Multi-Condition-Group-find-best-combination-for-Condition-Number-N-Is-True

Since I have to rename topics like "Multi-Condition Group Building Block" (not the first time today) and this is even a feature request (!), can you choose more detailed titles which would summarize your idea?

Just give it a thought: how can "Multi-Condition Group Building Block" be requested (again) if it's a baseline feature?

I’m not sure I follow the request, so you mean it would do it’s own kind of mini optimization to find the best ones?

It would be an expansion of this.

QUOTE:
Condition Number N is True
The second options instructs the Multi-Condition Group to resolve to true only if the specific numbered Condition Block is true. The first Condition Block is numbered 1. You may wonder why it would be useful to construct such a Multi-Condition Group, instead of just dropping the single desired Condition right onto the Entry/Exit. The payoff here comes when you make the How Many parameter optimizable. Then, during an Optimization, the optimizer will run through each of the Conditions, one by one, as the parameter is iterated. In the example below, an Optimization could determine which of the three oscillators performed best in the Strategy.

Instead of just determining which specific condition performs the best it would determine which combination of conditions works the best.

OK, thanks I get it now.

QUOTE:
Instead of just determining which specific condition performs the best it would determine which combination of conditions works the best.

This is a statistical question because you need to contrast improvement (i.e. change in significance) as you add or subtract terms in the multivariant (i.e. multiterm) regression equation.

Look up either stepwise regression or all possible subset regression in an advanced multivariant statistics textbook. SAS and the MatLab statistics toolbox will do the former. R will do either if you install the right advanced regression package. With R, it's possible to have a .NET application (like WL) pass the problem to R, and R can pass the solution back. This sort of analysis is common in research problems.

All possible subset regression is as the name implies. It tries all possible combinations of terms. So if you have 5 terms, then you have 5! (5 factorial) possible combinations to analyze. Stepwise regression tries to pick the most "likely" permutations based on contrasting improvements (but it might miss an important combination inadvertently).

It's good to see users interested in advanced statistical analysis. There's a great deal of opportunity to use it in stock trading.

---
Off topic, but you can achieve exactly the same result (as a multivariant regression equation) with neural networks (i.e. NeuroLab on WL). Some researchers would argue that for stock trading (which is a nonlinear, event-driven process), neural networks would work better than using traditional linear methods (like stepwise regression). So I have to ask, "Why don't you turn NeuroLab loose on this multivariant optimization problem instead?" It's a fair question.

Thanks for your answer / suggestion Superticker. May I ask how you would approach using NeuroLab for this type of situation? I can't intuitively think how to set up the optimization.

Lets say we have a strategy the we are trying to discover either:
a) combination of conditions(or entry rules) for an optimal entry given a set of conditions(indicator values) or
b) combination of conditions(or exit rules) for an optimal exit given the same conditions(indicator values)

QUOTE:
May I ask how you would approach using NeuroLab for this type of situation? I can't intuitively think how to set up the optimization.

Although I've used several different NNs back in the day, I haven't ever used NeuroLab on WL. It would be better if someone more experienced with NeuroLab took up your question since there would be specific issues along the way. (I'm even wondering if you should start a new topic for this?... since we are talking about NN models instead of linear models.)

But with any NN model, you feed independent variables (In this case, it's the metrics that make up your conditions.) and you get one (or more) optimized dependent variables out. You then use that one (or more) output variable to perform the trade when it reaches a certain threshold.

Somewhere between setting up the problem and running the simulation (backtest) you need to train it. That's the optimization step for the NN, not to be mixed up with the traditional optimization step for WL parameters.