Media Management Center      MediaInfoCenter      McCormick Fellows      Kellogg School of Management      Medill

Noted by the Readership Institute
Wednesday, September 10, 2008

Digital media expert: United Airlines story shows how software combined with human error can have frightening results

(Rich Gordon)

It started as a ripple in the world's ocean of information - a six-year-old news article that was inadvertently discovered on a Florida newspaper's Web site. That little ripple would ultimately build into a tsunami that brought down a small corner of the global economy.

The story of the Sun-Sentinel, Google News and United Airlines is a story of human error, to be sure. But it's also a story about some troubling realities: lives that are increasingly shaped by computer algorithms, a networked economy in which speed is valued over deliberation, and software that can't possibly account for every unforeseen combination of events.

Still, I don't think the most important lessons of this story are about Web sites, journalism, software or human judgment. The lasting lessons should relate to the regulatory oversight of the world's increasingly automated global economy. However blame for the United Airlines debacle is apportioned, something is fundamentally wrong when investors make substantial bets on a company's stock based on a news story that anyone following the company closely would know was incorrect. I'm left with many questions. Who was the first trader to act based on the six-year-old article? Why did that trader's action lead others to follow suit? To what extent were these trades executed by human beings who could have applied their knowledge or intution, and to what extent were they executed by software following (always imperfect) algorithms? Can regulators limit automated trading - and even if they can legally prohibit this practice, is it realistic to expect that violators can be identified and stopped?

From a personal standpoint, this story brings back some unsettling memories. Back in the late 1990s, when I directed The Miami Herald's online publishing operation, we once inadvertently posted Fidel Castro's obituary (pre-written for the day of his death, which still hasn't arrived a decade later) to the live Web site. This was before Google News, so the mistake went mostly undiscovered. We were able to remove the article after just four people had viewed it. Fortunately, no one told Romenesko - let alone made stock trades based on the news.

Because of my experience with Web publishing systems that operate based on a combination of algorithms and human judgment, I find it pretty easy to understand the sequence of events. The account that follows is based on the facts as they have been reported, along with some informed speculation on my part. I am indebted to Steve Myers of the Poynter Institute, who pulled together most of the pieces in a post late Tuesday on the Poynter Institute Web site. As near as I can tell, the chain of events involves at least two human errors and at least two flawed algorithms.
  1. At some point, perhaps years in the past, someone at the Sun-Sentinel published an article on the paper's Web site about United's filing for bankruptcy in 2002. Google released a screenshot of the article as it appeared on the Sun-Sentinel site. Typically, six-year-old stories would not remain permanently on the Sun-Sentinel's Web site - they would be automatically deleted after a couple of weeks and users would have to search the paper's paid archive to find them. I am guessing that this article was published as part of a package of stories on United Airlines and its troubles, and left there largely unnoticed for weeks, months or even years. The first human error: Whoever posted the story didn't put a publication date on the article. This might seem an unforgivable mistake, but I guarantee you there are many other articles on news Web sites today that don't have publication dates - or, even worse, that have inaccurate publication dates. For instance, some publishing software would put automatically insert today's date on an old article in which an editor made a small change today.

  2. In coverage of this story to date, the biggest unanswered question has been why this old article happened to get so much attention this week. I think the answer can be found in the timing of the key events. They occurred at a very interesting time (according to the Tribune Co.): between 1:00 and 1:36 a.m. Eastern time on Sunday morning, September 7. Based on my experience, this would be the time of the week when a local newspaper Web site would have its lowest total traffic. For reasons we don't now know, one or a few people (the Sun-Sentinel's server logs might contain information that could clarify this) were reading the article on the Sun-Sentinel's Web site around that time. Maybe a college student was doing her homework. Maybe a journalist was doing research for a story. Maybe a United employee - or a competitor - was checking out how the company's bankruptcy had been covered in Florida. Or, as the Wall Street Journal suggests, perhaps site visitors looking for information on travel delays found the bankruptcy article via the site's search function. However many people viewed the story (and it really could have been just one), that traffic was sufficient to rank the article fifth among the most-viewed Sun-Sentinel business news articles at the time (look on the bottom right of this screenshot from Google). The first flawed algorithm: The software that chooses the most-viewed stories on the Sun-Sentinel's Web site. It probably should base the ranking on a longer time span than usual during low-traffic periods, and/or cut off the most-viewed list at a viewership higher than this article had. One might also argue that the most-viewed list should ignore old, or undated, articles. But there's no reason in particular that old articles should be left off the most-viewed list if they really are suddenly popular. (Update, Sept. 11: The Tribune Co. has confirmed that my speculation was correct - just two clicks were enough to put the 2002 story on the most-viewed list in the Sun-Sentinel's business section.)

