The key to any succesful indicator is to provide smoothing (prevent whipsaws) and provide timeliness for detecting turning points. As you may have encountered this combination is often difficult to obtain using standard time domain methods like EMA/SMA. Any traditional indicator available is in fact a low-pass (trend like SMA) or combination (mean reversion like RSI) of filters. In my opinion digital filtering in the y beta):frequency domain is less prone to curve-fitting than most time domain machine learning methods (i.e. neural networks, regression, boosting) and is able to replicate any traditional indicator out there with less lag and optimal timeliness. Especially the coming of ever more potent machine learning methods like deep learning are an overfitting nightmare and totally unsuited for trading purposes. The direct filter approach by Wildi is an interesting alternative to traditional methods and has already won some of the most prestigious time series forecasting contests. http://www.neural-forecasting-competition.com/NN3/results.htm The best part is that digital filters can be exported as a set of filter coefficients for trading in traditional software packages like ninjatrader, tradestation and deltix. Convolution filters are not exotic in any way, however the way these are created make them of a totally new breed. I have been trading it with great succes now end-of-day and especially intraday. The key is to find good training periods and the right instrument. Any instrument with periodic movements can provide decent results. The nice part is that the filter characteristics can be tailored to any trading style (risk, frequency, period). To facilitate easy filter extraction I am developing a web application. Acces the application here (early beta): http://162.243.93.70/filter/ Features: -tailor optimal filters for EOD stocks -select in-sample and out-of-sample periods -export filter coefficients ( you could use the http://ninjatrader.com/support/forum/showthread.php?t=12404 indicators here and just replace the coefficients to trade) Upcoming features: -multivariate filter extraction -test against basket of stocks -intraday filter design capabilities -export as indicator to popular trading packages (Ninjatrader) -additional anti-overfit measures (the biggest enemy of traders) -replicate any imported indicator (but of course faster and probably more profitable) In the upcoming thread I will provide details on the technical aspects of signal extraction, how to prevent overfitting (the most essential part for any quantitative trader) and how to tailor the filter responses to your needs (i.e. high frequency versus low-frequency trading desire). Also I will provide source code on how to test/trade in different software packages. We will start with R and ninjatrader but please provide suggestions for other platforms. Let me know if you have any suggestions/questions, I hope you are excited to hear about a novel trading approach.
Jack Hershey is the only one that discovered how to see a turn before it even started but alas, he's not with us any more so the machines have some work to do. Maybe some AI computer can decode all his acronyms at some point. He made acronyms, then made abbreviations for them so it's going to take some serious computer power...
I hope that you show us using simple example how to generate coefficients for this filter. What happens if the instrument in question is not periodic or changes its characteristics?