I tried to follow your way of thinking.
So, I made a run using more bars for same period. 3x300=900 bars
M5. I work on it.

Ok, we're all cooked up here (stopped at 1 hour), and my eyes + brain has melted down.

I actually can't believe the results.  They're really good.  I stand corrected Sleytus.  Each time-frame generally stood up for at least another 5 days.  Some more.

See the chart attached below.

I've re-aligned the dates so that each In Sample result starts at the same point - day 1.  Each run had 4 "days" of data as In Sample.  Some had slightly less, which is possibly why a couple of the results are lower.

Once they cross the vertical dotted red line, they're "in market".  This was the date the EAS generation stopped, but continued using data in MT5, up until midnight 28th...which allowed the last period to run for 3 additional days.

The blue dotted line would be the profit from closed positions if each strategy were stopped on the 3rd day.  Note that I haven't taken into account any open positions that would be closed if that approach was taken.  But given the drawdowns were marginal during those periods, I don't believe it would have made too much impact on the net profit.

So, that's another very interesting result.  Bear in mind also that these results are converted back into AUD as the account currency.

And I'm now curious if different pairs react better to algo-trading - maybe AU doesn't, but CJ does.

Oh man, another rabbit hole!

So, Sleytus, thank you - you've taught me something tonight.

Good stuff Sleytus.  I will continue down the rabbit hole, and will experiment more with CJ and others.

Its certainly an awesome bit of software - there is no other that makes the generation so easy and quick.  Some of the periods I used y/day had over a million strategies tested and ascended over 4000 in about an hour - whittled down to the top 300 in the collection, and then down to the top 50 in the portfolio.  Incredible stuff. 

Happy trading.

My 2 cents. This exemplifies how OOS usage turns out-of-sample data into in-sample data. And using the same OOS data for all strats brings about another minus - that is biasing the strats to the underlying OOS data set.

That is another point for debate! Generator in essence still uses optimization, the fact that it throws indis with different parameter values at the data is still a form of optimization. If default values would be used then it is still a form of optimization - a passive one. Where's the right balance?

OOS is not useless, but to me it is acceptable only in checking the methodology, not as a filter for strats.

Yes -- no disagreement.  I used "optimization" too loosely and should have said "computations".

But no matter -- when using only the most minimal data to compute settings for current market conditions then concern about "over curve-fitting" goes away.  This is because we no longer care if the settings are over curve-fitted, since if they are then they are over curve-fitted for current market condtions and that is what we want.  All we are hoping for is for current market conditions to persist for another day or so.

Nice work Sleytus.

Here's another little insight...

On CJ, I was experimenting with SL level.  I'd originally had both SL and TP at 95 pips.  I'd noticed a couple of things over the last couple of days... SL was getting hit, not that often, but it was getting hit. 

And the other thing was that the overnight spreads can go to 200 (checked it via ICMarkets downloading bars - max was 200 over the last month).  Problem with that is that if you're short (it tends to be the Ask thats jacked up), and the spreads go beyond your SL, you get stopped out.  Not good.

So, I modified Mr Popovs code for a test to see what happens when SL is moved out massively...my theory was that I'd write a script that moved my open short position stop losses to 250 pips when/if the spreads started to get towards my 95 (example only).

What I found was initially interesting. 

If I moved SL out to 200-500 pips (huge yes I know), as you'd expect, they didn't get hit...but, what was surprising was that it didn't impact the results from the portfolio over the next 1-2 days. 

What this was showing was that the reversal/close rules were working really well at closing trades (maybe for a loss yes) but not closing them too soon (i.e. getting stopped out to soon). 

Note: I'm a huge advocate of SL and money management for sure. 

So, I'm going to experiment some more, to make sure that this does actually happen over a longer period and I'll report back here.

I'll be watching equity drawdowns, length of open position and correlation between MAE and MFE. 

When we launch a new portfolio, we set the Max Positions to 0, which still runs the exit signals, just won't open any new positions.  So, unless you get a massive move in the market - which is possible - the exit rules should come into play and close out positions "on-time"...

There's no OOS, no Reactor, no optimization.  Only the Generator using 300 bars IS.  You should see good performance for about 1 day after IS ends, and then it tapers-off.

Let me rephrase my question: you showed a pic of craptester results, a graph. On it you have marked in-market line, which represents the start of unseen data, the start of simulation of trading in other words, correct? Now, is it 100% clean sheet? I mean there are no trades, which are opened before the in-market line and remain open after the in-market line until they are closed?

Very quick test results, for one period only, so I'd recommend doing this across other periods:

Data Horizon End: 27/9/22
Bars: 384
Resulting Start Date: 21 Sept 00:00
Resulting End Date: 26 Sept 23:45

In the interests of time, I ran the Generator for 30 minutes. It generated 2212 strategies, of which it kept 300.  I moved top 50 based on SQN to Portfolio then exported to MT5. 

Set start in MT5 as 21 Sept.  End as 28 Sept.
Set Custom Forward date as 27 Sept.

Result is attached image below.
Net profit for the 27th was A$107 on size of 0.01 lots (EAS was 0.1 lots).
I used a SL of 220 pips and TP of 95 pips.
In the backtest, of 621 deals, there were 0 SLs, and 32 TPs triggered.
In the forward test, of 146 deals, there were 0 SLs and 0 TPs triggered.

I had not added any custom code to manage spread protection either (normally I set it to 50 pips).
While not repeated across multiple dates, due to time, it does show that this method does actually work.  Of course, sometimes it won't, thats the probability of it, and we can't foretell the sequence of wins and loses, but, it definitely seems to be an edge.

Using the forward test, it closes all deals at the end of the back test, and starts new deals as needed on the forward test - so, Footon, to you question, this is a clean data set.  No trades held over.

I think you've said it best -- it's not the HolyGrail and there are no guarantees -- but it's an edge with a very ***low*** DrawDown.

Furthermore, it is only possible because of EA Studio's ability to simply and quickly generate high quality ***portfolio*** EAs.  Otherwise it would be much too tedious and time-intensive to have to create or recalibrate custom strategies every day to accommodate tomorrow's market conditions.