Statistically testing every indicator with my Genetic Algoritm | Page 7 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

codification of the market / Genetic Algorithms 36 replies

S/R Area Algoritm 17 replies

What Constitutes a Statistically Significant Sample for Back Testing? 4 replies

Has anyone had luck with high frequency using either neural nets or genetic Pr? 3 replies

Genetic Algorithm for Metatrader 1 reply

Attachments: Statistically testing every indicator with my Genetic Algoritm
Exit Attachments

Statistically testing every indicator with my Genetic Algoritm

  • #121
  • Jul 8, 2014 4:50am Jul 8, 2014 4:50am
Quoting Macdon
{quote} Thank you once again Proximus. That's just about all I found on it too. I got the code from nonor's post on this thread http://www.forexfactory.com/showthre...23#post7568523 I'll have to read it all and digest it. Are you going to do any more tests? These 2 RSIs look very good - lots of PIPS and high strike rates of 46% and 57%. Since you did the original tests with 12 years data on the M30 charts, would it be worthwhile to test the indicators you originally tested on the data from 1st Jan 2010 on 1HR charts? Just a...
I am building now a trade manager, since all the EAs that i've built so far were unsuccesful, either i lack the skills to make an EA which would have "high intelligence" to calculate every variable that comes when one is trading, or there are just so many variables in place and a robot cannot look out for them all, whereas any trader with average intelligence is surpassing by far the intelligence of any EA no matter how well its coded.

And i vote for the second, you cannot program something which will have to work in an unpredictable enviroment, this is the biggest paradox ever, since programming is an exact science whereas trading and the markets are not.

So i am now shifting my attention to build an EA which has manual BUY/SELL inputs and only has money-management automated, everything else is discretionary.

So if i combine this research with my trade manager it would give me a very suitable trading enviroment

As for your question yes i will continue it after i finish this work.
"There's a sucker born every minute" - P.T. Barnum
  • #122
  • Jul 8, 2014 5:28am Jul 8, 2014 5:28am
Quoting Proximus
{quote} I am building now a trade manager, since all the EAs that i've built so far were unsuccesful, either i lack the skills to make an EA which would have "high intelligence" to calculate every variable that comes when one is trading, or there are just so many variables in place and a robot cannot look out for them all, whereas any trader with average intelligence is surpassing by far the intelligence of any EA no matter how well its coded. And i vote for the second, you cannot program something which will have to work in an unpredictable enviroment,...
Maybe this might give some coding ideas? http://www.dukascopy.com/fxcomm/fx-a...n=read&id=1099
  • #123
  • Edited 7:51am Jul 8, 2014 7:33am | Edited 7:51am
Quoting Macdon
{quote} Maybe this might give some coding ideas? http://www.dukascopy.com/fxcomm/fx-a...n=read&id=1099
I agree with a few points but not all.

1. Dynamic oscillators
Definitely, thats why i use a GA to spot the best local parameters, since no single parameter will work forever
2. Adaptive Stops and Profit
A little bit of dilemma here, as i mantain the 2:1 RR however, if its adaptive means that if lets say today we used a 50 pip SL and 100 pip TP because the volatility was low, and tomorrow is a GDP release for example we are required to use a 60 pip SL and 120 pip TP.

For which i have mixed feelings because it is good to be adaptive, but also if we change the Win & Loss size, we need good samples for those scenarios, so is we have like 3 types of trade sizes we need to have equally many trades from all of them to have a statistically robust system.

Since for example if we were to use a 50 SL, 100 TP for 500 trades and suddently the volatility becomes immense and we need to use a 500 pip SL 1000 pip TP and we particularly lose that trade, the 500 previous trades with different sizes wont compensate that, so you need 500 trades with 500 SL and 1000 TP aswell to have a good statistical sample for later analysis and performance evaluation.

So adaptive yes, but should not be too far away from your average size, i think the formula to describe this should be this:

REWARD * 2 * DELTA ~ RISK * DELTA

(where DELTA should be less than say 50 pips to not deviate too far away)


3. Risk Adjusted Position Size
Definitely agreed, risk management is important.

4. Dynamic Trade management
Trail stop or scaling out is a good tehnique but only if there is a big distance between open price and TP, otherwise it will just minimize profits
5. Generous Slippage Allowance in Back-testing
I do this, i always test my backtests with average spread *2 to incorporate spread widening and slippage which decreases entry/exit accuracy

I`m curious what do you guys think about this?
"There's a sucker born every minute" - P.T. Barnum
  • #124
  • Edited 6:01pm Jul 8, 2014 5:16pm | Edited 6:01pm
Quoting Proximus
2. Adaptive Stops and Profit A little bit of dilemma here, as i mantain the 2:1 RR however, if its adaptive means that if lets say today we used a 50 pip SL and 100 pip TP because the volatility was low, and tomorrow is a GDP release for example we are required to use a 60 pip SL and 120 pip TP. For which i have mixed feelings because it is good to be adaptive, but also if we change the Win & Loss size, we need good samples for those scenarios, so is we have like 3 types of trade sizes we need to have equally many trades from all of them to have...
You could code lower and upper limits for this. Do not take trades where the SL < 50 or SL > 100, say.

