The other day CFTC Chairman Timothy Massad gave a speech about “flash events” in futures markets that has attracted a lot of attention. Most of the attention was given to Massad’s claim that there had been 35 flash events in WTI futures this year, and between 9 and 25 events per year combined in corn, crude, e-minis, 30 year Treasuries, gold, and the Euro from 2010-2014. Flashy results indeed. But the method for identifying them is misleading, and makes big flash moves seem more likely than they really are.
These results, and specifically the WTI finding for 2015, is an artifact of the definition of a flash event (which Massad acknowledged is somewhat arbitrary):
[E]pisodes in which the price of a contract moved at least 200 basis points within a trading hour— but returned to within 75 basis points of the original or starting price within that same hour.
The problem is that the number of flash events will depend on volatility. Two percent moves are more likely in high volatility environments, or for high volatility contracts.
This is clearly what’s going on in oil. As this chart of the oil volatility index (OVX) shows, oil volatility was extremely low through most of 2014, but increased sharply in late-2014 through mid-2015, and then has picked up again in recent months:
With volatility in the 60-70 percent annualized range, you will have a much greater likelihood of a 200 basis point move (and a subsequent 125 bp or so reversal) than with 15 percent vols. The flashy 2015 crude oil results are a reflection of this year’s high underlying volatility, which has been fundamentals driven, rather than the microstructure of modern electronic markets.
The 200/75 basis point standard was chosen because that’s what happened in the Treasury market on 15 October, 2014. But a 200 basis point move in something like Treasuries, which have a volatility of around 10 percent, is a bigger number of standard deviation move than a 200 basis point move in crude, especially with a volatility of 70. So the more appropriate cutoff would have been standard deviations (sigmas) rather than percent. But if Massad had done that, he would have identified a lot fewer events, and his speech would have been met with yawns, rather than the attention it has received.
Let’s also put things in perspective. The contracts considered trade 17-23 hours per day. 252 days a year times (say) 20 hour per day times 6 contracts and 20 events/year gives the odds of a .06 percent of an event in any hour. Using a more realistic sigma standard would reduce the odds of an event comparable to the Treasury flash event to a much smaller number than that.
Put differently, the Treasury event was truly anomalous, and Massad’s way of analyzing the data makes it seem more common than it really is. To get a flashy, eye-catching result, Massad had to use a misleading standard to identify flash events. Objects in his mirror are smaller than they appear.
The taking off point for Massad’s speech was the report on the 2014 Treasury flash crash. Like the infamous May, 2010 equity flash crash, there was a sharp decline in liquidity leading up to the price break. Massad attributes this to the way algorithms are programmed:
We also know that as with humans, the modern algorithms have risk management capabilities embedded within them. So when there is a moment of sudden, unexpected volatility, it may not be surprising that some in the market pull back – potentially faster than a human can.
The report describes how on October 15, some algos pulled back by widening their spreads and others reduced the size of their trading interest. Whether such dynamics can further increase volatility in an already volatile period is a question worth asking, but a difficult one to answer. It is also very difficult for individual institutions of any type to remain in the book, opposing price headwinds, or worse, to try and catch the proverbial falling knife. For many, this decision can be the difference between risk mitigation and significant losses. Contrary to what some have suggested,
This makes perfect sense. Some algorithms-especially HFT algorithms-attempt to determine when order flow is becoming toxic (and hence adverse selection risks are elevated) and reduce exposures when they do. Holding depth constant, greater information flow makes prices more volatile, and the reduction in liquidity that the greater information flow causes makes prices even more volatile.
This means that looking at the depth reductions and associated increases in volatility focuses on a symptom, not the underlying cause. What deserves more attention is what causes the increase in the informativeness of order flow that makes the liquidity suppliers cut back. This hasn’t been done in any study, to my knowledge, nor is it likely to be possible to do so.
And as Massad notes, this phenomenon is not unique to electronic markets. Meat puppet market makers also take a powder when adverse selection risks rise:
Contrary to what some have suggested, I suspect it was difficult for market makers in the pre-electronic era to routinely maintain tight and deep spreads during volatile conditions. They likely took long coffee breaks.*
It’s beyond suspicion, actually. It happens. Look at the Crash of ’87 when locals fled the pits and OTC market makers stopped answering their phones.
These reductions in liquidity are inherent in any trading environment where private information is important, and the rate of information flow varies. Regardless of trading technology or market microstructure, liquidity suppliers will cut the sizes of their quotes, or stop quoting altogether, when order flow turns very toxic.
Given all this, Massad’s policy prescriptions are oddly disconnected from the flash phenomenon that prompted his talk:
The focus of our forthcoming proposals will be on the automation of order origination, transmission and execution – and the risks that may arise from such activity. These risks can come about due to malfunctioning algorithms, inadequate testing of algos, errors and similar problems. We are concerned about the potential for disruptive events and whether there are adequate measures to ensure effective compliance with risk controls and other requirements.
Now of course, you could have errors before, in the days of pit traders and specialists. You could have failures of systems in less sophisticated times. But generally the consequences were of lesser magnitude than what we may face today. And that’s in large part because the errors were easier to identify, arrest or cure before they caused widespread damage.
I expect that our proposals will include requirements for pre-trade risk controls and other measures with respect to automated trading. These will apply regardless of whether the automated trading is high or low frequency. We will not attempt to define high-frequency trading specifically. I expect that we will propose controls at the exchange level, and also at the clearing member and trading firm level.
That’s all great, but really beside the point. If rogue or fat-fingered algos were the problems in any of the alleged flash events Massad identified (including the Treasury event of a year ago), he would have been able to say so. But he admits that the causes of the various events are all unknown. So it’s a bait-and-switch to pose the problem of flash crashes, and then advance remedies that have nothing to do with them. It’s the regulatory equivalent of applying leeches.
In sum, Massad overstates the flash event problem, and offers policies that have nothing to do with them. The fact remains that these things are probably beyond a policy fix anyways. They inhere in nature of the trading of financial instruments when order flow can become toxic.
The crucial point is that these automated trading programs — like Hal — lack human judgment. When a crisis erupts and prices churn, computers do not simply “take a long coffee break”, as Mr Massad says, and wait for common sense to return; instead they tend to accelerate trading, fuelling those flash crash swings.
Sheesh. Please read, Gillian. Massad’s point is that the algos do take a metaphorical coffee break. They don’t speed up, they pull back.