A bit more research / results, albeit from a short period of time... around when to set Max Open Trades to Zero (I'll call this Zero hour)

So I ran 300 bars, for each day, for the period 29th Aug to 16th Sept...15 days in total, with 3 x Mons, 3 x Tues, etc, so 15 Portfolios in total, with 50 EAs in each Portfolio.  Generation ran for 60 minutes each time.

Admittedly, I fixed SL and TP at 220 and 95 respectively, which I've since found is not the best approach - but, I didn't want to re-run everything...

I then modified the portfolio code so that I could then run M5 Optimisation, Balance Max, to see what the portfolio performance was if the Max Open Trades was set to 0 at specific hours after it was launched (Zero Hour).  Trades were then left to close automatically by the EA rules. 

All Portfolios were set to launch at 01:00 GMT (just after the spread change).  The code also ensured that no trades were opened prior to the launch date/time, despite what was set as the start date in MT5.

The last date I used was the 16th Sept, because it could still be closing trades naturally, if the Zero Hour was set at say 1 week after launch.

My hypothesis was that 15m CJ portfolios perform well for about 24 hours, and at that point, I should set it to not open any new trades.

Based on the result though, the optimal time, for my settings, appears to be around 44 hours.  At that point, I'd set Max Open to zero and let the open positions close normally.

The graphs show what happened to each portfolio when the Max Open Trades was set to zero at 1 hr, 2hr, 3hr etc after launch - all the way out to 120 hours (all charts), and 277 hours for each of the Weekday Charts. 

It might also suggest that you can pick different Zero hours for different days of the week.

* Weekday graphs under a new post.

Some notes:
1) Totally agree, this is only a short period of time.  15 tests in total. 
2) Launch time will no doubt change the results - I chose 01:00 because the EAS Generator requires a date back in time, and it sets it to midnight (or last bar) and so by the time I would have generated, and re-launched, it could be done within 1 hour, and then a few minutes to test it and launch it - hence 01:00.
3) I fixed TL & SL - not the best approach
4) Very specific to my settings and my broker data (IC Markets), which I can export up to the last previous bar.
5) Recommend you test this yourselves!

timelleston -- I give you two thumbs up!!  Thanks for performing and sharing the results of your analysis...

One additional feature not to lose site of -- the extremely low DD.  The resultant strategies generated by EA Studio make it very safe to trade a portfolio EA with dozens of strategies.  I sometimes see the accumulated lotsize of all my open trades > 2 (which is a lot for my smallish account), but DD remains in-check.

timelleston -- a question about the 'All Hours All Days.png' chart.  The dashed-red line for Total doesn't look right to me.  Why does it drop so low, especially during the first 15 hours?  It looks like it drops lower than the sum of the individual colored lines, unless I'm not reading it correctly.

Hey Sleytus,

Around this concept I am wondering what the Ideal generator/process is?

It seems we are looking for a way of tuning to the immediate market and then switching EAs daily.

If our own time wasn’t an issue for example would switching EAs more regularly than a day be a benefit.

Presumably ideally we would have a real-time generator as you say that is generating based on recent bars only and revalidating what in current collection on recent bars only also. Then when have a full collection we put portfolio of 50 or so on with a set limit to only open trades for 24hrs and close gracefully after that period.

Then rinse and repeat for the next day.

Oh and also a way of checking the previous days performance also since that need to be good. Aka 3-5 trades and profit factor of >1.2?

I think I am going to try a couple ways of implementing this concept. Trying to stay true to the concept but also trying to automate as much as possible and reduce the manual effort I need to put in continually.

Just going to go straight to practicing on demo so I’ll share the results.

Also have you tried any demo trading with this yet timelleston or still fine tuning in mt5 backtester?

Yeah I guess we would want around a month of good demo results to feel confident that we are applying this concept suitably.

I hoping to get something on demo working and then I will try a few variants on other demo account also and see what is typically working best.

It’s a super interesting and logical concept, how to best implement it has lots of little questions for sure though. First test week starts on Monday evening for me. I am going to trade this Way Tuesday to Saturday as I think skipping Monday so that the most recent day of data was literally yesterday rather than pre weekend makes sense.

I’m going to do two things:

1 - Try to follow what is discussed as closely as possible, using a portfolio for CADJPY M15, CADJPY M30 and also GBPJPY M15, GBPJPY M30.

   Generate using 300 bars data (aiming for roughly 5 trades a day), test with tick data, put straight onto mock live and repeat every 24hrs

2 - Generate a few portfolios of 50 for each (so 150 EAs for each symbol period combination) based on most recent three weeks data and put on incubator demo.

    -Then use FXBlue to every 24hrs select the top 30 performers for the last day (needs 3+ trades abs pf >1.2) and then move those to mock live account.

For the second approach I would be looking to change to fresh porfolio EAs weekly as expected they would not last long but hopefully it gets enough stratgies through that perform well yesterday for trading today.

I think I can do this with a couple of hours work each evening and will aim to do for at least a couple of weeks before considering how I may need to tweak my settings.

Also just going to do using Dukascopy demo accounts btw

Hi sammjacks...

1. Yep -- you got it... 

2. I would say "yes" for a few reasons:
    a. Since the goal is "near real-time" then the most recently generated portfolio EA should better align with current market conditions.  However, suppose everything was fully automated and you could add a portfolio EA every minute -- is that practical?  Probably not because the number of open trades would grow and the original EAs would still need to execute until all their trades were closed.  I have updated portfolio EAs during a single session and it works fine, but there would come a point when it becomes a logistical nightmare.
    b. As timelleston has shown, portfolio EAs have a profitable lifetime of about 44 hours.  The earlier in the lifetime then the better the performance.  This again relates to the goal of trading portfolio EAs in **near** real-time.
    c. The time at which a portfolio EA is launched may make a difference -- e.g. I waffle between shortly after Tokyo Open and just before London Open.  I see differences but I don't know if there is a consistent pattern of one being better than the other.  Market conditions change and it wouldn't be too surprising if they are more likely to change at certain times.

