From time to time people ask me what sort of model I use to trade. I have mentioned my methods on this site in the past, and since I have changed techniques over the months and years, it is time to reveal my methods again. Just remember the following things:
1. I am certainly not the best trader around, and many of our users do much better than I. But since I understand most of our algorithms quite well, many people think I must be pretty good. I make no guarantees about your results with the methods I currently use.
2. No trading model works for more than a few months before it has to be retrained, reoptimized, or refurbished. The market is efficient and once good patterns become obvious, the market has to adapt to make those patterns obsolete.
3. Although I don’t mind telling you in great detail what I do, I don’t want to give out my EXACT charts. You may have to become somewhat proficient to use my methods outlined below.
4. You’ll need the NeuroShell DayTrader, the Neural Indicators addon, and maybe the Turning Points addon.
5. There are a zillion ways to build good models, and the one outlined below is not supposed to be optimum, just one I’m comfortable with and works well for me.
6. The models I make are not always right, but part of trading is knowing when to hold and when to fold. This is why experienced traders do much better with our software than math or neural net wizards, and this is the art you will have to learn even if you have a great model.
When I trade S&P eminis with the NeuroShell DayTrader I use the Neural Indicators addon and usually use the recurrent indicators (Usually Recur3, 4, or 5).
Those indicators have to be fed some input parameters, and I usually use one or more of the following (in order of my frequency of use):
Linear Time Regression: Coefficient of Regression (slope)
The N bar Lag of the above
PeakProp minus ValleyProb
PeakProbB minus ValleyProbB
The first indicator is found in the Regression category. The Lag is in the Basic category. Often I’ll use one slope unlagged and 1 or 2 lagged. The next two are spreads of the indicators found in the Turning Points addon.
I will use a reversal strategy with a bar sizes that I have varied from 1 minute to 30 minutes. I put one Neural Indicator in as a long rule, and another in as a short rule. Usually I will specify a time rule for entry as well, like Time=10am (use any time you feel comfortable with). Then instead of using a Neural Indicator in an exit rule I will use only a time rule, like Time=11am. I’ve been predicting the overnight gap alot until recently.
Then I use the optimizer to find all parameters except time, although I sometimes will fix the scale parameter of the Neural Indicators to exactly 100 bars and not let that optimize. I let the periods of the regression slopes and their lags vary from maybe 3 bars to a max of 20. For turning points I use window sizes from 50 to 300-500, and a use neighborhoods of from 3 to 10 usually.
I use intraday bars, usually optimizing on about 3 months of intraday data using continuous contracts from eSignal (ES #F=2), and then paper trading on the next week or two. I try to find a period of optimization for which the market showed no clear trends, volatile with lots of moves both ways. I rarely use the default amount of optimization, opting instead to set the amount of optimization time from 3 hours to 6 hours (I’m not too worried about over optimizing because 3 months of 5 minute bars is around 4800 bars, and Neural Indicators are slow anyway).
The regression slope indicator is very effective in trading strategy rules as well as neural nets (see the Dec 2000 Stocks and Commodities magazine tip called “Pattern Recognition” on this site).