You'd have to back test the system to get the optimum average SL and work from there. Perhaps use SL +/- 2*STDEVs for lower and upper limits? You could also use ATR or the High-Low difference for N days to work out a SL and TP.
  • #125
  • Jul 9, 2014 1:05am Jul 9, 2014 1:05am
Quoting Macdon
{quote} You could code lower and upper limits for this. Do not take trades where the SL < 50 or SL > 100, say. You'd have to back test the system to get the optimum average SL and work from there. Perhaps use SL +/- 2*STDEVs for lower and upper limits? You could also use ATR or the High-Low difference for N days to work out a SL and TP.
Yes i made a custom ATR which tracks that, however there is no optimum SL/TP setting, what is optimum for this week may be irrelevant next week.We need to be more adaptive than that, scaling our position by the ATR is logical, and i am willing to take the risk of uncertainty and diversity over the risk of inflexibility.

And the RR ratio will always be > or = to 2:1
"There's a sucker born every minute" - P.T. Barnum
  • #126
  • Jul 9, 2014 1:48am Jul 9, 2014 1:48am
Quoting Proximus
{quote} Yes i made a custom ATR which tracks that, however there is no optimum SL/TP setting, what is optimum for this week may be irrelevant next week.We need to be more adaptive than that, scaling our position by the ATR is logical, and i am willing to take the risk of uncertainty and diversity over the risk of inflexibility. And the RR ratio will always be > or = to 2:1
I'm also using ATR to determine if RR is acceptable. That way I never use fixed SL/Tp, those need to be dynamic as situation develops.
  • #127
  • Jul 9, 2014 2:49am Jul 9, 2014 2:49am
Quoting Szpons
{quote} I'm also using ATR to determine if RR is acceptable. That way I never use fixed SL/Tp, those need to be dynamic as situation develops.
What ATR setting do you use, since its really hard to quantify that which is optimal.I use the past 48 hours average.
"There's a sucker born every minute" - P.T. Barnum
  • #128
  • Jul 9, 2014 3:25am Jul 9, 2014 3:25am
Quoting Proximus
{quote} What ATR setting do you use, since its really hard to quantify that which is optimal.I use the past 48 hours average.
Since I trade mostly on higher timeframes I use past 12-18 days average (last 72-110 bars for 4H) depending on the pair
  • #129
  • Jul 9, 2014 8:33am Jul 9, 2014 8:33am
Hi Proximus,

I wonder if it's really a good way to test an indicator by basing both entries and exits on it. If the indicator triggers infrequently, for example RSI 90 and 10, the system will be holding trades forever and the results will just be random performance based on what the underlying markets did.

Also I'm curious about whether you rearrange the time series inside your clusters. Are the length of the time series inside your random clusters the same?
  • #130
  • Jul 10, 2014 7:56am Jul 10, 2014 7:56am
Quoting foobar
Hi Proximus, I wonder if it's really a good way to test an indicator by basing both entries and exits on it. If the indicator triggers infrequently, for example RSI 90 and 10, the system will be holding trades forever and the results will just be random performance based on what the underlying markets did. Also I'm curious about whether you rearrange the time series inside your clusters. Are the length of the time series inside your random clusters the same?
It was just an approach, nobody said its the most optimal one, of course for any indicator there should be a neutral zone besides BUY/SELL zone where you could exit the trade.

For RSI you could set it to 50, since 30 is BUY and 70 is SELL.However i tested this method where you "buy & hold" type, based on my theory that if the indicator is accurate, then it signals the best position to get in and the best position to get out.

So its logical to buy at buy signal exit buy at sell signal and vice verse, cos if the indicator is really accurate it should give the signals only when there is a real trend change, not at fakes.

As for intra-trade drawdown, that is due to fakeouts.Its not a healthy method from MM perspective, but it engulfs all noise and you only trade pure BUY & SELL signals.

The approach was really meant to measure the accuracy of the indicators not to provide a small risk trading system.
The length of the cluster has a random seed, however its mostly based on volatility cycles * by a random number to encompass an error margin.
Previously i used only random seeds but it turn out that it was innefective so i introduced a volatility based cluster.
"There's a sucker born every minute" - P.T. Barnum
  • #131
  • Dec 7, 2014 3:41pm Dec 7, 2014 3:41pm
Hey, why are my 3 favorite indicators not even listed or mentioned here? Just wondering... Something for you to think about, sir.
Please check out my blog at https://from-zero-to-hero . net
  • #132
  • Dec 16, 2014 12:16pm Dec 16, 2014 12:16pm
proximus how did you backtest your algorithm? was it on unknown data? or was it the same data that you used in the devellopement?
  • #133
  • Last Post: Dec 17, 2014 2:56am Dec 17, 2014 2:56am
I think one should have 2 different TPs. One is static, one is dynamic, whichever hits first.

In my opinion dynamic TP would be something like this:

For Long entries for example :
EntryPrice + (Sqrt(BarsSinceEntry) * (ATR + (Stdev(ATR) * x))
Here the x would be your TP multiplier (in reality that would be the Z-Score) which should be between 1 and 2.58).

This way if there is a spike in you favor, the trade would take the profit quicker, if the market is slow than it would hit you static TP eventually.

Just a thought....
0 traders viewing now