Morning James Lu - yes sounds like you have been curve fitting. I have an idea for you. Run the optimized strategy and another strategy that has a much lower level of optimization and compare the performance of the two. Living something is much more powerful than just reading about it. The other day I came across a really good video educational piece which clearly describes how to approach systematic trading - https://www.google.ch/url?sa=t&rct=...K4FOdmfGRrnPT0EcCT8mNw&bvm=bv.119745492,d.bGg Just sit back and absorb what this guy is saying. Then play it again a few times until the information has sunk in. Just trawl the internet and you will find some good pieces on curve fitting dangers. Regards and good trading
Something which has occupied my mind lately: When making an adaptive formula, where are the lines in the sand to distinguish between improving expectancy vs adding noise, ambiguity or somehow lessening statistical significance? There seems to be limits on how adaptive one should make systems, or is it just a fool's errand to even try?
Maybe first reconsider that some argue that there is no noise in market quotes: Each market quote represents actual, real, transactions; not some random "noise." Then, go from there?
My question is about noise that is added by using adaptive formulas as a small part of a greater trading plan. For instance, a regular SMA reduces noise as a function of its very definition, while an adaptive SMS could increase noise if ie. the period changed wildly every other bar. So if one strives to make something adaptive, there seems to be constraints to usability and how adaptive you can make it without affecting stability and statistical significance too much. Such a formula would be used in conjuction with the rest of the trading plan, so I agree that all the bars and even lesser timeframes could be useful facts. Another way to think about it would perhaps be to run many strategies in parallell and compare their results, making the entire trading plan adaptive instead of just parts of it.
I understand generally what you're saying, I've actually made such a model. I used both: An ensemble of SMA's of varying periods, and a dynamic SMA whose period was based upon a formula, whose terms where derived via an evolutionary algorithm. The formula represented a sort of volatility measurement. The system would flow between whichever sub-model was 'currently' performing best. After some testing, the dynamic SMA outperformed the ensemble of fixed SMA's. But, I don't consider any of the data noise.
Can you explain the statistical properties of kalman filter? I've looked at it before, but it seems hard to understand. Anyways, I'm generally not interested in just an average, but looking to quantify more properties of price action. @userque : When averaging, you effectively filter out smaller "frequency/noise" components. So even though nothing in pricedata is really noise, you're filtering it away anyways when employing averages.