Is optimization reliable?

I have checked a couple of trading systems here, which didn't had EA, then I wrote and tested it.
I was so unlucky as I found to much DD for me.
Regardless how much the win ( even is 100x of the initial capital) I can't risk more then 30% of my capital, but I don't feel good at watching the -20%.
For this reason I have this at the EA optimization, I got a result like this:

It is for control points, so it may differ from fine clicks, but I got an idea where to search the Holly Grail.
Lets assume the each click will show the same result.

It is used a SL and a TP too, which is target for optimisation, but the max range is around 100 pip. I saw results with 20 -30 pip SL and 80 pip TP and 90 pip SL and 80 pip TP.

From January 1 total trades 139 ( max 1 per day I think).

Profit factor 1.72. For me anything above 2.0 seems to be suspicious.
Drawdown 18.52%, it is an acceptable for me if is an 500% win chance.

I don't know the "Expected Payoff" what means here.

There is another trade at the second line with 18.67% DD, 2.33 profit factor, but 72 trades so half of it.
Maybe the strategy tester was able to find an over optimized parametrization.

Fine optimization takes a while... 13h more..


What are your thoughts?

Well, regardless of the fact that in the optimization process MT4 doesn't simulate all real world events, a general problem with optimization is overfitting issue, which makes most wonderful EAs performing great in backtesting worthless in practice.

Thanks for opinion.
To not over fit I didn't choose the top result.


For example here is a 29k result from initial 2k.

I wouldn't expect that, but a 6k it would be my target.

Anyhow I feel a luck to work the market the same way which result 29k.

At the optimizations results, which param , results can be considered or all are useless?

What if I choose -somehow- to fit the EUR/USD and the GBP/USD?
At least it shouldn't over fit only 1 series of data ( but 2)

I've learned that optimization can make anything look profitable. I actually made several dozens sales of an EA on the MQL5 marketplace because my EA looked so awesome, when in reality it performed like crap. Not that I was trying to scam people, I was just as naive in my 'understanding' that the past is a perfect example of the future.

I think optimization is good up to a point. I think it's good to determine appropriate lot sizing for your desired level of risk. I think it's good for isolating specific trading sessions to see which session your strategy performs best in. I also think it's good for choosing your filtering method. But like has already been mentioned, I think when you start modifying parameters of indicators you run the risk of curve fitting.

I have used a lot the backtest and optimitation of Metatrader.

After dozens of EAs, I am pretty sure that it is not a reliable tool.

Most of the time, the data provided is rubbish, and in a real account will never work

I agree with you.
Backtest data are very strange, not corresponds to selected time periods.

optimization result:


Data set overfit, because it is an exponential curve:

So this is useless can't expect in reality with other data

But how about those which made 17k to 47k range? - there are many.

Anyway from 1k to 17k not so bad

Hi all,


I am afraid that you dont have enough trades to tune any parameter.
100 trades are just enough to fix one parameter, and then, it's risky.

Balance is not a good criteria. PF is better. Or Expected Payoff.

Expected Payoff is the mean return : %win * winAverage - %loose * looseAverage.
It's relevant if you use a fixed lot size during all the backtest project.

Another point : dont increase lot size within equity level during backtest. Use fixed lots (except if you tune lots with ATR or VIX of course). But dont increase lots because balance is bigger. (unfortunately a consequence is that DD% does'nt have any sens yet). You have to compute yourselt DD% with : MaxDD / Initial Balance.

How many parameters do you have ( stop, take, any else ) ?
You need at least 100 trades * Number of parameters to launch an optimization.
( Personaly, I become nervous if I have only 100 trades per parameter ).

A vicious phenomenon appear during optimisation : the more the optimisation found tunned parameters which decrease the number of trades (i.e. filters ), the more the result could be over optimised.

Over optimisation learn the noise of the past. Then, there is no predictive value.
People think that the result will be a leatle lower. That's not the point : the result coul be negative in the future.

People think that most problems of optimisation are : different market condition, different executions, bad history datas...
They are right, but it's not the key point. Over optimisation (curve fitting) is the big issue...