I'm currently brainstorming about why the markets change and trying to take those subjective ideas into my risk monitoring engine.
Let me start this discussion with the following:
1. Bad coding is out of the question. (ex. OMS/OES is broken)
2. Test pitfalls / testing systems unrealistically is not part of the discussion. (ex. tests using limit orders with fills on a swing top/bottom.)
3. Trade costs (commish and equipment) and IT stuff. Let's keep this out of the discussion for now.
4. No one is smarter than the other. If I have a system, it's safe to assume that another person should too.
In anothers, this discussion is about a system performance and the market. All things surrounding it is not part of the discussion (It's important, but please keep it out)
-------------------------------------------------------
That said, let me start off this discussion....
Everyone uses the same source of data to develop models. So there's a high possibility that the multiple developers will come up with models that utilize the same underlying tendency. You have computers developing models and there's a lot of multiple layer optimization techniques that data mine and working on exposing every little edge they can find.
1. Let's take curve-fitting. Who would be making more money... A model that re-optimizes their parameter real-time or a model that re-optimizes after they see a loss. I've seen this done in person, by a larger hedge fund which is taking a "reasonable" system and constantly re-optimizing it parameters real-time so that all the trades done are always optimal. I've done a lot of tests with this, it adds a lot of value if you have a "reasonable" base.
In another words, it's a walk-forward enabled model. Re-optimizing on every tick is very consuming so there are a bunch of trigger based WFs like windowed-time, fitness-value, and market condition.
2. Fade systems. If you know an over-optimized system will fail, then why not trade the worst performing system out there? Obviously, these system continue to fail (statistically), most likely due to not getting an attention = doesn't change dramatically to a point where it reverses and makes money.
I've ran, A LOT of, tests regarding points 1 and 2. Best fit and worst fit does not work well or has issues. But there's no need to deal with Best or Worst. All we need is to find the GOOD or PROFITABLE fit.
Any ideas? or comments?
