The code predicts Ethereum prices for next day, 15 days and 30 days in future. It utilizes EMD, EEMD and CEEMDAN family of data decomposition algorithms. The input time series is broken down into various Intrinsic mode functions and a residue. The residue is further broken iteratively until no pattern is left. Finally, the aggregated sum of all data is taken. The resultant data exhibits an improved noise-to-signal ratio and results into a better forecast when integrated with any deep learning model such as LSTM, CNN, CNN-BiLSTM or Supervised model.
An advancement on the EEMD method, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) allows for a granular spectral separation of the Intrinsic Mode Functions and a more precise reconstruction of the original signal (IMFs). I utilize CEEMDAN and LSTM to forecast Ethereum prices multiple steps ahead.
The novelty in this particular scenario is that the residue has been iteratively deconstructed using CEEMDAN until there is no pattern remaining in the residue. The residue is typically ignored by authors in scientific publications, losing valuable data. In order to get rid of the residue, I had to split it down into smaller parts. This is a novel concept, and it is the responsibility of future researchers to incorporate it into their own work. Please provide a citation to this repository in your work.
Figures 1 and 2 illustrate the decomposition of Ethereum time series. The residue depicted in figure 1 is utilized as an input in Figure 2. The residue is iteratively decomposed at multiple levels. However in case of Ethereum, I am only including the first decomposition of the residue which removes most of the noise and results into a smooth pattern. Using the preceding iterative process, the residue can be reduced down to the point where no pattern remains.
