June 13 2019 updates.
We now use signP with weights.
So result = indicator1^weight1*indicator2^weight2..
In the past we used indicator1*weight1+indicator2*weight2
All weights should be set to 0.5,2, step 0.25, also for walk forward parameter weights.
Once the best indicators are chosen, we now take the top 250 systems of 50,000, not the top 1000.
Note 50,000 systems may be too high a figure as we have less possible duplicate combinations now. The c/us ration will start increasing it a clue to stop.
I am using 30,000 roughly. 50,000 is better if you can achieve it.
We are now picking the top systems by fitness with the metrics below.
Previously it was np/dd and pearson's combination.
We use crossover entry type only. In the past we used all 3 entry types.
We are back to using multiply operand. Recently we changed from * to +. Now we are back to * but with SignP
End results of all these changes
See Proving the methodology for more details.
Bottom line is this gave >50% increase in out of sample performance.
Artificial intelligence walk forward.
Typically there is a lot of merit to optimizing each input manually in TS. Its not essential but sometimes things come up.
For example if indicator1*weight1 gave best results when weight1=0, you know you've got a redundant indicator1.
Well AI should be able to tell you this.
So this is going to roll out slower over time. There will be a new WF called AI.
It will start by just doing some graphs, then later builds of GSB will draw its own conclusions on the graphs.
Other things under testing.
I am going to test results of walk forward using data pre 2018/2/28 instead of pre 6/30/2015