Phoenix optimization - page 2
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
OK, I don't get it now
Is it O.K. to use build 200 with imported Alpari history and downloading new data from Alpari (since it is Alpari's edition) or not? As I'm sweeping through the internet, it looks to me that there are different opinions about this issue...
Or do I have to switch back to build 198 to be able to test on Alpari? If this is the case, how do I do it?
Or is there a possiblity of using the download feature? Is the data reliable?
And one more general question: which data do you generally recommend? I don't have a live Forex account yet, so I can choose the one on which Phoenix works best... I know you have different issues to slove now but we will have to come to some kind of conclusion in this department eventually anyway ...
Thanks!
I think you could use Build 200 too if you wanted to.
You just need to make sure you don't have 2 sources of ticks 'polluting' the history files.
I have a MT4 which isn't connected to the internet (great idea Hendrick!) so no new data gets in.
I've deleted all tick data (all the *.hst and *.fxt files) first of this MT4 and then imported the history files the old fashion way (so NOT with the download button).the most important thing is that your optimization/testing platform is disconnected. so that you dont corrupt your history files.
After experimenting with all the different settings for Phoenix, my conclusion is that the "Percent" and "EnvelopePeriod" values have the biggest impact on when and where the signal to buy or sell is located. For now if we just focus on optimizing these two variables, we can get 90% completed.
Also has anyone looked into using Phoenix on 30min and 60min charts? I believe there is some potential for use as signal verification.
ericbach: good idea about checking different time frames for signal verification, i think that is something people could test w/o any more code changes after the "signal output" is printed to a log file. That is different than converting 4 indicators to 8 but it isn't by any means absolutely conclusive. To be moderately "conclusive" we would need to test which indicators make the most sense to use as confirming signals, and which time frame (M5, M30, H1) for each indicator.
regarding "which data to use" download vs Alpari, and which version to use, "Build 198 or build 200" I think the ultimate answer won't be found through backtesting. It will be found by testing the backtested optimized results on live data. We can get a hint while backtesting. We can backtest 03-2006 to 08-2006 to come up with settings, and then 9-1-06 to 9-29-06 to see how those settings perform. This time period is available for all to use using both methods. Some people also have October & November live data, while others do not. Getting settings from March to August (03 to 08) and then using them again backtested on Oct 1-Nov17 will tell us how those settings perform using data from a specific broker.
At this point I think we should seriously consider creating archives of Real vs Demo data from Oct1-Nov17 and use those 7 weeks as a "standard". With 10 brokers data (archived from live) we can determine which brokers Phoenix performs well with 'in the current market'.
There is an additional "optimization confirmation" step that I was told about yesterday by someone off thread. Run your live testing for a month, and then run a backtest to see if the trades live test match what the back test results should be. We wouldn't need to do this all of the time, but it could answer the "Build 198 vs Build 200" issue.
The timing for this data crisis couldn't be worse in terms of "confidence" but I am glad we can pull together as a community to figure this out instead of rely on Hendrick. We may be the first community to really solve this, but Phoenix is one of the more settings dependent (and profitable) EA.
Any thoughts on this? I might be able to get Oct 1-Nov17 on several brokers. There is a way to get additional "live" data: scroll backwards and you can see the download start.
ericbach: good idea about checking different time frames for signal verification, i think that is something people could test w/o any more code changes after the "signal output" is printed to a log file. That is different than converting 4 indicators to 8 but it isn't by any means absolutely conclusive. To be moderately "conclusive" we would need to test which indicators make the most sense to use as confirming signals, and which time frame (M5, M30, H1) for each indicator.
regarding "which data to use" download vs Alpari, and which version to use, "Build 198 or build 200" I think the ultimate answer won't be found through backtesting. It will be found by testing the backtested optimized results on live data. We can get a hint while backtesting. We can backtest 03-2006 to 08-2006 to come up with settings, and then 9-1-06 to 9-29-06 to see how those settings perform. This time period is available for all to use using both methods. Some people also have October & November live data, while others do not. Getting settings from March to August (03 to 08) and then using them again backtested on Oct 1-Nov17 will tell us how those settings perform using data from a specific broker.
At this point I think we should seriously consider creating archives of Real vs Demo data from Oct1-Nov17 and use those 7 weeks as a "standard". With 10 brokers data (archived from live) we can determine which brokers Phoenix performs well with 'in the current market'.
There is an additional "optimization confirmation" step that I was told about yesterday by someone off thread. Run your live testing for a month, and then run a backtest to see if the trades live test match what the back test results should be. We wouldn't need to do this all of the time, but it could answer the "Build 198 vs Build 200" issue.
