Discussion of article "Introduction to the Empirical Mode Decomposition Method" - page 5

New comment
 
MisterH:
Actually, this is not a good article. EMD is not a causal technique. This means that its past values change in real time, rendering it utterly and completely useless for trading. It's in the same category as Singular Spectrum Analysis, the Hodrick-Prescott filter and all types of splines. It looks o so good on a static chart, but in real time it is no better than a LWMA. Just place a SMA(1) on the result of your EMD line and you will see how bumpy it becomes... Nice from a research/scientific point of view, useless in trading.
If you are trying to use EMD (or probably any other analytical technique) as some sort of simple predictive price filter based on historical price data then I would agree it is pretty useless but I would not be quite so quick to totally dismiss the technique outright. There are various other ways in which decomposing non stationary data into component wave forms can  be useful and informative. In my experience EMD does a pretty good job of this
 

Hey everyone,

i am struggling to form a logical path to implement the EMD technique together with SVM-regressions. Most papers i read about (E)EMD-SVM (e.g. "Short-term prediction of stock index based on EMD and SVMs") decompose the complete time series first before implementing the SVM learning path.

But i noticed that if i add one additional dataset (t+1) to the time series, the EMD algorithm changes almost every single IMF value (even the number of IMF can change too (for the same date in the past)) than it was before.

Therefore, i am concerned that if i split my data set into a learning period (e.g. 2002-2010) and want to make out-of-sample forecasts (e.g. 2011) my EMD decomposed IMFs should only contain data from 2002-2010 to predict 2011, right? Predcting 2011 with IMF-time-series calculated with the EMD data set (2002-2011) would incorporate information from the "future" making my backtesting results not valid, right?

So for every one-step forward prediction my EMD must be calculated with the additional data points ... then the SVM-regressions can be performed to backtest such a model, right? This recursive method could be "BUMPY" as the MisterH mentioned above, making it useless for backtesting/trading strategy? 

 

Can we have a word about the "slightly different implementation of EMD" in the second enclosed includnik . pros cons cons cons differences

big + for the article

12345
New comment