HFT as an insight into where fintech is going
Markets are cyclical, Darwinian and subject to diminishing returns. Nothing new in that observation. But people do seem to forget it all the time.
Case in point: high-frequency trading (HFT). Circa 2009-2010, algorithmic trading techniques were considered a veritable goldmine for those who knew and understood how to deploy them against dumb human competitors.
But as with most innovations, markets evolved. Dumb competitors got wise and began to employ similar strategies to defend their wealth stores. A rush of new competitors meanwhile turned up to eat the lunches of all the first entrants. In no time at all, the lucrative profits to be had from HFT began to encounter limits (fast-as-light trading, regulation, et cetera) bringing forward a new equilibrium.
...MUCH MORENow, tellingly, with their profits under threat, HFT companies are looking to invest in new techniques to recapture their edge. Much of this is focused on big data analytics, and of course AI and machine learning.
In a recent blog post Streetwise Professor Craig Pirrong, a market structure expert from the University of Houston, makes a valuable point about the nature of these sorts of cyclical transitions. Namely, they always happen. So don’t be surprised when they do.
Remarking on Alex Osipovich’s story about HFT traders falling on hard times, Virtu’s bid for KCG, Quantlabs buying HFT company Teza and Interactive Brokers exiting options market-making, he notes, with our emphasis:
Alex’s story repeats Tabb Group data documenting a roughly 85 percent drop in HFT revenues in US equity trading. The Virtu-KCG proposed tie-up and the Quantlabs-Teza consummated one are indications of consolidation that is typical of maturing industries, and a shift it the business model of these firms. The Quantlabs-Teza story is particularly interesting. It suggests that it is no longer possible (or at least remunerative) to get a competitive edge via speed alone. Instead, the focus is shifting to extracting information from the vast flow of data generated in modern markets. Speed will matter here–he who analyzes faster, all else equal, will have an edge. But the margin for innovation will shift from hardware to data analytics software (presumably paired with specialized hardware optimized to use it).What’s interesting is what Pirrong goes on to say about internalisation, incentives to keep investing in information collection and the related Grossman Stiglitz paradox which can take hold.
The short story is this. As markets get increasingly competitive they encounter physical limits, such as fast-as-light trading. At this point it sometimes becomes cost effective to invest in protectionist inefficiencies. The illegal version of this comes in the shape of cornering, manipulation or front-running....
Then, in the cheap seats, the comments veered hard into artificial intelligence and algorithms, and Holy Hannah, hang on for the ride. As I quoted in another context:
“Bang! Bang! Bang! Bang! Four shots ripped into my groin, and I was off on the biggest adventure of my life … But first let me tell you a little about myself....
Here's an example, Ms Kaminska jumping into the scrum:-Max Shulman Sleep Till Noon (Doubleday, 1950)
...An AI can never be all knowing. At best it will operate like a quantum leap ziggy machine quoting probabilities all the time. Which means it can still be gamed or duped by another AI, or by the underlying data.
In some ways this is regressive. Informed trading used to be about certain information. 100% style probability that x company is going to announce this or that. Ai informed trading is based on processing all the surrounding variables and deducing that x company is going to do this eventually (if it hasn't done it already). Given that wider data is available to everyone who can be bothered to process it -- and it's a much more energy intensive way to go about getting privileged info -- there's no guarantee it's not already been absorbed into the price. There's also a helluva lot more chance that you might be wrong.
An AI can't compete with a human who knows the CEO plans to do x or y because he casually told him over dinner.
And if it can that's only because he's deduced it from some random behavioral stuff based on exceptionally private information. Which might also lead the AI up garden paths all the time.
The only way an AI can be a sure bet, is the same way a human brain can be sure in its forecasting. Actually cause the future they are predicting....That would be a metaphysical bingo.
The most sophisticated quant stuff is, at its core, just probability calculating. What the 'puters are trying to do is become the casino. Getting self-referential* again, you never play a negative expectation game.
And once the math is in your favor you want to place as many bets in positive expectation situations as you can so that even though you might lose an individual bet, at the end of the day, week, year, you've gotten enough action down that the math is in your favor regardless of volatility, variance or whatever. Then you lever up, looking over your shoulder to make sure that Meriwether and the rest of the LTCM crew aren't within a thousand miles and start the machines a-runnin'
Of course you also need to practice risk management/bet sizing and maybe have Ed Thorpe, one of the quantfathers and a pretty good blackjack theoretician; on speed dial.
Here's our intro to a 2011 post on that aspect of the game, "Dreamtime Finance (and the Kelly Criterion)":
I've been meaning to write about Kelly for a couple years and keep forgetting. Today I forget no more.Now the thing is: I could do the same with almost every one of Izabella's comments as she chats back and forth with the readers of her piece. Just amazing.
In probability theory the Kelly Criterion is a bet sizing technique used when the player has a quantifiable edge.
(When there is no edge the optimal bet size is $0.00)
The criterion will deliver the fastest growth rate balanced by reduced risk of ruin.
You can grow your pile faster but you increase the risk of ending up broke should you, for example, bet 100% of your net worth in a situation where you have anything less than a 100% chance of winning.
The criterion says bet roughly your advantage as a percentage of your current bankroll divided by the variance of the game/market/sports book etc..
Variance is the standard deviation of the game squared. In blackjack the s.d. is 1.15 so the square is 1.3225.
As blackjack is played in the U.S. the most a card counter can hope for is a 1/2% to 1% average advantage with much of that average accruing from the fact that you can get up from a negative table.
Divide by 1.3225 and you've got your bet size.
It's a tough way to grind out a living but hopefully this exercise will stop you from pulling a Leeson, betting all of Barings money and destroying the 233 year old bank.
I'll be back with more later this week.In the meantime here's a UWash paper with the formulas for equities investment.
Where most managers and traders screw up is in overestimating your advantage, the math is a lot more straightforward in blackjack....
You want to talk about creating the future to make your bets come in?
How about spoofing the spoofers?
This is the stuff that's going on right now!
Anyhoo, the reason I said I was getting self-referential was the fact I've harped on the negative expectations thing (and Professor Thorpe) a few times:
2013: "It's The Math Stupid: When Your Junk Bond Yield Is Lower Than Your Expected Rate of Default You Have Locked In a Loss (JNK; HYG)"
2014: "What Proportion of Your Bankroll Should You Bet? 'A New Interpretation of Information Rate'":The first rule of gambling is "Don't play negative expectation games."
When you are playing a negative expectation game the answer, of course, is zero.
2015: "The High Stakes History of Card Counting (And Its Uncertain Future)"
One of the rules of life:
NEVER EVER play a negative expectation game unless forced.
And a few others I can't remember at the moment.
Re: Ms Kaminska- Read her post.