Wouldn't it be great if you could sit back and let your computer pick your trades for you? Some people are doing just that, and they claim they're making money.

It's done with a kind of artificial intelligence software called neural networks. On certain levels the software can think and learn, even about things as complicated as the stock market. The only catch: It needs to be programmed correctly before it'll work.

Neural-net heads -- mostly math and engineering types -- swear their software can beat the market consistently. Take Scott McCormick, an operations director for a Detroit-area computer-based design firm who has degrees in math and aerospace engineering, and a minor in theoretical physics to boot. McCormick says he has used neural networks to guide his investing for more than nine years now, producing average annual returns of 50% or more. (Please see our

related story on how to get started using neural nets.)

Neural nets' usefulness lies in their ability to recognize patterns, McCormick says. "That's really the object. Is there a pattern that appears to be repeatable and is there an opportunity there?"

Those patterns aren't limited to trading. Banks use neural nets to analyze credit reports. Security firms use them to recognize facial patterns. While working for

NASA

, McCormick designed a neural network that could assist satellites in navigating as they approached the limits of our solar system.

When used by traders, neural nets work best when they focus on an individual stock rather than the market as a whole, says McCormick. "A stock is owned generally by the same kinds of people or by the same funds." Because of their similarities, both individuals and fund managers would tend to buy or sell a stock at the same time, he says. That's what creates the pattern.

I Think, I Am. I Think, I Am

How exactly do neural nets work? Picture a series of interconnected, nervelike cells. Neural-net users call these cells neurons, though they are really just blocks of programming script. Each neuron functions as a sort of minicalculator. The calculators break apart a problem and then set to work on it. When picking stocks, for example, the problem might be determining the relationship between the stock's volume, its price and its moving average (the average closing price plotted for a specific period, for example 200 days).

A neural net might crunch through a year's worth of historical data for a stock. It will then come up with a series of rules or predictions about when the stock might move in price. As a next step, the neurons test their predictions using current market data. And all the while they have the capacity to tweak the variables to improve their accuracy. They might, say, adjust the moving average period from 200 to 196 days. Then they'll see if that new time frame is a better predictor of the stock's price. That's what the software designers mean when they say that neural nets are able to learn.

Neural-net heads will tell you that comparing price, volume and moving average is pretty simple. Off-the-shelf neural programs let you do far more complex calculations. "You can take the moving average of the linear regression of a stochastic if you want," says McCormick. Linear regression and stochastics are both statistical analysis techniques commonly used by technical analysts.

The Nightly Grind

So how does McCormick use the data? Each evening, McCormick downloads end-of-day price data on about 100 stocks and enters this information into his neural net. The net then spits out maybe a half dozen stocks that it predicts will make a move the following day. Of course, the results depend on what indicators McCormick programs into his neural net. Those indicators may work for a while, then suddenly they'll cease to accurately predict a stock's movement. McCormick calls this phenomenon indicator drift, and he's not certain what causes it. Nevertheless, because of indicator drift, McCormick must routinely tweak the set of indicators he programs into his net.

When he's satisfied with the neural net's recommendations, McCormick uses them to fill out limit orders, which he enters through his online broker. Then he waits for the market to come to him. "I set limit entries. I'll put those orders in. If they fail, they fail."

Sounds good, but could everyone do it? Having designed neural nets for NASA, McCormick is, after all, a rocket scientist. McCormick says it helps if you didn't sleep through high school and college math classes. Beyond that, he says, "Neural nets won't make a bad trader good, they'll make a good trader better."

Mark Ingebretsen is editor-at-large with

Online Investor magazine. He has written for a wide variety of business and financial publications. Currently he holds no positions in the stocks of companies mentioned in this column. While Ingebretsen cannot provide investment advice or recommendations, he welcomes your feedback and invites you to send it to

mingebretsen@thestreet.com.