Expert Advisors optimization & backtesting

no I am using mt4

What MT4 writes in Experts audit ?

Found great free video and textual instructions on how to make EAs backtesting properly - see here.

Main idea:
1) Take 10 years history
2) Make optimization for the first 6-7 years
3) Run found parameters for the next 1-2 years - it is called "validation". If see bad results - adjust EA parameters and reoptimize again 6-7 years till "validation" phase is ok
4) Then run a final test on final 1-2 years - if OK - then you have stable EA parameters to run on live

So first optimize 2010 - 2017:



Then backtest 2017 - 2019:



Then backtest 2019 - 2020:



2010 - 2017 is nice setup but just found in historical data. But 2017 - 2020 two Excels prove it is stable setup. It was already some kind of real-life test because EA didn't know anything about 2017 - 2020 while backtesting 2010 - 2017.

This means a new paradigm for me... DO NOT TAKE THE BEST (MAX PROFIT, MIN DRAWDOWN) OPTIMIZATION PARAMETERS SET. But take the most STABLE parameters set.

Yes, this means additional time to optimizations/backtesting... but who said it would be easy here in forex?

Just play around with new backtesting approach and find out a stable martingale setup:



With martingale:



Without martingale:



Implemented in EA "virtual trading" to count looses and then optimized 7 years to get min looses series. Validated on 2 years and then additional testing for 1.5 years. Total 10,5 years stable martingale. I see the first time so stable martingale. Mostly any martingale fails after max 2 years in my previous backtests for any entry strategy. But here strategy very simple - DISTANCE X pips + ADX on bigger timeframe to detect RANGED or TRENDY market conditions.

What here is unique - EA trades "virtually" first then wait for needed loses count and then trades goes real. Before that EA just calculates "how that trade would go if were opened". Before the "virtual trading" feature in my EAs losing series was not possible to optimize in a good manner but now it is and I just made 1st attempt with GBPUSD...

Closed trades copy with 11% profit. Was down an additional 15% - that was ok according to 10 years backtest (total dd 32%). Closing copy because quite a big run-up 27%.

"Candles trend" strategy still on good run-up, for now ~76% up from the last ~30% drawdown:



This is nothing unexpected, from 10 years backtests I can see there were ~16 periods with run-up from 70% to 200%. So continue to wait till the next 15% drawdown to turn on a copy to my investor account.

Here current distribution:



All symbols according to individual backtests. Was closely monitoring GBPUSD because it reached max drawdown -15% according to backtests (see this post). And now it almost reached break-even with just -5% drawdown.

Meanwhile monitoring bitcoin price - when dropping to about 8000 USD then will move my investment account currency to BTC. So will diversify risks - if BTC still moves down but EAs earning or EAs losing and BTC up - then minimize looses. If BTC goes up and EAs profitable - then gaining more profits. If BTN down and EAs losing - well anyway that's my longterm investment. That's some kind of hedging - have 1 negative scenario versus 1 positive and 2 neutral.

Turned on trades copy from "Candles trend" in my investor account because 22% drawdown:

Still in the investing process...


what we can see that the 13EAs package had a 42% drawdown what is 10% more than the biggest dd in 10 years backtests history.

Seems correct is the approach: backtest, notice drawdown x%, and yearly profit y%. When go on live trading with that setup be ready to 2x% drawdown and 0.5y% profit. That's real-life...

Made optimization for 2010 - 2019, then forwardtest 2019 - 2020 and here question - what is good or bad?


From 2010 - 2019 max drawdown was 111 usd. In forwardtest 2019 - 2020 all resulted to drawdown 101 usd. Can be said "Forwardtest failed"... "not profitable entire 1.5 years". But drawdown is according to backtests... there were 1-year drawdown 111 USD... It might be the best time to turn on this setup because it run allready down so far - should go up. How to decide right here?

Some weird strategy but working...

The idea simple - opening SELL and BUY deals on GPBUSD with TP 75 and SL 70. When one hit TP, open another deal in the same direction but with twice volume. And keep opening opposite deal as well.



When 2 wins in a row - reset to initial volume (stop rising).

So this is the end of the year - let's make a summary of my trading so far.

Have a >9 accounts with more than 90 EAs running. The most interesting for me is the account that runs my 50% capital. On that account running 19 EAs starting from April 2019.

Here how it went... were some manual trades, some turned off EAs etc... now year ended with 19 running EAs on that account and here I analyze performance.


Here I made an analysis of my EAs magic numbers to see which particular EA was winning and which were bad:



So removed 2 losing EAs with DJI because their drawdowns were quite more than in backtests. All the rest losers still acceptable. Hope next year EURUSD EAs will recover.

Cashout 80% from profits. Will use them for stock trading strategies.. see here some nice ideas regarding stock trading. Will try next year.

Closed 3 accounts too... 2 wiped out (15% of my forex capital) and 1 on break even.

The main message here - do not rely on 1 symbol/strategy/setup - DIVERSIFY. You see that somehow worked at least for me. Happy new year!

