尔街因不知道从错误中汲取教训而臭名昭著。或许机器能做 得更好一些。
这是越来越多的投资者的愿望。他们希望人工智能科学能帮助他们进行投资决策。
运 用人工智能技术时，程序员不仅仅把计算机设定成根据特定的输入命令做出决定。他们让系统能够学习各种决策，并做出相应的调整。大多数试用这种方式的投资者 均在使用“机器学习”技术，这是人工智能的一个分支，计算机程序对数量巨大的数据进行分析并对未来做出预测。许多技术公司均在使用这种技术，谷歌 (Google Inc.)用其匹配网络搜索的结果，NetFlix Inc.用它预测哪些用户很可能租看电影。
在华尔街的人工智能竞赛 中，Rebellion Research成为一颗冉冉升起的新星。这家仅有约700万美元资本的纽约对冲基金一直在用自己开发的机器学习程序进行股票投资。据知情人士称，这家公 司由一小组20多岁的数学和计算机奇才运营，业绩骄人，自2007年开始投资到今年6月份，除去各种费用，其年度业绩平均比标准普尔500股指 (Standard & Poor's 500-stock index)高10%。像许多对冲基金一样，它的目标是年复一年跑赢大盘。
Rebellion 公司人工智能系统的开发者、现年27岁的格林博格(Spencer Greenberg)说，很显然人类并没有进步，但计算机和运算法则却越来越快、越来越强大。
据称，纽约州的Renaissance Technologies LLC等部分老道的对冲基金已经开始运用人工智能技术进行投资。但多年来，这些公司均只是另类。一些曾涉足人工智能技术的公司怀疑这项技术要投入使用还差 得远。
许多新公司正在使用机器学习的方式进行交易，Rebellion公司只是其中之 一。拥有1,000万资产的旧金山对冲基金Cerebellum Capital从2009年开始使用机器学习的方式进行投资。据知情人士透露，德克萨斯州的RGM Advisors LLC和芝加哥的Getco LLC等许多高频交易公司正在使用机器学习功能帮助它们的计算机系统高效地买卖股票。
支持者说，程序效率高，因为它们能在短时间内处理数 量巨大的数据，学习行之有效的方法，并随时调整自己的策略。相比之下，典型的定量方式或许只运用单一策略，或甚至一次结合使用多种策略，但或许不会基于程 序确定的最有效方式在这些策略之间进行调整或修订这些策略。
过去的成功并不意味着Rebellion公司将继续跑赢市场。由于使用了许多定量分析策略，如果市 场基本面的变化导致其计算机程序──“星”(Star)出错，其系统可能停止工作。
格林博格说，“星”能够基于市场及更广泛经济活力的变 化调整其策略，这使其拥有了一定的智能，它不局限于任何单一的投资方式。在特定条件下，基金会买入便宜的股票，其它条件下，基金会青睐价格迅速上涨的股 票，或同时执行这两种策略。
与利用人工智能协助进行快速交易的高频交易基金不同，Rebellion倾向于长期持股。它的平均持股时间为 四个月，但有些情况下会超过两年。它也不做空股票或使用杠杆（即借钱）。后者虽然能扩大收益但也能增加风险。
这个程序监控约30个可能影 响股票表现的因素，例如市盈率或利率等。
程序经常要处理十年以上的市场历史数据以及最新的市场活动以决定是否买入或卖出一支股票。当特定 的策略不可行时，程序会自动添加这一信息，学习并调整投资组合。
例如，它或许会发现数据表明某些市盈率低的股票很可能上涨并买入这些股 票。然后，如果程序发现，由于它跟踪的因素出现了变化，这种策略很可能失去效果，它将清空这些股票，
并买入它认为更好的股票。
每 天早上，“星”会推荐一张买入或售出股票的清单，它推荐的清单经常会毫无变化。一位交易员实施这些操作。公司说它从未否定过计算机程序提供的建议。从 2007年开始使用至今，除少数几次修修补补之外，这个程序基本保持了原样。Rebellion公司通常同时持有60至70支股票。
2005 年年中，格林博格从哥伦比亚大学(Columbia University)毕业并获得工程学学位不久后便开始设计这个程序。加入这项工作的还包括他高中时的伙伴、拥有金融和数学背景的弗莱斯 (Alexander Fleiss)、拥有作曲硕士学位的斯特奇斯(Jonathan Sturges)以及帮助设计这个人工智能程序的数学家牛顿(Jeremy Newton)。
2007年1月，以200万美元为资本，这 个程序开始选股。那年春天，它开始进入防守状态，转持公共事业公司等股票。据知情人士称，Rebellion在2007年升值17%，而当年道琼斯工业平 均指数(Dow Jones Industrial Average)的涨幅仅为6.4%。
2008年大多数时间，它一直处于守势，持有 黄金、石油和公共事业公司的股票。尽管如此，它也像大多数投资者一样损失惨重，缩水26%，但仍优于道指34%的跌幅。
2009年初， “星”开始买入银行和保险公司等跌幅极大的股票。这些公司将从经济复苏中受益。弗莱斯说，这个人工智能程序刚刚买入了不少价值股。2009年，这个基金升 值41%，是道指19%涨幅的两倍多。
这家公司目前的投资组合大体上以防守为主。据知情人士透露，黄金股是其最大的头寸之一。
起 初，这种防守的举动让愈发看好股市的弗莱斯担心。但迄今为止，事实证明这是明智之举。他说，我已经学会不去质疑人工智能。
Wall Street is notorious for not learning from its mistakes. Maybe machines can do better.
