A co-founder of Quantigic Solutions has presented what he calls “the freshmanlevel” answer to the question: why did dollar-neutral quant trading strategies, such as equities stat arb, fail so markedly during the COVID-19 market sell-off?
Zura Kakushadze, a professor at the Free University of Tblisi, is on reasonably solid ground with his premise. Not all quant trading strategies lost money during the Covid-related March sell-off, but enough of them did poorly enough to raise the question. In mid-March, as the extent of the joint epidemiological/financial disaster was becoming clear, Millennium announced that it had lost 2.7% in just the first 12 days of the month. Millennium was hardly alone.
And there are those who find this confusing. After all, the “neutral” in dollar-neutral strategies is often understood to mean market neutral. Absolute return. Prepared for anything. One expects long-only strategies to be punished by market routs. One even expects this of long-mostly strategies such as 130/30 and their kin. This leads us to another of Kakushadze’s points—there is no such thing as a certain lunch. The mystification about the “quant bust 2020” arises from a misunderstanding about what such funds and strategies are trying to do.
Why Does Stat Arb Work When it Does?
Dollar-neutral strategies are those that hold equal long and short positions as denominated in dollars. For purposes of his explanation, Kakushadze focuses on statistical arbitrage. Why, in normal times, does stat arb work?
In this connection, too, Kakushadze has a (skeptical) point to make about data science and machine learning. He says that both fund managers and investors sometimes make the mistake of believing “that dropouts with no requisite knowledge of the financial markets, whose skills are limited to knowing how to call off-the-shelf Python libraries, will magically decipher the stock market.” What was once known as data mining, albeit with a derogatory connotation, is now dressed up respectfully as data science. The belief that dollar neutral can be market neutral even in very non-normal times, is bound up with this overestimation of the powers of data and algorithms.
Stat arb portfolios can consist of thousands of stocks since no human/discretionary research has to be done. Humans can only keep a limited number of balls in the air, so a long/short fund with a component of human discretion will on the other hand have “a modest number of stocks.” The algo-driven stat Aarb portfolios with which Kakushadze is especially interested have a “medium” holding period. Positions are in general held for between 1 and 20 days. Thus, his study excluded high-frequency trading funds, which have holding periods measured in milliseconds or less, and it excluded funds with holding periods of months or years.
The stat arb portfolio works in a dollar-neutral way during normal times because the strategy expresses the principle of reversion to the mean. “[W]e have factor returns for each industry, and the return for stocks belonging to a given industry are assumed to be mean-reverting around the factor return for said industry.” Thus, some stocks on a given day are higher priced than they should be given the factors, and others are lower prices than they should. The strategy is: short the rich and buy the cheap.
But in Non-Normal Times
During non-normal times, when volatility spikes, this strategy becomes undependable. At such times, the long-only guys suffer first, but they do not suffer in silence. Their liquidations do not proceed in an orderly fashion. The normal correlations go out of whack, and the otherwise predictable factor returns get out of whack. There is no “snap-back” within the medium holding period. Further, as soon as some of the positions in the portfolio start to lose ground for this reason, the algorithms hold up on more losing positions.
There comes a point at which the portfolio has lost more than the risk-tolerance parameters allow, which means that humans step in. The humans start to liquidate, usually in a haphazard fashion, and these liquidations further throw off the correspondences.
This sort of avalanche effect is the reason (explicable to college freshmen) why the dollar-neutral strategies melt down in high-vol environments.
Machines, Dogs, and Cats
Machine learning cannot prevent such meltdowns and can sometimes make their consequences worse. After all, Kakushadze writes, machine learning was developed for such uses as distinguishing an image of a cat from an image of a dog. This is a matter in which the entity being tutored reaches a successful equilibrium because, as others before Kakushadze have observed, dogs don’t turn into cats when machines learn which is which. An excess price on the other hand can turn into a market price, or even into a bargain price, because the machines have decided it is an excess price.
Further, as Kakushadze puts it, “most ML/AI algorithms trained on cats and dogs would mistakenly identify hyenas as canines.”