YES YES YESS!!!!!!!!!
I'VE BEEN LOOKING FOR THIS KIND OF CONTENT MY WHOLE LIFE.
have been trading discretionarily for the past 1.5yrs and always fail miserably because of emotions
and recently switched to a more objective approach which is algo/systematic/automated trading but never been successful for the past 6months

and i was wondering what did i do wrong, now i know better after reading the contents of this thread, i didn't diversify.
and looks like u r more interested in building a more "price action" EAs than using indicators, just like Andrea Unger.
I'm intrigued! can i help u build EAs and post it here? ofc i'll follow ur rules, 10yrs of data and no more than 20%dd with 80%profit without compounding
but i lack resource (multicharts is the ideal platform, but must pay and i live in a 3rd world country), i'll be using tradingview instead, they got 10yrs worth of data on Daily chart.

edit: okay i just read ur terms & condition looks like using mt5, okay no worries i'll try that, but how to get 10 yrs of data?
edit2: nvm i found out that mt5 is very cool, can back-test from wayyy back, cant wait to build my first EA as i have a lot of ideas i wanna try ranging from single candles behavior to volatility trading strategies

this post is really insightful to do backtesting the right way, i also read darwinex's blog on backtesting. good contents man

6yrs testing & optimizing ur EA --> 2yrs validation (test1)--> 2yrs final test
if validation not good enough then back to optimizing
also if can, use timeframe on or above H1, because below H1 is more expose to over fitting & slippage & spreads,etc. (source: https://blog.darwinex.com/dos-donts-...4-backtesting/)
choose stable variable/parameters for ur EA, rather than 49 or 51 use 50 (but also if can, the result of 49 50 51 must be a smooth bell curve)

i.e. 49 profit: $100 ; 50 profit: $900 ; 51 profit: $400
this is an ugly bell curve, a smooth one u can use statistics (standard deviation) or just common sense: profits differents r not twice as much or as little.

a good bell curve is like this: 49 profit: $110 ; 50 profit: $99 ; 51 proft: $115
u wanna be using 50 even though 51 extract more profits, this is mainly a matter of stability rather than profitability which will expose you to: over fitting. (common sense number which is not expose to over fitting, the more common sense the number is, the more investors u'll get, cuz they don't like data mining)

Deleted

Deleted

80% traders lose. Ok. Chose 10 unexpierenced traders. Split capital to 10 parts. Copy their trades but in oposite direction. If trader open SELL - open BUY. Proportional deal volumes. Exit when trader close deal.

This approach should give positive outcome. Think not hard to find consistent losers to copy?

Why this wont work? Because of spreads? Timing issues?

If cant find good consistent losers then let's create some.

Let's make EA who is an ambitious "trader" and, "thinks" it can beat the market because it did it in the year 2019.

Use a simple entry/exit strategy: when RSI down enter BUY when RSI up enter SELL. Use SL30 and TP30. Use H4 timeframe.

Optimize EA to find the best ever setup for 2019. Here it is - our gorgeous super success results:

Attached Image (click to enlarge)

Attached Image (click to enlarge)

But we know this is bad "trader"... well... because it is overoptimized. That EA setup is considered to be bad according to best practices.

Let's see is it realy bad... so emulate forwardtest - our EA knows nothing about year 2020. Let's see how it goes... And here it is:

Attached Image (click to enlarge)

Attached Image (click to enlarge)

It is BAD

So we have loser EA. We allready at begining of 2020 know it is bad because it is overoptimized. So let's see how performs our reverse copy trading. We now run EA when bad trades are simulated in memory and oposite trades are done as real trades. Here it goes:





As we see - our bad "trader" wiped out all account. But we - smart investors earned 60% on that.

Remember - our reverse trading EA knows nothing about 2020 too. The same as looser. But "smart" EA knows that overoptimization does not works... so use it as advantage.

This is example how probabilities works. Now need to optimize/backtest/forwardtest about 100 various setups to ensure this idea is working and can be repeated several times.

It seems if EA lose entire account then reverse trading will earn about 40-70% but not possible 100%.

Ideas how to make bad EA:
1. Fixed SL/TP, should be more than 10 pips because 10 is too small (spreads etc). Seems starting from 30 pips works great
2. Small timeframe – less than H4. So get more noisy signals
3. More than 100 trades per year – we need overtrading

How long to run reverse trading? Till bad EA wipes out account or reverse trading earns 60%.

Please write more ideas what needs to be implemented for and realy bad EA so we can do opposite.

Great work, Joyny!

When I am less busy with other things I will definitely participate in this testing.

Here's a couple of threads to read about reversing a losing strategy. I'm not saying they apply in this case. I'm not familiar with how the 'bad trader' EA or 'smart' EA trades. It's just a few things to keep in mind.

https://www.forexfactory.com/thread/...43#post8832643

https://www.forexfactory.com/thread/...03#post8839403

The main point is that the losing strategy needs to be non-randomly losing otherwise reversing it will also produce a random result. Is this reversing EA also in the page1 JShifter combined EA?

Tnx. Good points!

No this is special ea made to be reversable. It runs "bad" logic in virtual mode and "smart" with real trades.

Yes transaction costs eat out profits of reverse trades therefore we need at least 30 pips sl/to range so spreads etc % is affordable. Still max what can get out is 60% yearly even if "bad" fails totally.

"Edge" here is that we with purpose create "bad" ea. It is overoptimized. But we still need portfolio of EAs because of fluctuations.

All battle is about fluctuations and we fight no more as for 1% in our favor. With rr 1:1 we need 51% profitable trades.

We need 100 different EAs setups here to test if concept "overoptimization is bad" is true. But here we need significant % to ensure "edge". If 51 vs 49 thats fluctuation too. We need to see that at least 70 overoptimazed EAs fail in forwardtests so thats no more fluctuations then but real edge.

Entire portfolio of 100 EAs with this method could generate no more than 20% yearly profits.. But stable year by year and more than sp500. For darwinex investors this might be ok.