At least in my experience finding order from past chaos by tuning parameters until curve fitted and expecting it to follow in futures is not working long term.
I think its fundamentally wrong to optimize because stock market is not a sine wave from signal generator.
There are extremely low frequencies present that in some cases may not even have done single phase shift from companys ipo up to rare but quick spikes.
If algo idea comes to mind that involves parameters that have close to direct effect on output signals im probably going to skip the idea.
Also dont think common university stuff and ai techniques that gets taught to millions of people can work out of the box.
If too many people with similar base knowledge and ideas enter market are they is going to have advantage or are they going to reduce inefficinecys to 0?
In my case best result came if using parameters that change some fundmental way of looking at input data and keeping flexible code instead of parameters in places where parameter has chance to interpret data or overoptimize to past.
Also may be good to have and run multiple variations of the same algo with dif parameters , equalize asset allocations(as same algo with dif parameters proabably will contribute more or less) or even use multiple algos with each having multiple parameter combos.
At the end the end result could be generated into single graph and it probably will look smoother than using 1 set of parameters.
So far i have slightly over half decade of research experience and most of it half time.
I think many trading algo devs goal is to get past the awful repeating cycle of idea-develop-test and replace with with modify-test + fully automated but it might best to take it as process not goal.
Even if you get it live market competition also evolves and adapts constantly and at some point the system might not work and could step back into idea-develop-test...
In my case the market inefficiencys will probably vanish profitability at some point as i see gradual almost linear decline in profitability in over 5 year historical out of sample tests.
I think current market conditions hold good potential for algo trading because historical tests over 2008 and current times show similar peaks with higher returns than normal.
For someone else their experience and knowledge may be completely different as there are infinite ways to interpret data, limited by imagination.
Last edited:
