- ago
I had the idea to let the Evolver find a cute system for 15-Min Data.

After Generation 2 I see this:


It is "Strategy 26" which looks more than promising.

With generation 3 the picture changes dramatically:


When I select "Strategy 26" and go "Open in New Strategy Window" I can run a few Backtests with exactly the same settings.

Results are:


The detailed backtest results look unsuspicious. No hint for data errors or similar.

This same effect repeats again and again.
I see similar "fluctuations" with EOD data, probably explainable with NSF positions,
but the effects are much larger with 15-Min data and a bigger number of trades.

What can this be?
Any ideas, hints, suggestions?
0
293
Solved
13 Replies

Reply

Bookmark

Sort
Glitch8
 ( 13.75% )
- ago
#1
I'd run the data through the Data Manager cleaning tool and look for anomalous spikes.
0
- ago
#2
It detected a few (< 700) spikes which looked quite Ok for this data under closer investigation.

I created a new DataSet with the spiky symbols removed.

Now the Checker says:


And still, there are these inconsistent APR values from run to run for the same strategy.


0
- ago
#3
I checked with other DataSets and Daily data also.
Conclusion:
Evolver is broken since build 29.

It finds some nice strategies with high target metric values and forgets them in the next generation.
0
Glitch8
 ( 13.75% )
- ago
#4
The only reason it will drop a strategy is that it doesn’t satisfy your established filter 🤷🏼‍♂️
0
- ago
#5
QUOTE:
The only reason it will drop a strategy is that it doesn’t satisfy your established filter

There is no active filter!
0
Glitch8
 ( 13.75% )
- ago
#6
Then the other possibility is there are too many strategies with the exact same gene code. The evolver will drop some to increase diversity.
0
- ago
#7
As you can see in the very first post, this happen also at the beginning, when there are few generations and many different Genes.

Here is an example using Nasdaq-100 and very small position sizes (to reduce NSF Positions):


Strategy 252 made it to a good rank.

Some generations later:

Strategy 252 has disappeared.
0
Glitch8
 ( 13.75% )
- ago
#8
Yes there were too many other strategies with the exact same gene code.
0
- ago
#9
It is hard to see the effect with EOD data and low NSF ratios.

With 15-Min data the effect turns Evolver useless.

What I noticed:
In the very first post, look at Strategy#5: It has a very low NSF ratio.
Still the number of positions changes dramatically between gen 2 and gen 3, 1982 vs. 6738.
I think such a difference is not explained by the NSF ratio...
0
Glitch8
 ( 13.75% )
- ago
#10
Also that strategy had a very high nsf ratio so the next generation’s run was probably not as good so it got dropped.
0
- ago
#11
I think I cought him in the act:

Here Strategy 104 was calculated to have a very high APR, low NSF Ratio and 32k Positions:


When Evolver reevaluated Strategy 104 it changed dramatically:

Please note: The Evolver "sees" only 8800 Positions.

Consequently it disappeared form the list.

But why are the results so different in the first place?

(And if this Strategy is run through a normal backtest the results are much closer to the better ones)

It looks like Evolver "loses" man of the positions when it evaluates a Strategy a second time....
(The same happens to Strategies 46 and 61 for example)
0
- ago
#12
Problem solved!
It was a data problem.
0
Best Answer
Glitch8
 ( 13.75% )
- ago
#13
Just as I suspected all along!
0

Reply

Bookmark

Sort