Behavioral finance has documented a wide range of stylized actions of investors, like loss aversion, regret aversion, overtrading, overconfidence, home bias and so on.

The Nobel Prize Richard Thaler have documented the tendency of the investor of overreacting to large economic events.

Financial advisors have long advised their clients to stay calm during any passing financial storm in their portfolios. Despite this, a percentage of investors tend to ‘freak out’ and sell off a large portion of their assets in certain adverse market environments. This situation is often discussed in the financial press and media, but is rarely defined or quantified.

Daniel Elkind and coworkers have brilliantly contributed to the behavioral finance literature with their research “When Do Investors Freak Out? Machine Learning Predictions of Panic Selling“, published in August 2021.

The authors have developed a method to identify panic selling and apply it to a novel large dataset of brokerage account information from 2003 to 2015 to examine panic selling and ‘freakout’ behavior.

They analyzed the investment activity of 653,455 individual brokerage accounts belonging to 298,556 households spanning the period 2003-2015.

They defined “freaking out” as occurring after a portfolio value decline of 90 percent in one month and the investor selling half of the portfolio within that month.

They defined re-entering the market as occurring when the portfolio recovered to 50% of its pre-liquidation value and the investor buys at least half of what was previously sold.

Following is a summary of their findings:

  1. There were 36,374 panic sells by 26,852 household investors (9% of all households) across a period of 13 years between January 2003 and December 2015.
  2. While panic sales are infrequent, with only 0.1% of the investors panic selling at any point in time, they occur at up to three times the baseline frequency when there are large negative market movements — a disproportionate number of households make panic sales when there are sharp market downturns, the phenomenon of freaking out
  3. Of households with at least one panic-selling event, 21,706 of them (81%) did so once within the sample period, while 3,081 (11%) did so twice.
  4. 31% of the investors who panic sold never returned to reinvest in risky assets. However, of those that did buy stocks again, 59% reentered the market within five months and another 13% returned within 10 months. 
  5. Investors who are male, or above the age of 45, or married, or have more dependents, or who self-identify as having excellent investment experience or knowledge tended to freak out with greater frequency. Whilst, the three occupational groups with the least risk of panic selling are ‘paralegal’, ‘minor’ and ‘social worker’.
  6. The occupational groups with the three highest risks of panic selling were self-employed, owners and real estate. 

In the chart below you can see the proportion of panic selling (blu bars) in relation to changes in S&P 500 (green line).

Elkind and coworkers also found that the median investor earned a zero to negative return after freaking out because while freaking out does protect investors during a crisis, such investors often wait too long to reinvest, causing them to miss out on significant profits when markets rebound. For example, an investor who liquidated at the start of the Great Financial Crisis and held out for more than 34 months after liquidation would have missed the post-2009 market rally and forgone potential profits. The bottom line is that freaking out is suboptimal behaviour. Of particular interest was that investors who identified themselves as having excellent investment experience freaked out more than twice as often as those who identified as having no experience — demonstrating that overconfidence is an all-too-human trait.

Elkind, Kaminski, Lo, Siah and Wong also used logistic regression and deep neural network techniques to develop machine learning models to predict when investors might panic sell in the near future. Their set of predictive features included the demographic characteristics of the investor, their portfolio histories, and current and past market conditions. Their best-performing deep neural network achieved a 70% true positive accuracy rate and an 81% true negative accuracy rate, “demonstrating that artificial intelligence techniques can assist in identifying individuals at risk of panic selling in the near future.” Among the most important predictive variables were:

  1. Being young or elderly decreased the risk of panic selling, as did being disabled or a minor
  2. Declaring oneself a member of the clergy, an owner or an executive increased the likelihood of panic selling, as did having self-declared excellent investment experience.
  3. The likelihood of a panic sale increases with the percentage of daily trades made by the investor. 
  4. An investor will be more likely to panic sell if options compose a larger proportion of the entire portfolio.