Featured Image

The CEOs All Warned Us

Data Science and Technology

By Ruggero Scorcioni  |  June 15, 2020

We are inundated daily with exciting and dramatic financial and economic news. Sometimes we forget that a healthy economy is not founded on the latest tech startup, the flashy IPOs, or surprising M&A. We know and trust that the majority of companies keep the economy healthy. They are the so-called “core” companies, not newsworthy by definition. They keep their businesses alive, their suppliers busy, and their clients satisfied. Not very exciting, no high margins, but they do the heavy lifting and support us all. It is unfortunate, but if one of them is in trouble, it does not move the needle. We don’t really notice. What if 5% are in trouble, would that make a difference? Would anybody notice? Probably not.

But what if the number of core companies (showing neutral performance) dropped 10% in a few days? Nobody noticed but this is exactly what happened on February 6, long before the stock market crashed due to the COVID-19 pandemic. February 6 gave us the first indication that the U.S. core economy was entering uncharted waters.

Segmenting Companies with Neutral, Positive, and Negative Outlooks

To identify the core or neutral companies versus those companies that were likely to over- or under-perform the index to which they were attached (NYSE or Nasdaq composite index), we developed a machine learning (ML) model trained on Factset CallStreet earnings transcripts. The model, called High Impact Transcripts (HIT), is based on one-month returns following the company’s earnings call, which we assume reflects the outlook presented in the earnings call. We define negative HIT (N-HIT) as the percentage of underperforming companies versus the long-term trend; P-HIT is the percentage of overperforming companies. The long-term trend is the daily average value of the percentage of companies with a negative/positive outlook computed over the 2013-2019 period. An N-HIT/P-HIT value of 20% means 20% more companies than expected are identified with a negative/positive outlook.

Our analysis is not based on a sentiment reading of earnings call transcripts. Rather, our model is based on more than 15 years of past earnings call transcripts in the NASDAQ and NYSE exchanges and the subsequent market returns of those companies’ stocks. The historical data spans more than 200,000 earnings calls. Fortunately, market events like COVID-19 or the 2008 global financial crisis are rare. This allows machine learning models to better identify “normal” market behavior than rare events. For this reason, the HIT model labels transcripts assuming a normal market.

Our backtesting analysis has shown that in recent years, for 2010-2017 specifically, this model has held up quite well. The chart below illustrates how well our positive and negative classifications perform in normal markets. Using HIT negative and positive classifications as separate trading strategies, at the end of 2017, the positive strategy sees returns 37% higher than the relative index; this corresponds to a 4% yearly return. The negative strategy gives returns 20% below the relative index, corresponding to a -2.7% yearly return.

Chart 1

February Earnings Calls Raised a Red Flag

During January 2020, our model performed as expected. However, during February the model drastically deviated from the long-term trend. In just a few days, the number of companies underperforming the index drastically increased over the expected value. It appeared as if CEOs knew something that no one was willing to acknowledge: the impact of COVID-19 would be deeper than the market had accounted.

The chart below shows in red the trajectory for the N-HIT during 2020 and its percent deviation from the mean. The gray dashed line represents the daily volume of earnings calls relative to the annual volume. At the beginning of January, there was a positive spike for N-HIT. Given the extremely low volume of calls at that time, it is not significant. Earnings calls volume started to increase at the beginning of February as we entered earnings season; as the daily volume of calls increased, the percentage of companies with a negative outlook increased. February 6 is the earliest date at which N-HIT deviated from the norm. N-HIT kept increasing until the beginning of April. The black arrows indicate the February 19 S&P 500 peak and the March 23 bottom.

N-HIT Signal NEW

The Negative Outlook Remains Elevated

The chart below adds the yearly N-HIT signals for the period 2013-2019, displaying one line per year. Prior to February 6 of this year, the 2020 N-HIT behaved within the range of the past seven years, represented by overlapping regions of both sets of lines. In contrast, starting February 6, the N-HIT signal has remained higher than almost any day of the past seven years. This remains the case as we enter June even as the N-HIT signal peaked at the beginning of April and slowly decreased during May.

N-HIT Signal 2 NEW


Despite the continued highly negative sentiment in earnings calls, the S&P 500 has recovered by more than 35% from the bottom in just a couple of months, although the index remains down 5.9% for the year. Will the market revert to levels more in line with the negative sentiment or will the CEOs adapt their business forecasts to the “new” normal? Whatever the future holds, the N-HIT could provide valuable insights to understand when an economic downturn may be lurking around the corner.

New call-to-action

Ruggero Scorcioni, PhD

VP, Principal Machine Learning Engineer, Cognitive Computing

Dr. Ruggero Scorcioni is Vice President, Principal Machine Learning Engineer, Cognitive Computing at FactSet. He joined in 2018 and prior to that, teamed up with Nobel laureate, Gerald Edelman, as an associate fellow at The Neurosciences Institute in San Diego, California, where he created a model of the mammalian visual cortex implemented on GPUs. He also founded a startup called PolarSleep and joined HRL Labs as Principal Investigator of neuromorphic computing. Dr. Scorcioni earned a Ph.D. in Computational Neuroscience from George Mason University.