Curve fitting or optimization: Statistical versus Phenomenological optimization

When we try to optimize using an optimization algorithm are we doing curve fitting and our model is it going to work in the future?

This is a major problem, there are authors that suggest when we test our system for a long period we should not change its parameters anymore. Others prefer to optimize everyday the parameter of the system.

Where is the truth?

I will try to see the things under other perspective. When we try to optimize a system, we are doing basically a statistical optimization. We optimize return on account, maximum winners, Profit/Loss ratio etc. And we optimize statistically with an algorithm the system.

So how robust are the results when they face the real markets. Well, this is tough question.

Sometimes it works, sometimes it does not. That is the reality.

On the other hand not testing a strategy is suicidal. Some authors state that they limit the variables of the system so the over-fitting is less probable concentrating on the most important things like trend direction and support and resistance zones.

Statistical versus Phenomenological optimization

Phenomenological Theory. A theory which expresses mathematically the results of observed phenomena without paying detailed attention to their fundamental significance

With this kind of optimization we choose to have not one system for all market conditions but some systems phenomenologically optimized for a certain condition.

And we need some criteria to switch between the systems. The Fractal dimension graph index is an useful tool in the phenomenological approach.

For example we can use one system when we have antipersitent fractal characteristics and another when the characteristics of the price are persistent. In fact the adaptive robots are a hype, when they know when to stop trading a particular strategy and when they start again.