Our Gee Whiz examples show the kind of results that are possible and that you could see with our neural nets and genetic algorithm optimizer using just simple inputs. In every case the screen shots you see are backtesting over an "out-of-sample" time period, which is the period in time just after the neural net was trained or just after optimization took place. In other words, the system did not "see" this data when the model was being built; the backtest simulates trading that would have occurred had you build the model earlier and then traded it during the dates shown on the chart. Where present, equity curves are returns for one share, not re-invested, and not including initial capital.

Disclaimers: The Gee Whiz models might not provide good results very much longer than shown, so there is no guarantee that a model that performed well "out-of-sample" will not lose money later. It happens. Furthermore, the same inputs will not necessarily work as well with other ticker symbols and in different time periods. However, the inputs and models are not specific to one type of instrument, meaning that a stock model is just as likely to work for a futures or forex contract as another stock. Not all of the models we build work this well, and therefore your results may vary as well.

GM

You could have made a lot of money on GM stock, single stock futures, or options on this neural network model, because it "figured out" that GM would be declining in 2005. The network was trained with three standard indicators (MACD Signal, RSI, and Williams %R) which were not even optimized by our genetic algorithm optimizer. It was set to predict the change in open 5 days in advance, and produced a 68.5% return on account during the out-of-sample backtest period.

INTC

In this prediction of INTC, we used the following standard indicators as inputs: Stochastic %K, Stochastic Slow %D, and Williams %R. The net was trained to predict the percent change in open 5 bars ahead. The return produced was 100.3% annualized on unseen out-of-sample data.

WMT

Here are the out-of-sample signals given by a neural net trained to predict WMT prices 3 days in advance. Inputs were slopes of regression lines through previous data and were not optimized. The signals produced a 24.8% return on account during the out-of-sample backtest period.

JNJ

This neural net was trained to predict JNJ prices (percent change in open) 3 days in advance. Inputs were slopes of regression lines through previous data, the same ones we used to predict WMT. These inputs were also not optimized, and produced a 28.4% annualized return out-of-sample.

CAT

In this example we inserted eight spreads between moving average pairs. We instructed the optimizer to train neural nets with combinations of those eight, but limiting total neural network inputs to fiveof them. The optimizer decided upon four that worked best. The model actually made a higher annualized return on the out-of-sample set (104.3%) than it did during the training period.

DIA, SPY, QQQQ

In the model we loaded three of the most popular ETFs: DIA, QQQQ, and SPY. We built a trading strategy that used one of our add-ons (Adaptive Net Indicators) as a trading rule. Inputs to the Adaptive Net Indicators were just the close and a few previous closes. (We usually recommend against using just previous closes as inputs, so even we were surprised when it worked!) The task was to predict the change in close 4 bars ahead. We then optimized that rule over all three instruments, i.e., we made sure the rule was the same for all three ETFs. Then we tested the optimized rule on the future out-of-sample data. The annualized returns were DIA - 44.5%, QQQQ - 20.5%, and SPY - 12.9%. The chart shows the DIA signals and equity curve.
 

E-Mini

Here is a simple moving average crossover strategy we applied to one S&P emini contract. You can pay thousands for software that just does crossover strategies like this from other vendors. This one took less than 5 minutes to build with no previous experience with this type of strategy. We crossed our built-in adaptive moving average over a traditional moving average, and did no optimization. The strategy made $4500 from the morning of 2/21/06 through the afternoon of 3/13/2006, for an annualized return of about 2043% (assuming a margin of $4000). The ratio of gross profit to gross loss was 2.31 and the ratio of average win to average loss was 1.92. Over that same period the emini lost an annualized 11.9%

Here is a closeup view of the trading in the previous emini chart, which uses 5 minute bars and the NeuroShell DayTrader Professional.

FOREX

Since so many people are interested on FX trading these days, we built a model for the Euro, which seems to be a favorite. This model is very similar to other FX pair models we have built in that they all use other FX pairs as neural net inputs. In this case we applied a linear prediction indicator based on previous performance to US Dollar currency pairs of the following countries: Australia, Canada, UK, Singapore, Japan, Switzerland, and the Euro itself. The model was trained and optimized with daily data from 1/1/2004 through 1/1/2006. The graph shows out of sample trades generating a $6720 profit from 1/2/2006 through 2/10/2006. Assumptions were $100,000 point values, $1000 margin, and $30 spread one way.

One interesting aspect of this model is the contribution the neural net attributed to each pair:

42.41% Swiss Franc
32.26% Euro
07.01% British Pound
06.81% Canadian Dollar
04.38% Singapore Dollar
03.99% Australian Dollar
03.15% Japanese Yen