Comparison between HTM and other methods for Time Series predictions

question
community

#1

Hi,

I have wrote my own implementation of HTM system and i think i might be ready to make a comparison with existing methods. What methods should i check for that comparison? Are any of them available in MATLAB toolbox, or in R?


#2

That’s great! There are several possible techniques you can compare with. The classic one is ARIMA, which is available in R (and probably Matlab). You could also compare against TDNN or LSTM. See our paper [1] for more details. The code we used to implement the simulations is available here:

You’ll probably want to look into this subdirectory (includes R ARIMA scripts):

–Subutai

[1] Cui, Y., Ahmad, S., & Hawkins, J. (2016) Continuous online sequence learning with an unsupervised neural network model. Neural Comput., 28, 2474–2504.


#3

I got a problem predicting 5 steps ahead in hotgym comparing with Arima model (for 5 steps ahead my implementation is closely to ARIMA results, and its quite bad).
What might have gone wrong?

For one step ahead i have :


Where MAPE metric for my HTM system is 40% higher than ARIMA.

While for 5 steps ahead:


MAPE is 5% higher.

I use 950 mini-columns, and 12 cells per column(because of system limitations).
Also i use same SDR classifier(with softmax function) with only change SDR length(column size = prediction steps).
I would appreciate any tip, or direction.


#4

Are these two different models? If so, where do the model params differ?


#5

Model’s parameters are same in both cases. But for 5 steps ahead i get these results. The only difference is as i mentioned above for SDR classifier. Is there any code-pseudocode for SDR classifier (and decoder) procedure? Maybe i am missing something, when prediction steps become > 1?


#6

Do you need to change the SDR classifier at all? What happens when you don’t?


#7

I got same bad results… Where should i focus for this issue?