Minority Report, the 2002 Steven Spielberg film starring Tom Cruise, is set in 2054 when psychic “precogs” see crimes before they’re committed, allowing police to intervene pre-emptively.
If a client’s greatest crime is panic selling, new research may be as welcome to advisors as a clairvoyant in a sensory deprivation chamber: What if it were possible, as with the precogs, to identify the panic sellers before they acted?
Researchers from the Massachusetts Institute of Technology have developed a machine learning model for this purpose. “Subtle patterns in portfolio history, past market movements, and demographic profile can be exploited by deep neural networks to accurately predict if an investor will panic sell in the near future,” the paper published in August states.
To develop this predictive capacity, the researchers examined a dataset of more than 650,000 accounts from one of the largest brokerage firms in the U.S. between 2003 and 2015. The profile they developed of the panic seller may not be what most readers expect.
The authors defined a panic sale as a 90% decline in a household account’s equity assets in one month, with at least 50% of the decline due to trading. Panic selling during periods of market uncertainty is given the highly technical term “freakout.”
The good news is that freaking out is uncommon. The authors counted 36,374 panic sales by 26,852 households (9% of all households) over the 13-year study period, accounting for only 0.13% of the more than 25 million data points they examined.
Panic sales were most common among those with small accounts (43.2% among accounts with less than US$20,000), but there were still 696 such events among millionaires.
Investors over 45 were more likely to panic sell than younger investors. Men were slightly more prone to freaking out than women, but women were more likely to panic sell in general (outside of volatile periods). Those with more dependents were more likely to liquidate their portfolios.
Interestingly, those who self-declared their investing experience and knowledge to be excellent were more likely to panic sell. (Advisors may be nodding along, recognizing that more humble clients are better at following advice and staying invested.)
Broken down by occupation, panic sellers were most commonly listed as “self-employed,” “owners” and “real estate,” while the least likely to panic sell were “paralegal,” “minor” and “social worker.”
Using this demographic information, as well as market conditions and portfolio history, the authors developed a machine learning model to predict panic selling one month in advance. Prediction is difficult because panic sales are so rare, but the best-performing deep neural network achieved a 69.5% true positive accuracy rate (predicting when panic selling would occur in the next month) and an 81.2% true negative accuracy rate (predicting when there would be no panic selling).
The researchers also found that panic selling isn’t all bad: it can be beneficial as a stop-loss mechanism in plunging markets. Too often, however, investors sell when the market is improving, or they wait too long get back in. Almost one-third of investors never returned to the market after panic selling.
The disposition effect describes investors’ tendency to buy stocks with strong recent performance and hold those stocks when they become losing investments — which the authors pointed out is basically the opposite of panic selling. Good advisors also advise clients against overtrading, as various studies support a buy-and-hold strategy.
“When Do Investors Freak Out? Machine Learning Predictions of Panic Selling” by Daniel Elkind, Kathryn Kaminski, Andrew W. Lo, Kien Wei Siah and Chi Heem Wong from the Massachusetts Institute of Technology, published on the SSRN network, Aug. 9, 2021.