[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Optimized sys. diff results diff stocks.



PureBytes Links

Trading Reference Links

A trading system works only by capitalizing on some characteristic of 
a market that is repeatable - "if XX occurs, then the price is highly 
likely to rise". Without that, trading is just gambling.

While it may be possible to find some characteristic that is so 
universal that it works on all markets, I have never found it. In 
fact, I suspect that if it ever existed, it would have been 
recognized by lots of others and then "arbitraged away". If such a 
characteristic exists and continues to exist, eventually everybody 
would learn about it and everyone would make money trading.

But individual markets clearly have "personalities" that create 
repeatable events on which you can build a profitable trading system 
for that market. Even the "efficient market hypothesis" economists 
now admit this.

So most of the work of developing a trading system is in finding some 
characteristic of a market that is consistent enough to be tradable. 
The coding is usually the easy part. Once you find one of these, you 
usually do not want to tell others about it because if a lot of 
people know about it and trade it, it will disappear.

The problem is that you never know how long the characteristic will 
persist. So at least such a system should tell you when it is no 
longer working well. In the example, "if XX occurs, then the price is 
highly likely to rise", if this stops happening we would like to know 
it and stop trading during that period. Sometimes you can test to 
look for those periods and just avoid them. For example, it is pretty 
well known that the S&P futures trade differently during the lunch 
hour that at other times so maybe you would like to avoid trading 
then.

You can usually make a system more robust by using the correct 
"dimensions". For example, many systems hold up better over time by 
using the percent change in price rather than the raw price as the 
input variable.

Next best is one that adaptively adjusts the parameters over time to 
attempt to track the shifts in the market. You can do this by 
"measuring" what the market has been doing and adapt the parameters 
accordingly. Using your example, you might measure what the 
retracements have been running recently and use that value in the 
system. Then if the retracements shift from 20% to 15%, your system 
might still keep working.

Lastly, you can just re-optimize the system idea for each market 
fairly frequently. This can lead to curve fitting if you are not 
careful. It is a big help to make the input parameters "orthogonal" 
as much as possible. For example with a system that uses two moving 
averages, your might normally use the two lengths as inputs:

    Input: Length1(15), Length2(45);

But if you optimize on the two lengths, you often will get very 
strange interactions between the values.

But if you keep one length some multiple of the other:

    Input: Length1(15), Ratio(3);
    Length2 = Ratio * Length1;

You can now optimize on Length1 to shift both lengths proportionately 
then optimize on Ratio to vary the relationship between the two 
Lengths. This often results in greatly reducing the interaction 
between the two input parameters so that when you re-optimize, you 
might only have to vary one of the parameters.

Bob Fulks


At 4:22 PM -0600 11/20/99, Jimmy Snowden wrote:

>It has been my belief that a system should work from one stock to
>another without changing the inputs. I would like to know if this is
>a rule that might be incorrect sometimes. Does anyone use a system
>that they optimize for each stock or other symbol and find it
>acceptable to trade over time.
>
>My thinking is that if a system is optimized for a single stock and
>works well as new data comes in RT then eventually when the stock's
>trading pattern changes just optimized again on more current data. If
>this is so and the system works for several months with no changes is
>it viable or am I kidding myself on this. Yes the system works on a
>broad range of inputs but there is significant difference from one
>stock to another in the best range of inputs.
>
>Here is where I am and my logic. I built a system based on Up trend
>is true if retracements are less than up legs. If so and the stock
>reverses to the upside it goes long. If the stock makes a peak it
>exits. Yes you miss up moves if the stock doesn't have a retracement.
>I watch lots of stocks to make up for this.
>
>The short sell system is the same thing turned around.
>
>I have optimized the system for each stock and for each time frame I
>chart the stock on. My thinking is each stock has different cycles
>and I want the most profit from each swing. The system is a screen
>for strong stocks and stocks that lend themselves to the system. What
>I am after is to only trade the stocks that optimize out with the
>following criteria.
>
>1. Either high percentage of successful trades or high win loss ratio
>2. Losses on bad entries are small and less than the average winning
>trade 3. The largest win is not nearly all of the net profit
>
>I am using 60 minute charts to trade and I watch daily and 10 minute
>charts to calm myself on the daily and get my blood pressure up on
>the 10 minute charts. I want the biggest bang for my extremely
>margined buck. I buy only one or two stocks at a time. If I am
>stopped out I will buy as much as my account will allow on the next
>buy signal. Stops kick in so fast that losses are still kept small on
>the larger position.
>
>I optimized all stocks about a month ago and signals are still right
>on. Yes I know the market has been up for a month but the short side
>is working also. The system is really two separate systems one for
>long and another for short.
>
>I am attaching two system reports on the last two stocks I traded. If
>they don't optimize at least this well I don't trade them.
>
>I am very interested in any input before I either get rich or go broke.