"Me, I spent years working on systems only to find that because I pull the trigger (on a whim, I suppose) it makes my system discretionary..."
Yesterday was painful. I got a buy signal after the market went straight up after 10 am, then pulled back a bit and just clipped my sell stop. Had it been last week, I wouldn't have bought after a straight run up.
I spent the day trying different strategies like keying off a moving average instead of price, or waiting for %R to drop below 25% for a buy signal. These would have worked well for yesterday, but when I ran the system over the last five years of data, the results are no better than my original system.
I noticed that some people use one set of parameters when they go long, and a different set when they go short. When I did that, I got about a 20% improvement in profits. Since my system has only one main parameter, plus another for the stop level, I don't think that doubling the number of parameters will do too much damage.
Regarding discretionary trading, I think I can use a system to calc the buy/sell levels, and then go back over the data and see where the system tends to fail. And then under similar conditions again, I can choose to trade the system on a given day, not trade it, or given that it is a breakout system, perhaps take the opposite side if there appears to be significant resistance above the breakout level (during a prolonged drawdown, taking the opposite approach must have some merit I guess).
I am somewhat concerned about historical testing. Clearly the market changes from time to time (esp. volatility) and parameters sets will change as well. I'm wondering how to best choose a set of parameters. Perhaps, starting from 1982, one could optimize parameters on 4 consecutive years, and then test the parameters on the following year, which was not used in the fitting procedure. For example, I could determine a best set of parameters for 1982-1985, and then test them on 1986 data. Next determine another set from 1983-1986 and test them on 1987 data. At least there would be a rigorous procedure for calculating the parameters and I would know how well this procedure worked. |
