Definitions of These Two Genetic Options Settings?

What are these two Genetic Options settings: (hopefully in layman’s terms)

1. Max Tree Depth

2. Decimation coefficient

 

 

+1, i would like to know every single settings inside the genetic generation also please,

 

PLEASE MORE IN-DEPTH THEN THE SQ GUIDE..

 

thanks

Ah, the Decimation Coefficient is very interesting then…ThanX for the help.

 

I don’t understand the description of Tree Depth, though.

So if i’ll select “5” as the Max Tree Depth,

the strategies will have Up To “5” rules for entry and Up To “5” rules for exit did i understood it correctly?,

if i selected “5” it does not mean if it will find a strategy even with “2” rules it should be added to my databank right? not only with “5” yes?

 

Did not get the benefit of having Decimation coefficient other then “1”

 

 

thanks notch!

Decimation coefficient generates more initial strategies (in generation 1), and then chooses the 100 best strategies out of those (or whatever you set for “Population Size”) to continue onward in subsequent generations.

 

This seems like an extremely important setting to me, as you can start with a much higher-quality base of randomly generated strategies. I currently have it set at “4”.

 

I have also been using the “initial population conditions” to filter the initial generation. I have been doing so very lightly (I think?) in order not to unknowingly reduce some strategies that have elements that may be extremely profitable but show up much more frequently in strategies with lower standards.

 

For example, here are settings I am using for now (but maybe they are still too tight?):

 

Number of trades < 50 [five-year IS period]

Profit Factor < 0.6

Stability < 0.2

% Stagnation > 50%

 

Some other factors that I was considering:

 

Net Profit < -$7,000 [using NP would weed out strategies that pass the PF test but have a larger number of trades but still don't have much hope of being transformed into something profitable]

Symmetry < 10% [Not sure about this, but should produce more robust strategies?]

Payout Ratio < 0.4 [would bias results against systems with lower Win%]

 

I briefly tried applying these settings to OOS results, but I don’t think it produced much better initial strategies.

 

Even with these lax settings, it takes a good amount of time to derive an initial population of 400 strategies, but I think the time spent *may* be worth it.

Just use more loose criteria,

those are what i use always:

these are my settings over at the GenGen 😉

lets see if its any good

 

Those criteria I listed above I use in the “Genetic Options” tab, in the “Initial population conditions” so that it is a bit picky for the 1st generation population.

 

I have similar conditions to yours in the “Ranking Options”.

ThanX for your reply notch; I have subsequently adopted an approach similar to yours, as my approach was not generating many strategies and those strategies were not visibly better.

 

I still use the “decimation coefficient”, but the only condition I have for including generated strategies is “Number of trades > 40” (over IS of five years).

 

I then used Ranking Options as you do to guide the model generations to my overall criteria. This makes 100% sense.

 

I differ from you by using Number of Trades. I believe that the Weighted Fitness in Ranking Options will account for better Net Profit, but that initial strategies that have below a certain minimum Number of Trades are not good candidates to produce an eventual good strategy. They will likely be weeded out through the Genetic process, and thus you are left with a smaller pool of viable initial strategies that can eventually be used to good effect. And since Number of Trades is not used in my weighted fitness, whereas other criteria are, I think it’s the only criteria that should be used in my initial generation phase.

Do you use those criteria inside the Ranking tab or inside the Genetic tab?, and is there any difference between where to set the criteria here or there?

@Karish:

 

I currently use the following criteria inside the Genetic Options:

#/Trades < 40

 

Decimation coefficient = 5 [However, considering Notch’s post above, I am again reducing this to 1]

 

In Ranking Options, I use the following simple weighted Fitness parameters:

Net Profit: 1 weight

Profit Factor: 3 weight

R/DD ratio: 10 weight

@notch: ThanX for your warnings…

 

As noted above, I’m only using a low “#/Trades” criteria [40 for a five-year period]

I’m guessing (but don’t have experience of course) that strategies with ultra-low #/trades don’t yield any significant contributions to genetic development.

 

My weighted Fitness criteria are very simple (nearly identical to using only R/DD).

 

I briefly used Stability = 3 and % Stagnation = 1, but dropped them.

 

I honestly have questions about the importance of each of those parameters. (I know those more knowledgeable than me will argue differently! I’m just not convinced.) As long as R/DD is good, these parameters shouldn’t be of too much concern. (?)

i think so too,i tend to use the RDD as my main thing to follow, the higher the better,

Here are my thoughts about Stagnation and Stability…please tell me how far off I am.

 

Obviously, the market is not totally random and therefor does not produce a fully random order of trades (as does our Robustness Test of Randomize Trade Order). BUT if it did produce a totally random order of trades, then we could completely forgo the values of Stagnation and Stability, because Drawdown from our robustness testing would completely dominate in importance. If trade order was random, Stagnation and Stability would be quite random, depending heavily on Win % and R/DD.

 

Because the market has some sort of structure to it, strategies do not produce totally random trades. But really, how much structure does the market provide to force a particular strategy into certain pre-destined Stagnation and Stability values? If we re-run the same historical period with some minor deviations, it will likely cause a major change in both values.

 

However, R/DD is not as vulnerable to these changes. We have tests to determine the standard deviation of R/DD and can choose strategies that have the largest probability of having high R/DD.

 

I don’t feel like the above is a complete thought, but it’s a bit late… 🙂

No, because the model with lower Stagnation and higher Stability is more likely to show those same traits in the future, but my question is, “how much more likely?”.

 

Perhaps there are other measures that can provide a better long-term prediction (or likelihood of values) of Stagnation and Stability.

 

I’m guessing Stagnation and Stability are heavily influenced by things like Win% and Average Win/Loss (?). These later numbers would tend to vary much less given sufficient sample sizes. I haven’t examined most of the Robustness testing…maybe there is already a test for Stagnation and Stability?

 

I guess my overall point is that any one historical run in the market can produce significant variance from the norm in Stagnation and/or Stability (as well as other traits, such as R/DD — for which we have a variance test).

Yeah you the truth is that when i use pop size 100 & # of gen 25 with coeff 1, i get more results rather then what i was using before

100+ pops size with # of gen 999 and coeff 5~10,

 

thats odd. 🙂