Facebook owner Meta scored its biggest ever one-day loss in market value on February 2 this year, shedding a fifth of its worth after a disappointing earnings announcement. Two days later, Amazon recorded the biggest one-day gain—a 13% rise—after a positive announcement.

Such moves are the stuff of nightmares for options traders, who must hedge against the price moves such surprises can cause.

Quants at Citi have built a model that can help.

“It gives us a better idea of how our book is going to perform in an earnings event for both expected earnings moves and unexpected moves,” says Henry Yeh, the bank’s head of single stock flow derivatives trading.

Citi’s model forecasts the shape of implied volatility surfaces for its single stock options book under multiple scenarios, each relating to different spot moves after an announcement.

The model uses gradient boosting, a type of machine learning that iteratively improves a tree-based classification model, training on spot price data and changes in the shape of the implied volatility surface. Citi’s quants retrain the model weekly.

A team led by Anissa Dhoubi, Citi’s global head of equities quantitative analysis, started work on the project in 2020—initially submitting the solution to an MIT research initiative before building the model in-house and rolling it out to the firm’s US traders at the start of this year. The bank is now trialing the tool in Europe and Asia.

Every day, the tool emails Citi’s traders with forecasts of implied volatility surfaces for different spot moves in names with earnings due.

The forecasts tell traders how different moves in the underlying would affect the P&L, as well as the sensitivities of the firm’s options book—its delta, gamma and so on.

It gives us a better idea of how our book is going to perform in an earnings event.

Henry Yeh, Citi

On any given day, as many as 20 names in an options dealer’s book might have an earnings announcement due, says Yeh. And it can take as much as an hour for a trader to figure out the likely impact of each on the volatility surface. With the bank’s new tool, the task takes two minutes.

The biggest advantage is “simplification,” Yeh says. “The number one challenge in flow trading is to take a large amount of information and process it seamlessly and efficiently.” The new model frees up traders to spend time on higher-value activities, he says.

The tool also helps traders judge how to quote for new trades coming into the book shortly before announcements, says Thomas Fouret, North America head of equity derivatives quantitative analysis.

“If a trader gets a call 15 minutes before the closing bell for a quote on a name with an upcoming event, they want to be able to say how the book would look if they took the trade on. Is it a risk they are willing to take? How much should they charge?”

Implied volatility typically falls after an earnings announcement but rises on unexpected news, such as a CEO stepping down. The machine learning model is better than parametric models at accounting for current market conditions when forecasting the effect of such surprises, Citi’s quants say.

“If the earnings event was expected to lead to a 6% move and a stock moves 6%, our traditional methods do a pretty good job of handling such a scenario,” Yeh says. “If spot happens to move 12% because of some positive or negative news—that’s when the machine learning model shines.”

Retail nudge

Citi started work on its earnings forecasting tool partly because big price moves on earnings days have become more common in recent years. Changes greater than 20% around earnings occurred more than a hundred times in 2018, 2019 and 2020, compared with around 60 times a year in the preceding eight years.

Partly this is due to the rapid growth in retail options trading. A study from researchers at MIT and Stanford shows pre-earnings trading in Nasdaq single stock options by retail investors more than tripled from 2019 to 2021.

“Forecasting volatility around earnings has always been tricky,” says Garrett DeSimone, head of quantitative research at options data firm OptionMetrics. “Earnings are known to cause jumps in stock prices, which can alter the volatility surface dramatically… [both] the level of volatility [and] the shape of the volatility smile.”

Retail traders can drive up the volatility of popular stocks ahead of earnings, exacerbating the post-earnings stock response, DeSimone adds.

Market-makers that are most accurate in predicting implied volatilities are able to make tighter prices, says Mark Higgins, co-founder and chief analytics officer of Beacon Platform, a cloud-based trading and pricing engine, and a former co-head of quantitative research at JP Morgan. “That leads to higher market-making revenues.”

Citi’s project follows steps from others in the industry to automate the work of traders. JP Morgan has developed a suite of machine learning hedging tools that allows it to automate or semi-automate hedging of flow and some exotic derivatives books.

Jason Sippel, JP Morgan’s head of global equities, described that project to Risk.net, a sibling publication of WatersTechnology, in February as a “top-level reimagining” of what the firm wants trading to be.