JP Morgan is using machine learning to automate the hedging of some equity options, a move that one quant calls a “game-changer”. 

The bank started using machine learning to hedge a portion of its index vanilla flow book last year. Since then, it has been able to hedge its exposures faster, and quote higher volumes as a result. 

“The real advantage is we are able to increase volumes quoted – because we are faster,” says Hans Buehler, global head of equities analytics, automation and optimization at JP Morgan in London. “If you have to manually manage this, you have to divert somebody’s time and sit them down to focus on it.”

One senior quant calls JP Morgan’s approach a “base-level rethink” of hedging, which he says will benefit illiquid markets in particular. He estimates the technique has the potential to cut hedging costs for certain commodity derivatives by as much as 80%.

“There are lots of places in the market where there is either illiquidity in the hedging instruments you have or large transaction costs or products that have risks that are unhedgeable,” says Mark Higgins, chief operating officer and co-founder of Beacon Platform in New York and co-head of JP Morgan’s quantitative research team until 2014.

“In those places, it will be a real paradigm change in how people can approach optimal hedging,” he says.

It will be a real paradigm change in how people can approach optimal hedging

Mark Higgins, Beacon Platform

JP Morgan’s new hedging program uses complex statistical regression, a type of machine learning that tries to find statistical relationships between variables by trawling through large amounts of data. The technique relies on historical market data rather than risk sensitivities – or Greeks – to estimate hedging costs, a dramatic shift from the popular Black-Scholes model. 

“Black-Scholes Greeks were very useful in the 1980s because we didn’t have a ton of data and we didn’t have a ton of computing power. So this was an approximation that worked very well for a long time. Today, we have much more data. If you revisit the problem of hedging derivatives now, I don’t think you would sit there and build the Black-Scholes model,” says Buehler.

Machine learning models consider many more variables and data points when making hedging decisions, and can generate more accurate hedges at greater speeds, he says.

“We can incorporate much more information into that process. Because it is purely data driven, we can use signals, market information and flow information.”

For instance, commonly used volatility models, such as the local volatility model, struggle to effectively capture the impact of transaction costs on profit and loss (P&L).

“The impact of transaction costs on a local volatility model is very hard to do in an analytical sense,” says Buehler. “In our approach, it is basically built in. So you know immediately what the implementation costs are, which in classical finance is hard to achieve.”

JP Morgan has used similar machine learning models to provide optimal execution for clients in cash equities for nearly two years. The bank plans to roll out comparable technology for hedging single stocks, baskets and light exotics next year.  

Quants are embracing so-called model-free machine learning techniques, such as complex statistical regression, to solve sticky problems. These approaches attempt to identify patterns in data without necessarily trying to explain the results within an existing model framework. More advanced methodologies are also in the works.

Deep Hedging

Buehler was one of the co-authors of a recently published paper on so-called deep hedging. He says the contents of the paper are part of an “ambitious project” at the bank, which will allow it to hedge positions multiple time-steps ahead. This means it can provide hedges along a path of times rather than just a single period of time in the future.

The research has already attracted interest from others in the industry who are eager to apply it in their own businesses. Beacon Platform, for instance, is researching the application of deep hedging in commodity derivatives and variable annuities.

For example, an investor that buys the rights to run a natural gas storage facility would need to determine the best way to delta-hedge their exposure to commodity prices. They could hedge at the storage location, where transaction costs might be high, or take on some basis risk and hedge at a more liquid hub location. Deep hedging techniques can tell the investor how to distribute their hedges between these locations.

“Deep hedging for that situation gives you a quantitative model that, when you actually introduce basis volatility and transaction costs at different locations, tells you interesting things about how you spread the delta-hedges in different locations and gives you a better business outcome,” says Higgins.

The same holds true for variable annuities sold by life insurance companies, which provide a variable payout at retirement based on the investments of the buyer. These products carry hard-to-hedge risks, such as mortality or early redemption. Deep hedging can be used to more accurately model these variables.

Higgins says his firm’s research found that for natural gas storage, deep hedging lowered transaction costs by 50–80% compared to the standard risk-neutral hedges, depending on the market structure, while the standard deviation of P&L was 50–90% smaller.

Other quants say these sort of techniques could mark a frontier in pricing and hedging derivatives.

“There is no fundamental physical law that governs risk factor dynamics,” says Alexei Kondratyev, a managing director at Standard Chartered in London.

“Our best hope is to take the vast amount of available data and analyze it in its entirety, without imposing any convenient mathematical models. This is why keeping the hedging model free and Greek free is such an appealing proposition.”

This is not to say the Greeks will be consigned to the scrapheap. Buehler says they still form the backbone of risk limit-setting for books that are automatically hedged.

“We still use the Greeks as a risk control. We have limits on vega, we have limits on term structure and so on. But in terms of the hedging decision, it’s no longer used,” he says.