  3. By the Tribune's account, the "Googlebot" spider or crawler visited the Tribune's business section around 1:37 a.m. Eastern time Sunday (10:37 p.m. Pacific time Saturday). The Googlebot found the United article on the list of most viewed business articles and recognized it - presumably via a comparison with its record of the site on its previous visit (20 minutes earlier, according to Myers' Poynter post) - as a newly posted news article. The result: Google News listed the article (screenshot provided by the Tribune Co.) with a Sept. 6 date. The second flawed algorithm: The Google software that decides whether an article published to a news Web site is really news. It probably should ignore undated articles. But I'm betting that there are other news sites that at least occasionally fail to include publication timestamps - or, even worse, are dated inaccurately. From Google's perspective, it's not unreasonable to assume that an article newly appearing on a news site's most-viewed list is, in fact, news.

  4. According to a Tribune article published this morning, people continued to view the article during the day on Sunday, keeping it on the list of most-viewed articles for the day. This was a circular, self-fulfilling prophecy: People read the article because it was listed among the most-viewed articles, which reinforced its traffic numbers - again, at a time of the week (Sunday) when usage of a local newspaper's business section would be low. Traffic generated by Google News - especially, people who had signed up for alerts from Google for news about United Airlines - would have added to the traffic.

  5. On Monday morning, a reporter for Income Securities Advisors, a news service that provides information to "Bloomberg terminals" used by many professionals in financial markets, did a Google News search for "bankruptcy 2008" and found the United article. She wrote a quick bulletin, which was reviewed by an editor before being posted to Bloomberg around 10:53 a.m. Eastern time on Monday. The second (and perhaps third) human error: Neither the reporter nor her editor apparently had the knowledge or background to recognize the clues that it was an old article - for instance, outdated financial figures. Richard Lehmann, president of Income Securities Advisors, said he couldn't fault the reporter, who was just doing her job, Myers wrote. (The Washington Post, though, quoted Lehmann as saying, "It would have been nice if the reporter had been more grounded in what's going on out there in the world.") It's a basic principle of good journalism that some articles - for instance, those that could move financial markets - deserve extra scrutiny because the risks posed by inaccurate information are so high. A story about a bankruptcy of a company as large and well-known as United ought to get an extra layer of review, but given the premium placed by investors on getting information fast, it is perhaps understandable why that didn't take place.

  6. Now comes the part of the story that to me is the most important. The distribution of the story via the Bloomberg network caused some traders to dump the UAL stock. Others then saw that the stock price was falling and sold UAL as well. Driving this process were more algorithms - software that can generate stock trades automatically based on rules. For instance, "sell UAL if it drops more than 5 points," or "short-sell UAL if trading volume exceeds this limit." As the Journal reported this morning, "The damage was exacerbated by the growing use on Wall Street of automated programs that trigger stock trades without any human interaction. The so-called algorithmic trading mechanisms, which buy and sell stocks based on news headlines and earnings data, were responsible for roughly a quarter of New York Stock Exchange trades in the last week of August." Is there any limitation on automated trades such as these? Are there protections in place - either through software or regulatory requirements - that would prevent a future information ripple from having even greater consequences than this dramatic plunge in the value of a single company's stock?
Oceanographers have a term for an enormous wave that appears suddenly and unpredictably in the middle of the ocean. They call it a "rogue wave." In the ocean where information and high finance converge, the massive selling of UAL stock on Monday was definitely one of those waves. Many different things had to go wrong at exactly the right time for these enormous consequences to result.

Still, it seems clear that there are other scenarios in which inaccurate or misinterpreted information published to the Web could affect investors' behavior. It also seems clear that there is a group of investors these days who succeed or fail based on how quickly they can act. These investors must increasingly rely on software to trigger trades automatically based on market trends or news events. Could the next string of flawed algorithms and human error have an even more catastrophic result? I'm hoping to hear that regulators are trying to make sure the answer is no.


By Rich Gordon (richgor-at-northwestern.edu)
Rich Gordon is an associate professor at the Medill School and director of new communities at the Media Management Center at Northwestern University.

AddThis Social Bookmark Button

Permalink
Posted at 9:40 AM
Email this post:


Comments:

Post a Comment


Links to this post:
Create a Link


Get Smart Blog Main Page







Search the Noted by the Readership Institute blog

©2008 Readership Institute • 301 Fisk Hall • Northwestern University • 1845 Sheridan Road • Evanston, IL 60208-2110
phone: 847.491.9900 • fax: 847.491.5619 • email: institute@readership.org