I am interested to do time series forecasting (1.png) with artificial neural networks. In order to decrease data complexity I have decided to remove the linear component (2.png). Also I did some Fourier analysis and I have removed some sinusoids (3.png).
My question is - do you have other ideas, which can be applied in order data complexity to be reduced?
What other kinds of preprocessing are you using in time series forecasting?
Here's an important article that shows the typical errors in timeseries prediction: https://towardsdatascience.com/how-not-to-use-machine-learning-for-time-series-forecasting-avoiding-the-pitfalls-19f9d7adf424
No. Of inbuilt algo library is available in R and python.
You can try there, You must able to get the desire results.
Have you used MATLAB for the same ?
Hi, Dr. Balabanov,
If time series methods fail to satisfy you, perhaps Markov models will give you a more satisfactory model. I've been experimenting with data sets which are as chaotic as the third (lower right-hand side of your post) series you posted, and optimal Markov analysis has identified some very strong results. Here are two examples:
- https://odajournal.com/2019/02/18/optimal-markov-model-relating-two-time-lagged-outcomes
- https://odajournal.com/2019/02/07/confirming-the-efficacy-of-weighting-in-optimal-markov-analysis-modeling-serial-symptom-ratings
Time series seems to be non periodic. So fourier analysis is unnecessary. Some normalisation like min-max may be tried before removal of the trend.
The standardization that Maximilian speaks of is called "ipsative standardization". This standardization is crucial in most serial designs, except (it seems) for Markov.
Ipsative standardization works when multiple series are to be combined, so long as each series is standardized separately:
- https://odajournal.com/2013/10/23/ipsative-standardization-is-essential-in-the-analysis-of-serial-data
Of course, ipsative standardization works well when there is only one series.
- https://odajournal.com/2013/11/07/ascertaining-an-individual-patients-symptom-dominance-hierarchy-analysis-of-raw-longitudinal-data-induces-simpsons-paradox
However, even ipsatively standardized data are susceptible to paradoxical confounding--something to look out for.
- https://odajournal.com/2013/09/19/the-use-of-unconfounded-climatic-data-improves-atmospheric-prediction
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series