tingiskhan / pyfilter Goto Github PK

The proposals Linearized, ModeFinding and Unscented are not using the "correct" pre-weighting. Fix this

pytorch is still missing searchsorted functionality, and in order to use the GPU efficiently we either need to implement one of our own, or find an existing solution.

Could you please share which GPU you are using when you run and save the example notebooks?

According to the comments in older versions you were using a GTX 1070ti. I have just run the SMC2 fit in stochastic-volatility.ipynb with a Titan V and my result is actually slower than that in the notebook in master branch at ~9it/s vs ~10it/s.

The Titan should be 2x faster than the 1070ti in fp32 so I hope you have upgraded and I am comparing against a different GPU!

Implement a better proposal for log-concave observation models, e.g. stochastic volatility models, using a linearized proposal

Look at using the torch.nn module in order to save/load models.

All of the Notebooks have not been updated to reflect the new functionality of the library. Also add example of consistency of algorithms

Hello Khan,

I wanted to tell you that I still think that this is an amazing package. I was suprised to see that you could make some updates, even so you became a father a short time ago.

Are there plans to add a structural time series like in Tensorflow Probability as well?

https://www.tensorflow.org/probability/examples/Structural_Time_Series_Modeling_Case_Studies_Atmospheric_CO2_and_Electricity_Demand

Is it possible to make predictions with the lynx hare example with VI, or the nutria example? And how could the nutria example be used for other data like weather?

And one more question - would it make sense to add a AR component-parameter as well to the stochastic volatility example to improve the prediction? Or even better add a parameter for resistance like in the SVL example -

https://github.com/mfrdixon/ML_Finance_Codes/blob/master/Chapter7-Probabilistic-Sequence-Modeling/ML_in_Finance-Kalman_Filters.ipynb

rho = pm.Uniform(name='rho', lower=-1., upper=1.)

I am still studying all this and try to improve my skills, and hope to make some contribution to the pyfilter as well in the future. Because I can say that I like it a lot.

Sometimes there is an error relating to the resampling where the resampling scheme generates an index that is 1 greater than the length of the array.

The optimal proposal distribution for 1D linear observation models seems to work, wheras the one for n-D does not work.

Nested filters (like SMC2 and NESS) don't seem to work for n-D models as well as Unscented proposal/filter.

Update the Notebooks

Consider using an empty dimension for state space models of only 1 dimension, such that we could remove the special case of 1 dimensional models

Hi tingiskhan,

thanks for your effort.
I tried to run the stochastic-volatility.ipynb, but I am getting an error

student_t = DistributionWrapper(StudentT, df=Prior(Exponential, rate=0.1))
obs = AffineObservations((go, fo), (Prior(Normal, loc=0.0, scale=1.0)), student_t)

Error message:
raise ValueError("The parameter {} has invalid values".format(param))
ValueError: The parameter df has invalid values

Could you be so kind to explain me what I did wrong? Thanks again

Hello Khan! How are you?
I came across your package and I think it is a really great thing.
I wanted to install the package with pip but somehow it was not really working out ....
In the end I got it running by simply copying the folder into site-packages and now it runs like a charm it seems.

Right now I run it with cpu but it should be possible to install pytorch cuda.

My question is - do you use the program with success for live trading?
Are you doing normal trend following - or some kind of pairs trading?
Did you also try it with normal close price?
Are you always forecasting a whole week - or is it possible to use the program with 1 step as well?

Best regards to Sweden - Stockholm is a great city :-D