High Frequency Trading and market sociality: “it’s really machines all the way down”?
October 21, 2010
Following up on Daniel’s point on HFT missing from mythical accounts of Wall St, and continuing thoughts from an earlier post on the May 2010 stock market plunge: here is an interesting commentary from Ars Technica on High Frequency Trading and the crash. Bringing a Computer Science perspective, the article considers HFT as a case of large system:
the market as described in the SEC report looks like an awful lot like a giant, multithreaded software application. And on May 6, the market did what every piece of multithreaded software eventually does in response to just the wrong mix of execution conditions and inputs: it crashed.
The author then narrates the mechanism of the crash, based on the SEC report, as one where thousands of algorithms are interpreting each other’s response. The idea here is not just that machine trading might generate its own panics, but we also get a picture of how derivative markets and equities markets are linked–via the activity of HFTs and other algorithms–and how humans fit into this operation. For example,
The algorithms that buy and sell stocks in the equities market were using the previous few minutes’ action in the derivatives market as inputs to guide their trading, and when they registered the giant sell-off described above, some of them had safety controls that told them to stop trading so that the humans could take a look to see what was wrong. And as these algorithms pulled out of the market, the market got more illiquid and prices dropped faster.
A CS perspective can provide us with a new description which entails new questions:
to be a single multithreaded app, as opposed to an unrelated collection of multithreaded apps, the different threads must somehow interact with one another. In other words, the threads must share and jointly modify some kind of state. What state do the various apps and algorithms that run on Wall Street’s machines share? At the very least, every part of the market shares the quote feed.
The price of, say, AAPL at any given moment is a numerical value that represents the output of one set of concurrently running processes, and it also acts as the input for another set of processes. AAPL, then, is one of many hundreds of thousands of global variables that the market-as-software uses for message-passing among its billions of simultaneously running threads. Does it really matter that those threads are running on separate machines at different institutions?
It is worth reading the Comments section where people consider the proposition and debate whether HFT is really a multithreaded application or something else–a “complex adaptive system”, an “ecosystem”, a “capacitor”, a case of “control systems theory”, and so on.
Even if it’s not “machines all the way down”, as the author at one point suggests, these are discussions to which sociologists of finance should pay attention because they actually deal with the substance of current financial markets, as social spaces with interactions. The intricacies of how machines deal with each other should really become part of the fundamental understanding of market behavior (for a succinct argument see e.g. Callon and Muniesa 2005).