Does this mean you can do the same? It depends on you. You may do even better if you spend more time and use better inputs to the net than we did. We aren't professional traders, and don't have all day to build systems and trade.

Will our net will hold up in the future? Maybe, maybe not. The market changes, and models often need to change with it. We aren't claiming that neural nets are magic, just that they are a great help when you haven't been successful with traditional systems.

Does this mean that 30% per year is the best that can be done? Absolutely not. In previous years when the market was highly trending or very cyclic, we had better models, and so did some of our users. What we are showing you here is just an example which worked well through 2005 when the market was generally flat and even many professionals had a hard time making money.

What it does show is that it is possible to build neural network models that do better than the overall market, and probably better than you can do with your own biological neural nets.

PS - although we used a neural net in this example, NeuroShell is just as capable of building good models without nets by applying our genetic algorithm optimizer to traditional indicator and rule based systems.

Example 2

Here's another neural model we built for and use in our own trading. The net signal is from one of the net types you can find in our Neural Indicators add-on. When the signal gets close to 1 it means it is time to buy and a signal near -1 means it is time to sell. This model was trained January 20, 2006 on data going back to late October 2001, so the signals you see on the chart starting January 23 are out-of-sample signals. We like to use SPY because from the signals generated we can trade some ETFs, some Emini futures, and even options on ETFs or Eminis.

Example 3

Here is a chart made by one of our users. The trading strategy is a "panel of experts" which evaluates four different neural net predictions to produce buy and sell signals. The nets were trained and optimized on data prior to 2007, so the 2007 signals are out-of-sample signals. Therefore the equity curve restarts in 2007.

Example 4

NeuroShell can help you build market-neutral pairs trading models, even when the raw prices do not cross. In these models, when one stock is short, the other is long with roughly the same dollar amount invested. This example shows a backtest of trades for BMY when paired with WYE. The equity curve is combined net profit for both stocks. Notice how well the model does in volatile bear markets.

Example 5

This is a model we built in September 2008, and then watched it untouched for 7 months during the 2008-2009 crash. First we created just two sophisticated indicators from NeuroShell Trader. We used ChaosHunter to build the model from those two based on SPY alone, but executed the model in NeuroShell ready for trading. (Note that NeuroShell and ChaosHunter are two distinct products that we sell separately, but work synergistically together). The model as executed in NeuroShell contains 19 Rydex funds (ETFs) including the one shown and assumes $10,000 invested in each one just before September 22. Some of the 19 did well, and some did not, but the combined net profit during the next 7 months was $105,298, or 55.42%. On an annual basis, that would be 95% return during the worst market period since 1929. The same model also did well on other ETF fund families.

Example 6

In this model we built a neural net with one hour bars, trading 100 shares, holding out data after March 12, 2009 (after March 12, you can see the equity curve restart as the neural net starts trading on data it hasn't "seen" before). As you can see the model did very well in the "out of sample" period after March 12. The inputs to the neural net were some that we use in several of our examples. This model was originally built to trade the overnight gap, but we made a mistake, and the neural net, with a mind of its own, made a good model in spite of our mistake.