<- Back to Tasks

Weighted Fitness (multiple goals) - Fitness calculation uses columns outside of the selection

It appears as though the Weighted Fitness (mutliple goals) option for fitness scoring does not operate as I would expect.




For example, when using a custom databank column that only outputs 0 (to test this), fitness scores of 0.5 - 1 are still being output and seemingly based on the amount of net profit a strategy made. The same trait was initial spotted when I was using only Winning Percent in the selection of columns and I noticed that many strategies had a higher fitness score simply because their net profit was higher, even though the Winning Percent was lower.


This is completely misleading - fitness should be calculated with exactly what has been selected, nothing more, nothing less.


Is it possible to understand how exactly fitness scoring is calculated per databank column as well as how the aggregate fitness is calculated when Weighted Fitness is in use?

Attachments
  • Votes +5
  • Project StrategyQuant X
  • Type Bug
  • Status Fixed
  • Priority Normal

History

WH
#1

Stormin_Norman

28.10.2020 13:35

Task created

WH
#2

Stormin_Norman

28.10.2020 13:46
Voted for this task.
AT
#3

AngelTalavera

29.10.2020 00:25
Voted for this task.
JK
#4

Insanity82007

29.10.2020 12:36
Voted for this task.
JK
#5

Insanity82007

29.10.2020 12:37
I've also noticed weighted fitness not looking correct. I use weighted fitness with same weighting between Stability and Ret/DD and I'll often see a higher fitness for a strategy with lower stats for both ret/dd and stability than another strategy with better stats.
b
#6

bentra

29.10.2020 23:22
Voted for this task.
JH
#7

Jabezz

30.10.2020 06:19
Voted for this task.
TT
#8

Tamas

02.11.2020 12:06

Status changed from New to Waiting for information

Can you please attach some strategies + printscreens where I can see the problem ? 
WH
#9

Stormin_Norman

03.11.2020 16:55

Attachment bugged_fitness.jpg added

Attachment correct_fitness.jpg added

Attachment custom_fitness_option.jpg added

Attachment Strategy 12444 fitness bugged.sqx added

Attachment Strategy 423285 fitness ok.sqx added

custom_fitness_option.jpg
custom_fitness_option.jpg
(126.97 KiB)
Strategy 423285 fitness ok.sqx
Strategy 423285 fitness ok.sqx
(68.51 KiB)
bugged_fitness.jpg
bugged_fitness.jpg
(146.90 KiB)
correct_fitness.jpg
correct_fitness.jpg
(156.90 KiB)
Strategy 12444 fitness bugged.sqx
Strategy 12444 fitness bugged.sqx
(70.37 KiB)
I have uploaded evidence which shows that fitness is completely random sometimes. Strategies also attached as well as the custom fitness columns in use (only the 3 columns selected are used - none are hidden). Note that this issue happens on other custom fitness columns too, it is not specific to those selected in this specific example.
b
#10

bentra

03.11.2020 17:06

Attachment Screenshot 40.png added

Attachment Screenshot 39.png added

Screenshot 39.png
Screenshot 39.png
(224.61 KiB)
Screenshot 40.png
Screenshot 40.png
(103.25 KiB)
Is the fitness column supposed to be the same as "weighted fitness?" 

The "best strategy" in the dash seems to be correctly the highest sharpe ratio. But "fitness" column shows something else so it seems like they are two different things...

TT
#11

Tamas

09.11.2020 09:30

Status changed from Waiting for information to In progress

TT
#12

Tamas

16.11.2020 09:48

Status changed from In progress to Fixed

Weighted fitness is computed only from the stats selected in UI.


If the stat value range is 0 - 1 then yes, the fitness is the same as stat value. Good example is SharpeRatio


We made some changes to improve the fitness computation.

If you still have a problem with fitness, please attach a bunch of problematic strategies to reproduce the issue.

WH
#13

Stormin_Norman

16.11.2020 11:50
"Weighted fitness is computed only from the stats selected in UI."


If this is the case, can you please explain why in the provided examples the fitness suddenly jumps up from 0.8x to 1 whilst having far worse stats?


"If the stat value range is 0 - 1 then yes, the fitness is the same as stat value."


What about for stats where the value is not between 0-1? A good example for this would be Net Profit.

TT
#14

Tamas

16.11.2020 12:39

Attachment transform.jpg added

transform.jpg
transform.jpg
(96.02 KiB)
For these values the fitness must be transformed to range 0-1


Attached a printscreen with a logic of transform method.



WH
#15

Stormin_Norman

16.11.2020 13:54
Thanks for the insight into the transformation logic. That's helpful for future questions/concerns.


However, I'm still not sure why the fitness changed from 0.8x to 1 in my example systems. Do you know exactly why?

TT
#16

Tamas

16.11.2020 15:10

 fitness changed from 0.8x to 1

>>>>>>>>>>>>>>>>>>


Fitness is calculated with the same logic every time so the result must be the same.


Please attach the problematic strategies and write step by step how to reproduce the problem.


WH
#17

Stormin_Norman

16.11.2020 15:53
That's exactly what I attached in the previous comment with print screens + SQX strategy files. The strategies were taken from the same result set (same custom fitness columns used).


I understand that the fitness is the same calculation each time, I just do not understand how these strategies are rated with maximum fitness (1).

h
#18

hankeys

16.11.2020 19:31

i really dont know how it really works and i am not using fitness at all, but from my work i can see, that even i am ranking only by RDD i can see unsorted fitness if i sort by RDD


isnt it because the fitness is calculated for each island sepparately or after each gen. restart from scratch...so the fitness values are uncomparable, because they have origin from different island or generation?
b
#19

bentra

17.11.2020 04:26
@ Stormin, I think they mean they may have fixed it and if you find more problems with build v130 then post again.
Votes: +5

Drop files to upload

or

choose files

Max size: 5MB

Not allowed: exe, msi, application, reg, php, js, htaccess, htpasswd, gitignore

...
Wait please