3. I'm trading it on a smallish Live account.  DD is amazingly low and it grows slowly -- so, after I attach the EA to a chart I don't monitor and don't lose sleep...

Agree with everything you've said there Sleytus... trade volume is low for sure, even in the aggregate, so may not be statistically valid.  A few days ago I did try 500 bars M15 as well.

I'm playing around M30 still - 300 bars M30 is 600 bars M15 (equiv) so we'll actually be fairly close here.  I'm just working through a similar series of backtests...  A quick check on M30 and it generally takes around 12 bars for it to have traded around 100 times in total...a similar problem.

What are you guys using for your stop loss range typically. I’ve just gone with 10-100, my approach is that since the logic seems sound if I don’t get reasonable performance after a week or two then I’ll be looking to play with the generator settings and acceptance criteria and repeat the process.

I’ve just gone for pf >1.2, r/d > 1, rsquared>70 and trades between 30 and 50 over 300 bars.

Hey Sleytus,

Will be starting some testing this week. Been doing a bit of backtesting today and been getting poor results generally so how I am applying this at the moment seems to be off.

Basically what I have done it run the generator with 1000 bars on CADJPY M15.

The setting so have used is:

Win/loss > 0.6
Min trades =40
Max trades =90
Min profit factor = 1.2
Min R/D = 1

Then when I have enough strategies in the collection I will rerun through the validator but just for the most recent 300 bars and then visually cull the strategies not ascending in the last 1/3 of the chart. 

Add the rest to portfolio and then test for the following day.

So far I’ve been very unprofitable.

Any pointers? It could be that I am not running the generator for long enough so working with collection that are too small? Pretty sure that is one of my issues of not the main one as I’m much under the portfolio size of 50 that has been discussed previous

Trying to get a balance of good acceptance criteria but not overly strict which will limit getting a decent amount of strategies after an hour or twos generation time.

It seems we want to ideally be able to complete this process as quickly as possible, 2-4hrs tops seems like a sensible window?

I have some thoughts / comments -- in no particular order...

* I think timelleston's Back/Forward analysis has revealed something most interesting -- that is, strategies have a profitable lifetime of about 44 hours (probably due to changes in market conditions).  This has some important implications.  It also explains why regardless how noble one's effort is in creating EAs they rarely work out.  It's not that their algebra becomes stale but, rather, market conditions change.  If market conditions didn't change we'd all be gazillionaires. 

* It's nice to be able to put a number on this.  44 hours isn't carved in stone, but it gives us a ballpark figure to work with.  It implies that a strategy is most profitable within a day (or so) of its creation.  And from personal experience that feels about right.

* In my trading and testing not every day will be profitable -- as timelleston's Excel charts have shown.  In my experience a "bad" day is usually one where the portfolio EA treads water, and a "good" day is one where the profit grows gradually.  And this is fine -- it just means that we need patience and a consistent workflow (and smart money management).

* I've tried other pairs and timeframes -- CADJPY and M15 or M30 are definitely the best.  I don't know why that is, but it makes sense to me that different instruments would perform differently.  Related to this, I think the "window of profitability" is narrow.  Fiddling with EA Studio configuration settings could make the difference between success and failure -- that is, it could be you're losing money today but a slight change in configuration settings improves overall results.  For example, switching from AUDUSD to CADJPY makes all the difference -- who'd have thought?  Right Tim? 

Definitely made a huge difference...so this works on different pairs differently.

I've been trying (mostly unsuccessfully I might add) to limit the losses that can occur (observed on the original analysis).  I've been re-generating strategies for dates used in the original analysis, but using some of the different EAS settings, such as variable SL/TP (10-100) and Reverse, or, Reverse / Ignore. 

Reverse Position seems to make the most logical sense to me...if it's not going the right way, on an Opposite Entry Signal, reverse the position.  Keep your stop losses tight. 

I've also added some custom code to go to close all trades if the profit drops below a certain amount...thats worked quite well (in the long run).

Getting some fairly scattered results though to this approach.

As an aside, accepting that curve fitting is probably unavoidable, I thought I'd try to go the other way, and fully optimise and then normalise the results and see what happens...figuring that over the course of the next 96 bars or so, if the market stayed roughly the same, it would be ok.

So I generated a portfolio as per normal, and then downloaded the collection, re-uploaded it to validator, and tried to Normalise, and then Normalise and Optimise.

Both results were less successful from the generator only.  They failed further and they failed faster.

That sort of makes sense though.

Optimisation, in this instance, makes it fit the curve even tighter over the last 300 bars.  It makes it even twitchier. When you launch, the slightest change in the market makes it fail faster, as this new data/scenario has never been experienced by the previous data set - it was overfit. 

My result: It failed even faster.  And it failed further too.

So now I'm trying to find ways to not have any "optimisation" done by the platform - other than strategy selection/levels within the strategy.  Even down to the SL/TP.  I'm slightly nervous of getting it to figure out what SL/TP to use - as that seems to be fit for that particular portion of the curve.  In my mind, if I set a specific and fixed SL/TP (or no TP) then thats a variable removed.  But, I might be totally wrong too. 

But hey, thats the beauty of this platform hey?  The ability to test ideas in almost realtime...

And I'm not saying that Opt/Norm/WFO etc is not valuable, but we're attempting to capture that slice of time where we don't need to optimise.