The timing for this data crisis couldn't be worse in terms of "confidence" but I am glad we can pull together as a community to figure this out instead of rely on Hendrick. We may be the first community to really solve this, but Phoenix is one of the more settings dependent (and profitable) EA.
Any thoughts on this? I might be able to get Oct 1-Nov17 on several brokers. There is a way to get additional "live" data: scroll backwards and you can see the download start.It seems that everyone is forgetting that any testing or optimizing as according to Hendrick needs to be offline, which means DO NOT DOWNLOAD DATA FROM THE DOWNLOAD BUTTON.
See my post in the FAQ section for more information.
OK, thanks for the response. What if I'm using Alpari MT4 (therefore downloading Alpari data) and also Alpari history? I have been told that since both history and live data are from Alpari, there shouldn't be any conflict and the data should be consistent... Am I right?
This is for backtesting generally, I understand that only comparing backtest data and collected live data is the only way of testing which data suits best exactly to Phoenix.
It seems that everyone is forgetting that any testing or optimizing as according to Hendrick needs to be offline, which means DO NOT DOWNLOAD DATA FROM THE DOWNLOAD BUTTON. See my post in the FAQ section for more information.
It seems that everyone is forgetting that any testing or optimizing as according to Hendrick needs to be offline, which means DO NOT DOWNLOAD DATA FROM THE DOWNLOAD BUTTON. See my post in the FAQ section for more information.
The rationale provided for the recommendation was to prevent live data overwriting historical data. The download button did not exist at the time the recommendation was made - there was simply no alternative.
I recommend we compare the two and make an informed decision - especially since the data from DOWNLOAD button apparently does not belong to any broker. I want to confirm this on a low level after I finish this code piece. (nonPhoenix)
Attached are backtest results on North Finance MT4 Strategy Tester. Tick data was downloaded using new download feature in the History Center on Build 200. When trying to download tick data on IBFX, there were data gaps and data stopped at 09/29/2006.
I tested each Mode in Phoenix 5.6.04 as downloaded and with Daraknor suggested code changed + Count_Signal=2. http://www.forexfactory.com/forexforum/showpost.php?p=141935&postcount=901
You can see settings at top of each report.
Date range - 5/1/2006 to 11/17/2006
Deposit - $5,000, Risk 0.1
As downloaded
P5.6.04-Mode1 - Profit 13,997.71 - MD 29.51% - short 64(75%) - long 57(82.6%)
P5.6.04-Mode2 - Profit 47,186.22 - MD 30.22% - short 105(73.08%) - long 103(75.93%)
P5.6.04-Mode3 - Profit 613.16 - MD 41.28% - short 216(60.19%) - long 192(63/54%)
With code change and Signal_Count=2
P5.6.04-Mode1-2 - Profit 2,316.31- MD 42.43% - short 32(53.13%) - long 29(62.07%)
P5.6.04-Mode2-2 - Profit 1,184.47- MD 67.72% - short 84(50%) - long 83(65.06%)
P5.6.04-Mode3-2 - Profit 2,834.33 - MD 26.14% - short 117(46.15%) - long 120(52.50%)
If settings are not correct for comparative testing, please advise.
I'm curious as to how much difference in results between broker MT4 platforms and new build 200 tick data compared to alpari data.
Wackena
Wackena,
Thank you for the great data. When you say "code change" do you mean that the signal_count=2 was tick based or bar based? Interesting we had 2 consecutive losses several times at the charts I glanced over, but no more than that. I will sit down and study them at length soon. I think optimizing for minimal consecutive losses is more important than optimizing for less drawdown. It doesn't necessarily mean that you get more money in backtest, but it may mean we have higher accuracy with prediction. In the next experimental release I plan to offer "Set SL X pip profit at Y pip profit". I would set it to 0 or 1 normally; 1 because a breakeven trade counts as a loss for statistics. More of my personal code merged into Phoenix...
Wackena, Thank you for the great data. When you say "code change" do you mean that the signal_count=2 was tick based or bar based? Interesting we had 2 consecutive losses several times at the charts I glanced over, but no more than that. I will sit down and study them at length soon. I think optimizing for minimal consecutive losses is more important than optimizing for less drawdown. It doesn't necessarily mean that you get more money in backtest, but it may mean we have higher accuracy with prediction. In the next experimental release I plan to offer "Set SL X pip profit at Y pip profit". I would set it to 0 or 1 normally; 1 because a breakeven trade counts as a loss for statistics. More of my personal code merged into Phoenix...
daraknor,
As per your post, Bar based. 900 seconds ( one H15 Bar) minimum interval between orders.
Wackena