In the time series prediction, both lstm and sfm have good performance. I don’t know if pymc can consider adding these two time-predicted models to the framework.
In the process of studying stock forecasting, I generated a lot of useful models based on sfm, and I think that lstm’s architecture has a lot of room for improvement. Pymc’s time series forecasting method should contain two methods.I want to implement this idea in GSoC 2019. The model code for sfm is here (https://github.com/z331565360/State-Frequency-Memory-stock-prediction), and interestingly this code is also based on theano.
It depends on how you want to go about to use it, I see there are 2 possible ways, both are a bit challenging:
- Set up LSTM or SFM neural net as an approximation for autoencoding variational Bayes, similar idea: https://docs.pymc.io/notebooks/lda-advi-aevb.html. The challenge is to make sure the parameter clone is performed correctly, and you need a good logp function for the observed with the LSTM or SFM parameters are input.
- Use LSTM or SFM layer in a PyMC3 model. The challenge here is that Keras recurrent layer might not work out of the box, and inference could be extremely difficult.
Thank you very much for your reply, and I have benefited from the questions you have raised. I can use theano to implement lstm and sfm, and do not need to use keras. The sfm is actually a lstm process with the addition of Fourier transform. But how to include the API and how it works in pymc3 is really worth pondering. I will continue to study, I will inform you if there is any progress.