Optimatization (whens enough)

[mccmatthew]

In my current quest to develop a futures trading system in TS, I have been trying to limit the amount of optimatization on the indicators, profit targets , and stops.

It is easy to get carried away with it, but in what moderation and guidelines are there to keep the damaging effects of over optimatization at bay.



thank you,

 

[ChrisM]

General rule as You, know about optimization is: as little as possible, but this does not help in many strategies.
First You must understand optimized inputs.
Example: if You optimize position system based on indicators, very short periods oscilators naturally will produce significant amount of whipsaws etc., but strategy may be still OK.
Reliable system should be profitable with average (random) inputs. If any change in input`s value lowers significantly the system`s profit, do not trade that.
Also, system should stay profitable through series of inputs, the best if not changing profitability too much.
Also - take look at Adjusted Net Profit in TS analysis. If this result is far from Net Profit (i.e. negative) , that usually means that the system is "curve fitting".

 

[o_2_b_sean]

I feel that optimization can be useful to discover certain things about the stock or market that one is trading. For example, say you are creating a point and figure type system with "block sizes" or some other type of static or percentage based block size system, in such cases optimization can help you "discover" the ideal block size. This could be useful to filter out minor whipsaws, for instance.

Another example. Say you wanted for test for some good linear regression or moving average lengths that might work with your system, this can easily be done with optimization.

I think optimization is less harmful when done over longer time frames. This way the results are much more likely to be representative of actual market behavior that might actually repeat in the future. Optimization on a ten day time frame, in my experience, is almost always useless when applied to a year's worth of data, etc

good luck

 

[mccmatthew]

Thanks for your inputs they have helped greatly, the problem I am running into is when running the optimatization on stochastic for the system (es,5min) the results differ greatly and almost random from one setting to the next.

Example: A 23,7,7 will produce a far different result from a 23,7,6 there is no true pattern when looking at the results.


Is this normal when testing short time frame indicators, and can these results be trusted.

 

[o_2_b_sean]

Originally posted by mccmatthew
Is this normal when testing short time frame indicators, and can these results be trusted.

Yes, probably normal. Personally, I would never use a system based on an indicator like stochastics. You will probably have better luck writing you own indicators rather than a canned one. Start opening up all the functions and reading the code. Also, as someone pointed out before, the more you focus on the direct price data (as opposed to indicator/oscillators) the more you stand a chance of finding a system that works. My advice to you is focus on the price data and think along these lines: "how can I filter out what matters and discard the rest". That might mean looking at multiple time frames for the same symbol. Also thinking in ratios helps because as the markets get more or less volatile a fixed size filter will start to fall apart.

but what do I know...

happy hunting

 

[lundy]

i wonder if anyone has experimented with the opposite of optimization? to tell u when NOT to trade. or, if anyone has experimented with cross optimization in different time frames, different models, etc.

 

[vikana]

another way to look at it, is to determine how many parameters you realistically can have. If you only have one time series (e.g. price), you really shouldn't optimize more than one variable. Two is probably ok, but you have to be careful.

To restate: your system probably should not have more than one variable that is candidate for optimization. Anything beyond that and you risk over fitting. I'm sure you can get scientific with the dimension of the time series ... the above reflects my own experience

 

[rtstrading]

"It is easy to get carried away with it, but in what moderation and guidelines are there to keep the damaging effects of over optimatization at bay."

Good article in the May issue of Futures magazine by Murray Ruggiero about system testing. It would be worth your while to run down a copy and read it!

In general, the use of mulitple canned indicators w/optimization will not work into the future. You are simply curve fitting to the data set at hand. "The most successful system developers base systems on patterns(price or chart) or fundamental relationships."

"With a computer and trading software, it's quick and easy to test technical indicators, hundreds of which come pre-built in most software. Any edge that can be gained by trading classic indicators out of the box is likely negated by the tens of thousands of traders who have come before you."

I have found the use of price patterns, as suggested by Michael Harris( www.tradingpatterns.com ), as a superior method for having future success in designing trading systems.

Good luck in your search!

 

[ChrisM]

Originally posted by mccmatthew


Example: A 23,7,7 will produce a far different result from a 23,7,6 there is no true pattern when looking at the results.

This is exactly warning signal. In most cases strategy will lose money.

 

[mccmatthew]

Thank you for everyones input, in further thinking and research I have been convinced that the path of a robust profitable trading system might not lie in the "canned indicators" that we are all showed and read about in almost everything relating to trading.


Which brought my attention to charting patterns(no indicators), these systems seem to be more robust and harder to back into trap doors like other indicator based systems. Not to dilute the validity of those systems it just takes a great understanding, experience, and tasteful manner in which to implement them.

So once again thank you everyone,


Matt