Trading with neural nets is very much an art, not a science. Therefore there are many ways to build profitable systems. I am going to tell you about the successful systems I’ve built with daily bars for those of you who feel you need a jump start. There are many other ways, but if you want a cookbook of one man’s methods, here it is:
1. This first item is the most important of this entire tip. If you get this right, the rest will be easy. Choose several (3 to 10) volatile stocks or other issues that move up and down regularly. Don’t pick your favorite stock or issue; don’t start with what you WANT to predict; start with what is EASY to predict. You want about 4 years of history of SIMILAR types of moves. If you choose the right issue, many indicators will predict it well. If you pick the wrong issue, it will be hard to predict it with anything. Stay away from “bull market” or “fad” stocks. Don’t try to pick issues driven primarily by fear, greed, or strong fundamental factors for which you have no indicators.
2. We’re going to be building a trading strategy that looks for three indications that the stock is moving. This is similar to consulting three “experts.” These three “experts” can be your favorite indicators, breakout indicators, moving average crossovers, or neural nets. I use like to use three neural nets, but I’ve used a neural net and two “breakout” indicators. Here’s how I build the first neural net “expert”:
a. Predict the 1 day momentum or percent change in open 2 days in advance. Look at another tip on this web page for why this is a good output.
b. Pick only about 5 indicators from the Price Momentum category. Include a CCI or RSI. I have even built good nets using 5 RSI indicators with different time periods. When your 5 indicators are graphed, they should all look very different. Don’t use similar ones.
c. Use a training set size of 3-4 years, with a minimum of 2 years. If I can do well over the last 6 months of out of sample simulated trading, I am ready to go forward. I am not looking for something so perfect that it would have worked over the past 10 years. I think market conditions change too rapidly for that.
d. Train your net on profit and look for about 30% or more return on the evaluation period.
3. Now, decide on two other “experts” (indicators or nets). If you build two more nets like I do, build them exactly as the first net, but use inputs from another category of indicators. I make my second net from the volume category. For my third net, I use regression slopes (the slope of the line through the last 5 opens. Then 5 lags of that using different lag periods of no more than 20 days).
4. Now build a trading strategy that buys long when two of three rules are true. Your three rules correspond to the three experts (indicators or nets) you decided to use. For nets, your rules would be like: buy if the out of sample signal is greater than 0 (or 1 or whatever threshold you decide).
5. Your sell strategy would also have three rules.
6. If you are selling short, you may want to insist that all three rules are pointing to a drop, not just two.
7. Backtest to make sure your strategy would have been profitable (>30%) at least over the last 6 months.
9. Now add the other stocks or issues to your chart and see if they look profitable too. Delete the ones that are not. If you don’t have enough machine to handle 3 to 10 stocks and all the associated nets, then save a chart for each issue (do this by adding one issue at a time, and then removing the last one before saving).
11. Start trading, dividing your capital amongst the issues that looked profitable on the backtest and for which you saved charts. Start with just 100 shares of each issue; that way you can feel comfortable starting out because you aren’t risking your life’s savings. You are going to lose a few, but you should get about 60-70% of your trades right. Increase your capital invested as you become more comfortable with the system. Keep trading until the market conditions change for an issue and the signals are not very good any longer.
Happy Trading
Steve Ward