Question on genetic optimisation - page 6

 
Angela >> :

I have adjusted the best optimisation and forecasting parameters from the beginning of the year, the picture is not impressive:

...

More than 800 runs are not encouraging, the optimization results on the June story are worse than when tested with the initial parameters on the same period.

If I understood the phrase expressed a bit earlier correctly, did the optimisation give worse options than with the parameters before optimisation? This may be the case if some constraints are set on the optimisation, e.g. drawdown. Otherwise, there is something wrong. Also check that it does not happen that you optimize a small area (because of performance problems), and then do the forward test on a larger period. You need to stick to the inverse proportion: e.g. you optimise on a few months, then test on the next month, not vice versa. Otherwise you are giving the impression that the optimization results of June were tested in several months ;-).

 
marketeer писал(а) >>

If I understood the phrase expressed a bit earlier correctly, did the optimisation give worse options than with the parameters before optimisation? This may be the case if some constraints are set on the optimisation, e.g. drawdown. Otherwise, there is something wrong. Also check that it does not happen that you optimize a small area (because of performance problems), and then do the forward test on a larger period. You need to stick to the inverse proportion: e.g. you optimise on a few months, then test on the next month, not vice versa. Otherwise, you are giving the impression that the optimization results for June were checked for the next several months ;-).

Yes you understood correctly, only I used GA, and I have a suspicion that the best options were not found, and the best ones, including those with parameters that were before optimization, were discarded, I did not set any restrictions.

I do optimization for June, forward for July and 12 days of August. I limit optimization to one month history, because it takes more time when increasing the history, and I plan to reoptimize each week of the history, on which the optimization is done in one month.

 
marketeer >> :

That's the thing about closed bars: all four OHLCs are the same


Please give me an example from the quote archive, it's not clear what you mean...

 
OrlandoMagic >>:
Насколько я понимаю, у баров кроме цены открытия и закрытия есть еще максимальное и минимальное значения, которые могут оказать влияние... Ну да автору виднее...
>> :

That's the thing about closed bars, all four OHLC's are the same

OrlandoMagic wrote :>>

Please give me an example, from the quotes archive, it's not clear what you mean...

All readings of any non-zero bar (including max and min) will be the same in the tester as they came from the server online. It's not clear what's not clear.

 
But the highs and lows do not coincide with the opening and closing?
 
They may overlap, but they are usually different. What's that got to do with it?
 

Testing will not be correct because of this. The Expert Advisor can look at the opening bar, at the closing bar, at the arithmetic average. But it will have to perform trades on the whole range that is in this bar. This is why people test on all ticks. If there is a problem with the speed of such testing, one should find out where the program is wasting time and simplify it. This can be done, for example, by commenting out blocks of the algorithm one by one. As far as I understand, testing by opening prices and checkpoints is used only when testing an idea...

 

You are wrong. It all depends on the algorithm of the Expert Advisor. If the Expert Advisor works on completed bars, it makes no difference whether it received this bar from the tester or from online.

Yes, there are Expert Advisors that work on every tick - you really cannot test them on open prices.

 

OK, I'm wrong. Test it however you like.

 

How do I treat this? Optimisation was done from 1 May 2008 to 1 May 2009. Doing a forward test from 1.01.2008 to 1.05.2008 and from 1.05.2009 to today. The opposite picture is different, so what should I believe? How will my TS behave in reality, if tests on both sides of the optimization range show opposite results? A run in the tester with parameters obtained through the optimization range is also different from the results obtained in the optimization itself. Anyway, the further I get, the less confidence I give in this optimization.

