Kenneth Chang has reported in the NY Times: On my computer screen, a juicy bid pops up a buyer willing to pay top dollar for an item I’m selling. In the next moment, as I fumble to type and click, the bid flashes away. A competitor, moving as fast as a computer, has already closed the deal. And, in this case, it probably was a computer.
I’m sitting in a roomful of workstations at I.B.M.’s Thomas J. Watson Research Center in Yorktown Heights, N.Y., participating in an experiment on how well people compete against computers in auctions.
Researchers at the International Business Machines Corporation expect transactions like this to become common, and in a limited way, automated bidding robots or “bots” already exist. Auction sites like eBay allow people to specify the most they are willing to pay for an item, and their bids automatically rise as others bid higher.
But I.B.M. envisions more sophisticated situations and has established an Institute for Advanced Commerce to study them. Automated software, computer scientists at I.B.M. predict, will search different suppliers and negotiate prices with them. “We’re wondering what a world looks like when there are a billion of these software agents transacting business on our behalf,” said Dr. Steve R. White, who heads the research.
In the future, a purchasing agent looking to order pencils may enter some data say, quantity, maximum price, date needed and set the software loose on the Web.
That situation raises questions as well. Do these bidding bots bid differently from people? Could they change the behavior of financial markets? There is some evidence that they may; some blame computerized program trading, in part, for the stock market crash of 1987.
“They’re much faster than people,” Dr. White said of bidding robots, “and they’re much stupider than people.”
Pitting the bots against one another, the I.B.M. researchers found they often got into cutthroat price wars, relentlessly undercutting one another until one figures out that it can make more money by selling a few items at high profit and pops the price back up. Then the competitors follow suit, and the cycle repeats.”
The first time this kind of roller coaster price war occurred, the researchers thought it was “some dumb artifact of the simulation,” Dr. White said. But the same behavior kept on showing up in other simulations with different bidding rules. The bots, it appeared, were doomed to repeat the past, because they did not remember it.
The researchers then cured that self-defeating behavior by having them keep track of past pricing trends. “We’re trying to give them enough smarts so that we trust them,” Dr. White said.
The next question was, What happens when the bots go up against people? In a series of experiments, the I.B.M. researchers have looked at the fairly simple “double auction” similar to the way stocks are traded.
In a double auction, sellers put in “ask” bids specifying what price they’re hoping to receive for their wares. The buyers similarly bid what they’re willing to pay, and transactions are completed when a buyer accepts a seller’s price, or a seller agrees to a buyer’s bid.
In such auctions, the price quickly and almost inevitably converges to an easily calculable ideal. But how to gain the most in the fluctuations is unknown. “You just do the best you can,” said Dr. Jason Shachat, a visiting scientist at I.B.M. “We don’t know what perfect play is.”
This was the kind of experiment I joined. Twelve players were divided into six sellers and six buyers. For each group, half were people five others and and half were bots. The sellers as I am playing are peddling a mythical commodity, although the cost to produce each unit varies. The buyers are similarly given what is in essence the resale value of each item they purchase.
The human players sit in front of computer workstations. In the next room, another computer coordinates the auction, collecting and dispatching bid information to players. The bidding bots were 10 miles away at another I.B.M. building, communicating with the auction computer via the computer network.
In three-minute-long trading sessions, a flurry of buys and sells is followed by a lull when there are no more profitable deals to be made. Several times, I’m thwarted by typing in the wrong number, and by the time I enter the correct bid, the attractive price is long gone. “Snatched again by a bot,” I think.
I end up bidding $66 much of the time, in part because it seems to be close to the going rate, but also because it’s quick to type.
After about a dozen rounds, I retire to the back room, where the researchers analyze my performance.
I did reasonably well, they tell me, better than the other two human sellers and one of the bots.
But on average, the bot players were about 5 percent more profitable than the people, and that differential has persisted each time the experiment has been run.
The experiment was set up so that one of the bots shared my production costs, and that bot wound up making 10 percent more profit than I did.
Even though the bots are programmed with rather simple strategies, they excel in the double auctions, because they can pounce on someone else’s mistake and because they never make the kind of egregious errors selling something at a loss, for example that the human players are prey to.
And an increase in profit is “strong economic incentive to employ bots,” Dr. White said. “You will be at a severe disadvantage if you aren’t.”
Others think it will still be a while before the bots make major economic decisions.
The I.B.M. work is interesting, said Dr. John O. Ledyard, a professor of economics and social science at the California Institute of Technology, but he is skeptical that the bot advantage will play out in “thin” markets where there are only a few buyers or sellers, and where good deals are more a matter of skill than speed.
“That requires another level of intelligence, which is a harder problem,” he said. “Speed matters where you can take advantage of mistakes, but speed can hurt in negotiations.”