## Tuesday, December 22, 2015

### Looking beyond the label algorithmic trading'

At the EMF 2015 conference, I attended a talk by Pradeep Yadav, where he presented a paper: Raman, Robe, Yadav, 2015. This paper analyses algorithmic traders ("AT") and manual traders ("MT") on one of the world's largest electronic limit order book exchanges, the National Stock Exchange ("NSE").

NSE is an ideal laboratory for studying these questions, as it is a simple plain limit order book market, without the confusion caused by market makers. In addition, the overall equity market structure is simple, with two exchanges (NSE and BSE) where NSE has dominant market share. The complexities of the fragmented order flow and multiple trading venues, of the US, is not present. At NSE, both spot and futures trade in the same time zone, with orders emanating from the same co-location facility, which also facilitates research. NSE data has thus shaped up to be a very nice foundation for understanding markets, in recent years, with some of the cleanest microstructure work getting done here.

#### The puzzle

Pradeep Yadav and his coauthors find that under conditions of market stress, ATs withdraw from the market. This immediately lends itself to a pejorative interpretation: "ATs are good for liquidity under good times, but are quick to withdraw when the going gets difficult, and this is creating a new set of problems".

I wondered how this squares with a powerful result from Aggarwal & Thomas, 2014. This is a modern causal econometrics paper, and in this they find that the incidence of mini flash crashes goes down when there is more AT. They look for mini flash crashes defined as 2%, 5% and 10% declines of the price within a five minute window (see Table 6, page 28). All three coefficients are negative; more AT gives fewer flash crashes. For the 2% and 10% case, the differences are not statistically significant, but for the case of a 5% drop of prices in 5 minutes, bigger AT gives a statistically and economically significant decline in the incidence of flash crashes.

Both papers seem to have persuasive empirical strategies. How do we square the results? How is it that ATs are more likely to step away in difficult times (Raman, Robe, Yadav) but at the same time how it is that when there is more AT, mini flash crashes are less frequent (Aggarwal, Thomas)?

#### A better classification system

In order to figure out what's going on, I think we should break with the classification AT vs. MT. Instead, it's better to think in terms of simple, mechanistic trading strategies vs. complex strategies that involve human judgment. For the purpose of argument, let's call these "SI" for simple vs. "CO" for complex strategies.

Let's start with the old world, before algorithmic trading. In that world, we very much had many humans running SI strategies and many humans running CO strategies. Technical analysis', and other trend following mechanistic strategies, were around well before algorithmic trading came along!

As Friedman, 1953, reminded us, there is a Darwinian process at work where speculators who lose money tend to exit the market. Because markets are competitive, the dumb adherence to a SI strategy would induce losses, and the people who did this would exit the market. Hence, the only sensible approach for a trader who uses a SI strategy is to either stop some times (i.e. have a "kill switch") or switch to a CO strategy at certain times.

In the good old days, SI speculators had "kill switches". When the market got weird, they would just stop trading. Nobody expected a simple trend following speculator to behave unchanged when volatility changed or when big news broke. We had trading floors where a boss would shut down some strategies from time to time. This was akin to a "kill switch" applied to a large number of SI strategies.

All that has happened with algorithmic trading is that we now have powerful clerks, i.e. computers, who are the foot soldiers implementing SI strategies. Nothing else has changed. Humans are still in charge!

Some human traders keep a swarm of SI strategies under leash, and when market conditions get difficult, they hit the kill switch. Some of them switch to CO strategies when the going gets difficult, and because CO strategies are much harder to program, they may well do this trading by hand. Some bosses of trading floors yank hundreds of SI strategies when the going gets weird.

#### Resolving the puzzle

There were always SI strategies and CO strategies. These have been around ever since organised financial trading began. In the older data, we are not able to disentangle the two.

In recent years, the SI strategies have gotten automated. We have reduced the use of humans in mechanistic tasks, and got computers to do this clerical work. For the first time in human history, we are now seeing a flag on orders where orders from SI strategies are now called "AT" orders. The CO strategies continue to be mostly done by hand, as it's quite hard doing this programming.

There is nothing wrong or unusual in SI strategies backing out of the market when conditions become confusing. SI strategies can only work in peaceful times. A trader who ran SI strategies all the time would exit the market as his wealth would run out (Friedman, 1953).

SI strategies done through AT give us more eyeballs looking at the millions of traded products in the modern exchange environment. When there's a dislocation in a market (e.g. a crash in the futures price), immediately, hundreds of traders come through with a mechanistic response (reverse cash and carry arbitrage), which stabilises the price. In contrast, in the manual world, the field of view of each human was limited, and when a little crash got started, there were fewer people available to interfere with it. This gave more mini flash crashes in the pre-AT world. This is how both statements are correct:

1. In times of market stress, the AT orders shy away (Raman, Robe, Yadav, 2015)
2. Greater AT intensity reduces the incidence of mini flash crashes (Aggarwal & Thomas, 2014).

### References

Nidhi Aggarwal, Susan Thomas. The causal impact of algorithmic trading on market quality, Working Paper, 2014.

Milton Friedman. The case for flexible exchange rates. In Essays in Positive Economics, The University of Chicago Press, 1953.

Vikas Raman, Michel A. Robe, Pradeep K. Yadav. Man vs. Machine: Liquidity Provision and Market Fragility, Working Paper, 2015.

Please note: LaTeX mathematics works. This means that if you want to say $10 you have to say \$10.