Computers around the world are ready to sell, sell, sell! I don’t mean on-line stores, but stop loss orders. At first a stop loss order sounds great, if not a necessity: if you hold stock at $100 and you want to protect yourself, you set up a standing stop loss order that if the stock drops below $98, you automatically sell. That way you are “guaranteed” to keep your equity. Many financial sites claim that you’ve got to use these orders to be a smart investor. However, with a tiny bit more analysis, you’ll realize that they’re utter insanity, because the policy does not scale. I claim that it caused the 995 point drop in the DOW yesterday. Here is coverage from CNBC — notice how fast the drop came and went:
What do I mean that “the policy does not scale?” Imagine that a significant portion (say 20%) of a particular stock is under stop loss orders, but then a mistake (CNBC claims that this flash crash way have been caused by a sell mistake) causes the stock to drop a bit. Then 20% of the stock gets suddenly sold in an automated stampede that causes the stock to drop like a rock. From a game theoretic perspective, this is a great example of a strategy that seems ideal for an individual (automatically sell when the price drops a bit), but when expanded to a large group leads to a horrible result. This effect is so bad and so strong that for many years the NYSE had trading curbs that automatically stopped many types of computer orders if the market dropped (but for some reason curbs were stopped in 2007).
If you are a long term trader with a stop limit order, you may have been automatically caught up in what essentially amounts to a crazy day-trading glitch. Fortunately for you NYSE canceled many “clearly erroneous” trades. Alas, that will probably let this amazing drop go by with a minimal amount of analysis.
So what does this have to do with cloud computing? If you implement any sort of automated policy by autonomous agents you must make sure that your policy scales if a large portion of the population adopts it. A traditional example is deadlock acquisition: if you acquire locks in a simple greedy way, it is easy to end up in “starvation” situations. We computer scientists have known this for decades, but because of Internet and cloud computing, we need the ability to apply this wisdom to a much wider set of problems.
With wonderful irony, as the market is wildly fluctuating in the CNBC clip, Cramer talks about making $500,000 if he had placed a buy limit order on P&G. Automated buy limit orders probably led to the amazing rapid rebound. In addition, this points to a policy that does scale: buy low and sell high. A more nuanced version of this policy is: determine a goal distribution of your holdings and periodically correct your actual holdings to match. If applied universally, such counter-cyclic policies can lead to system stability.