Strategy Tester Report
Alpari-Demo (Build 225)


The symbol GBPUSD (Great Britain Pound vs US Dollar)
Period 5 Minutes (M5) 2008.01.02 09:00 - 2008.04.30 23:59 (2008.01.01 - 2008.05.01)
Model By open prices (only for Expert Advisors with explicit bar opening control)
Parameters Lot=0.1; StopLoss=11111; TrailingStop=0;
Bars in history 25288 Modelled ticks 49411 Simulation quality n/a
Chart mismatch errors 0
Initial deposit 1000.00
Net profit -29.65 Total profit 256.40 Total loss -286.05
Profitability 0.90 Expected payoff -2.47
Absolute drawdown 109.85 Maximum drawdown 270.25 (23.29%) Relative drawdown 23.29% (270.25)
Total trades 12 Short positions (% win) 0 (0.00%) Long positions (% win) 12 (58.33%)
Profitable trades (% of all) 7 (58.33%) Loss trades (% of all) 5 (41.67%)
Largest profitable trade 76.20 losing deal -122.80
Average profitable deal 36.63 losing trade -57.21
Maximum number continuous wins (profit) 3 (116.60) continuous losses (loss) 3 (-153.45)
Maximum continuous profits (number of wins) 116.60 (3) Continuous loss (number of losses) -153.45 (3)
Average continuous winnings 2 Continuous loss 2

Strategy Tester Report
Alpari-Demo (Build 225)


Symbol GBPUSD (Great Britain Pound vs US Dollar)
Period 5 Minutes (M5) 2009.05.01 00:00 - 2009.08.26 18:09 (2009.05.01 - 2009.08.27)
Model By open prices (only for Expert Advisors with explicit bar opening control)
Parameters Lot=0.1; StopLoss=11111; TrailingStop=0;
Bars in history 24900 Modelled ticks 48796 Simulation quality n/a
Chart mismatch errors 0
Initial deposit 1000.00
Net profit 425.30 Total profit 467.70 Total loss -42.40
Profitability 11.03 Expectation of winning 30.38
Absolute drawdown 0.90 Maximum drawdown 89.60 (7.19%) Relative drawdown 7.19% (89.60)
Total trades 14 Short positions (% win) 0 (0.00%) Long positions (% win) 14 (92.86%)
Profitable trades (% of all) 13 (92.86%) Loss trades (% of all) 1 (7.14%)
Largest profitable trade 96.55 losing transaction -42.40
Average profitable deal 35.98 Deal loss -42.40
Maximum number continuous wins (profit) 10 (426.40) Continuous losses (loss) 1 (-42.40)
Maximum Continuous Profit (number of wins) 426.40 (10) Continuous loss (number of losses) -42.40 (1)
Average continuous winnings 7 Continuous loss 1

Optimization results

Passage Profit Total trades Profitability Expectancy Drawdown$ Drawdown%

25 ____656.40____ 22_________ 3.28_________ 29.84_______ 176.40___ 13.90%

Optimization range test results

Strategy Tester Report
Alpari-Demo (Build 225)


Symbol GBPUSD (Great Britain Pound vs US Dollar)
Period 5 Minutes (M5) 2008.05.01 00:00 - 2009.04.30 23:59 (2008.05.01 - 2009.05.01)
Model By open prices (only for Expert Advisors with explicit bar opening control)
Parameters Lot=0.1; StopLoss=11111; TrailingStop=0;
Bars in history 74060 Modelled ticks 146623 Modeling quality n/a
Chart mismatch errors 0
Initial deposit 1000.00
Net profit 519.04 Total profit 849.08 Total loss -330.04
Profitability 2.57 Expected payoff 23.59
Absolute drawdown 60.44 Maximum drawdown 218.16 (17.09%) Relative drawdown 17.09% (218.16)
Total trades 22 Short positions (% win) 0 (0.00%) Long positions (% win) 22 (63.64%)
Profitable trades (% of all) 14 (63.64%) Loss trades (% of all) 8 (36.36%)
Largest profitable trade 151.16 losing transaction -78.80
Average profitable deal 60.65 losing trade -41.26
Maximum continuous wins (profit) 5 (341.80) Continuous losses (loss) 2 (-47.20)
Maximum Continuous Profit (number of wins) 341.80 (5) Continuous loss (number of losses) -78.80 (1)
Average continuous winnings 2 Continuous loss 1