’s head of alternative Alpha Research, said that audio can capture more than text. “Even though you may have a sophisticated machine that captures semantics, it is still only capturing semantics.”

The AI is able to detect “microtremors”, which are invisible to the human ear.

Robeco has been adding audio signals gathered by AI to its strategies since the beginning of this year. Robeco manages more than $80 billion in algorithmically-driven funds, making them one of largest quants. Chen stated that it has increased returns and he expects more investors to follow.

Audio represents a whole new level of the cat-and-mouse game between fund managers, executives and other professionals.

Yin Luo is the head of quantitative research for Wolfe Research. She said, “We found transcripts to be of tremendous value.” “The overall sentiment has become more positive, which is a problem for us and others. . . “Because the management of the company knows that their messages are being analyzed.”

Since the advent of NLP and the introduction of a more positive, companies have adjusted their language in order to play the algorithms.

Luo co-authored a paper earlier this year that found that using traditional NLP and audio analysis to distinguish between companies was an effective method of differentiation as their filings became increasingly “standardised”.

Although the costs have decreased, this approach can be expensive. Robeco invested three years in new technology infrastructures before even beginning to work on audio analysis.

Chen tried to use audio for years before joining Robeco but the technology wasn’t advanced enough. While the information available has improved, there still are limitations.

The most reliable way to avoid jumping into conclusions based on personalities – some executives are just more effusive by nature than others – is to compare different speeches made by the same person over time. This can make it difficult to assess the performance of new leaders, at a time that insight is most useful.

One executive from a company providing NLP analysis said, “A CEO change can mess up the overall sentiment [analysis].” This disruption effect must be more pronounced with voice.

The developers must also be careful not to introduce their own biases in algorithms that are based on audio. This is because differences like gender, class, or race may be more apparent than they would be in text.

Chen said that although “we are careful to ensure the conscious biases we’re aware don’t get in,” there may still be unconscious ones. Robeco has a diverse and large research team.

The results of algorithms can be misleading if they are used to analyze someone who is speaking in a language other than their native language. An interpretation that may work in one language might not work in another.

Pope predicted that investor relations teams will start training executives to monitor tone of voice and other behaviours that are missed by transcripts, just as companies have tried adapting to text analysis. Voice analysis is difficult for trained actors, who are able to stay in character. But executives may have a harder time replicating this.

He said, “Very few people are adept at modulating their voice.” It’s easier for us when we choose our words carefully. Since we were young, we’ve been doing this to avoid getting into trouble.