On Friday the Wall Street Journal reported that May 6’s stock market ‘flash crash’ was caused by a single trade made by a computer trading algorithm at a mutual fund company. The article highlights a potential flaw in the trading algorithm: that while most trading algorithms take into account trade volume, prices, and time to make the trade when making decisions, this algorithm considered only trade volume. Without the safeguards of pricing and timing decisions, the trade sparked a death spiral of high volume trading from high-frequency trading firms, causing the original algorithm to see even faster.
While the article highlights the concerns with automated trading algorithms and some potential flaws in this particular algorithm, it overlooks one of the inherent shortcomings of algorithms designed to make money for a specific entity: failure to optimize globally. When an environment, such as the stock market, has just a few computerized algorithms, they can operate autonomously with little effect on the rest of the players. But when the environment is dominated by computerized decision making, one decision can set off a chain reaction. The original ill-fated trade that launched the ‘flash crash’ set off a cascade of trading by high-frequency trading firms. These firms employ their own computerized trading algorithms to make many trades over a short period of time. The high volume of trading generated by the high-frequency trading firms then triggered the original trading algorithm to execute its trades even faster since the only factor it considered was trade volume.
While the automated trading algorithm that triggered the crash was clearly flawed by not considering enough data in its decision making, the bigger concern should be how the interaction between several trading algorithms led to the Dow Jones Industrial Average’s fastest decline ever. The crash was not caused by some young trader on the floor with a bad hunch or by the reaction to a world event. This was started by a computer that was created to be smarter, faster, and better than a human trader.
Sometimes the local optimum of a company (making huge profits) runs counter to the global optimum (maintaining the overall health and balance of the stock market). Computerized trading algorithms are not something that can be regulated since every company uses proprietary algorithms and no one is willing to give up their trade secrets or the potential competitive advantage that comes from being able to execute trades before humans can even process the data. And often the adverse effects of a decision like the May 6 trade could never be known in advance. But by adding guards to the software and understanding the potential impact their algorithms can have on the entire market, maybe trading algorithm developers and the operation researchers they work with can minimize the negative consequences of their local optimization.