Tom retired early after selling his investment advisory business in 1994. He had made a good living investing for overseas life insurance companies, but now he is a full time neural network trader. He has gotten to the point where he trades in large blocks, thinking nothing of taking positions over a million dollars. Tom spends about seven months of the year at his home on the warm, tropical island of St. John in the U.S. Virgin Islands. The other half of the year he lives at his other home in New Jersey. In between, he does a lot of traveling in Europe, but he always takes his laptop with the NeuroShell Trader on it. He gets his data from the Internet in Europe. In July, Tom will take up residence full time in the Virgin Islands.

Tom is a hedge trader: for every long position he takes, he has a short position in a different stock at the same time. Typically, he has four open long positions, and four open short positions, according to network predictions. In one of his favorite stocks, he has made 125% per year using an idea from the examples in the NeuroShell Trader. “It is uncanny how good my nets have been,” he says slowly, almost as if he can’t believe it himself.

Tom has purchased all of the Ward Systems Group product line, but he now uses the NeuroShell Trader exclusively, mostly because of its ability to do out-of-sample testing effortlessly. Data preparation and testing that took hours with general purpose network products now take only a minute or two. We were surprised when we were able to get quite a lot of information out of Tom about how he builds his models.

First of all, Tom uses pretty standard indicators for inputs - no big secret inputs here. He trades the oil service companies frequently, and his models for these issues are inter-market models where he takes advantage of oil prices, building relative strength spread indicators. The NeuroShell Trader easily allows him to load different types of data in his nets. His inputs do vary by issue, however. Tom shies away from stocks that are what we like to call “bull market” stocks (stocks that have done nothing but go up in recent memory). Unlike Bud, Tom says he usually uses only one neural net model per stock, but occasionally he does consult a parallel net or his gut feeling to override a network decision.

Tom develops his models starting with relatively short training periods of as little as nine months. He likes to find models that pick winning trades 60-65% of the time in the out-of-sample periods, which he likes to be as long as two years. However, he has traded models that were only 52% correct because when they picked correct trades, they were big winners, and the bad trades were only small losers. The NeuroShell Trader shows Tom multiple out-of-sample periods with its walk forward testing built-in. He uses the Trader's built-in backtesting to simulate trading and calculate the profits.

If a nine month training period doesn’t work well, Tom likes to extend the training to one, two, three, and up to sometimes six years to see if the out-of-sample performance improves.

Update: January 2000

Tom has just contacted us to purchase Adaptive Net Indicators, and we obtained the following update on his trading during 1999.

Tom basically traded two stocks during 1999, Biogen and Exxon. He made a return of about 150% on Biogen, even though he stopped trading it in May before the big run-up. He was getting about 70% of his long trades right, and 65% of his short trades right.

On Exxon, his return was 63%, with 65% of the longs correct, and 65% of the shorts correct. He stopped trading briefly after the merger with Mobil, but the nets continued to do well.

Tom said he appreciates that Ward Systems develops to meet user needs. He still enjoys his life in the Virgin Islands.

Detailed analysis from one of Tom's charts

Update: February 2004

Tom has just told us that since November 2003 he has been using what he calls his "best model ever." He built it with our Turning Points add-on, filtered with some simple conditional rules. In the last 2 weeks over 90% of his trades made money with very little drawdown. He is taking quick entries on 5 minute bars, and his average trade lasts between 45 minutes and an hour.