That is the hope of an increasing number of investors who are turning to the science of artificial intelligence to make investment decisions.
With artificial intelligence, programmers don't just set up computers to make decisions in response to certain inputs. They enable the systems to learn from decisions, and adapt. Most investors trying the approach are using 'machine learning,' a branch of artificial intelligence in which a computer program analyzes huge chunks of data and makes predictions about the future. It is used by tech companies such as Google Inc. to match Web searches with results and NetFlix Inc. to predict which movies users are likely to rent.
One upstart in the AI race on Wall Street is Rebellion Research, a tiny New York hedge fund with about $7 million in capital that has been using a machine-learning program it developed to invest in stocks. Run by a small team of twentysomething math and computer whizzes, Rebellion has a solid track record, topping the Standard & Poor's 500-stock index by an average of 10% a year, after fees, since its 2007 launch through June, according to people familiar with the fund. Like many hedge funds, its goal is to beat the broader market year after year.
'It's pretty clear that human beings aren't improving,' said Spencer Greenberg, 27 years old and the brains behind Rebellion's AI system. 'But computers and algorithms are only getting faster and more robust.'
Some sophisticated hedge funds such as Renaissance Technologies LLC, based in East Setauket, N.Y., are said to have deployed AI to invest. But for years, these firms were the exception. Some firms that have dabbled in AI are skeptical it is anywhere close to working.
Rebellion is part of a new wave of firms using machine learning to trade. Cerebellum Capital, a San Francisco hedge fund with $10 million in assets, started using machine learning to invest in 2009. A number of high-frequency trading firms, such as RGM Advisors LLC in Austin, Texas, and Getco LLC in Chicago, are using machine learning to help their computer systems trade in and out of stocks efficiently, according to people familiar with the firms.
The programs are effective, advocates say, because they can crunch huge amounts of data in short periods, 'learn' what works, and adjust their strategies on the fly. In contrast, the typical quantitative approach may employ a single strategy or even a combination of strategies at once, but may not move between them or modify them based on what the program determines works best.
Past success doesn't mean Rebellion will continue to beat the market. As with many quant strategies, its system could stop working if market fundamentals change in ways that trip up its computer program, known as 'Star.'
What makes Star intelligent, says Mr. Greenberg, is its ability to adjust its strategy based on shifting dynamics in the market and broader economy. The program isn't wed to any single investing approach. Under certain conditions, the fund will buy cheap stocks, in others it will favor stocks with swiftly rising prices -- or both at the same time.
Unlike the high-frequency funds that use artificial intelligence to aid rapid trading, Rebellion tends to hold stocks for long periods -- on average four months but in some instances more than two years. It also doesn't short stocks or use leverage, or borrowed money, which can amplify returns but also boost risks.
The program monitors about 30 factors that can affect a stock's performance, such as price-to-earnings ratios or interest rates.
The program regularly crunches more than a decade of historical market data and the latest market action to size up whether to buy or sell a stock. When certain strategies stop working, the program automatically incorporates that information, 'learns,' and adjusts the portfolio.
For instance, it may detect data indicating stocks with low price-to-earning ratios are likely to rise and load up on those stocks. Then, if the program later finds that the strategy is likely to lose steam, based on shifts in the factors it tracks, it will dump those stocks and buy stocks it deems more favorable.
Every morning, Star recommends a list of stocks to buy or sell -- often it offers no changes at all. A human trader implements the moves. The firm says it never overrules the computer program, which is largely the same system they started with in 2007, with a few nips and tucks. Rebellion typically holds about 60 to 70 stocks at any time.
Mr. Greenberg started designing Star in mid-2005, soon after he graduated from Columbia University with an engineering degree. He was joined by Alexander Fleiss, a high-school friend with a background in finance and math, as well as Jonathan Sturges, who has a master's degree in music composition, and Jeremy Newton, a mathematician who helped design the AI program.
In January 2007, with $2 million in capital, the program started picking stocks. That spring, it started moving into defensive positions such as utilities. Rebellion gained 17% in 2007, compared with the 6.4% gain by the Dow Jones Industrial Average, according to people familiar with the fund.
It stayed defensive throughout most of 2008, holding gold, oil and utility stocks. Still, it lost money like most investors, sliding 26% but topping the 34% decline by the Dow industrials.
In early 2009, Star started to buy beaten-down stocks such as banks and insurers, which would benefit from a recovery. 'He just loaded up on value stocks,' said Mr. Fleiss, referring to the AI program. The fund gained 41% in 2009, more than doubling the Dow's 19% gain.
The firm's current portfolio is largely defensive. One of its biggest positions is in gold stocks, according to people familiar with the fund.
The defensive move at first worried Mr. Fleiss, who had grown bullish. But it has proven a smart move so far. 'I've learned not to question the AI,' he said.