lavinei / pybats Goto Github PK

Dear Isaac,

First off, I'm so glad this package exists and it has lots of cool features. I am currently looking at univariate count data with lots of zeroes and wanted to compare the Poisson DGLM forecasting performance with that of a DCMM. However, it seems that currently the analysis_dcmm function requires some input for the covariates X, and inputting None (like the default in the regular analysis function) throws an error of 'NoneType object not subscriptable'. I was able to 'monkey patch' a few of the functions in the package to get it to work for my case, but I think this is a capability that would probably be desirable for others.

I can create a reproducible example of the error if needed! Thanks!

Sincerely,
Brian

It would be helpful if a reference to, or how, the define_glm_params (and the others) get their parameters.

For large prior_length the linear regression looks natural and the small size correction g/(1+g) is very close to 1, but I am not sure why g=int(n/2) (t-student moments suggest g=n).

Also, the returned "n" is - except the case of degeneracy in regressors - the number of degrees of freedom (number of observations minus the number of covariates) in the linear regression, while a naïve approach, applying a constant DLM model suggests the number of observations or that plus one.

If the reference are given, probably these objections are responded there.

Getting a ValueError: lam value too large when building model using poisson distribution.

I have outlined the details in this SO question: https://stackoverflow.com/questions/68073589/pybats-forecast-error-when-using-poisson-family-distribution

Any assistance would be greatly appreciated! Thanks

in other words, estimating X and call forecast modules for point and path forecasting.

It seems that there is some issue with the plots. When I run the 1-day example in the docs, I found the graph attached. I think that the forecast and observation labels have been swapped (wrongly), isn't it? It is like the confindence interval has been plot for the observations.

In the docs the labels seem to be properly placed. What is the source of this difference? I am running exactly the same code.

Regards

Is it possible to use pybats to forecast with only univariate time series data? It seems like X needs to be populated when making a model. I was thinking that you could just put the time index or just put a lagged version of Y in for X.

Separately, I saw other questions asking how to forecast for Y_{t+k} given only information of X_{t}. It wasn't clear to me how you could do this without keeping a separate forecaster for X. It would be nice to be able to make forecasts without data until the next observation (usually when you get access to both X_{t+k} and Y_{t+k}.

When tested with the examples it works beautifully but whenever asked to forecast more data than available it doesn't do anything.

Am I missing something or should i use the generated model for that.

I'm unable to generate a forecast_path on a model with holiday regressors. The example below uses slightly modified code fro the order forecast example.

test_df = data.loc[forecast_start:forecast_end]

model = analysis(
    train_df.Sales.values, 
    X=None, 
    family='poisson',
    seasPeriods=[7], 
    seasHarmComponents=[[1,2,3]],
    prior_length=prior_length, 
    dates=train_df.index, 
    holidays=USFederalHolidayCalendar.rules,
    rho=rho,
    ret = ['model']
)

model.forecast_path(k=10)

This results in the following error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-14-f1091b7e298d> in <module>
     15 )
     16 
---> 17 model.forecast_path(k=10)

/jupyter/packages/pybats/dglm.py in forecast_path(self, k, X, nsamps, copula, phi_mu, phi_sigma, phi_psi, analytic, phi_samps, **kwargs)
    287         else:
    288             if copula:
--> 289                 return forecast_path_copula(self, k, X, nsamps, **kwargs)
    290             else:
    291                 return forecast_path(self, k, X, nsamps)

/jupyter/packages/pybats/forecast.py in forecast_path_copula(mod, k, X, nsamps, t_dist, y, nu, return_cov)
    185         # Plug in the correct F values
    186         if mod.nregn > 0:
--> 187             F = update_F(mod, X[i, :], F=F)
    188         # if mod.nregn > 0:
    189         #     F[mod.iregn] = X[i,:].reshape(mod.nregn,1)

TypeError: 'NoneType' object is not subscriptable

Hello, Lavinei,

I've decided to apply your approach to my own data and had no success with the error above:

/usr/local/lib/python3.6/dist-packages/pybats/dglm.py in simulate_from_sampling_model(self, p, nsamps)
415
416 def simulate_from_sampling_model(self, p, nsamps):
--> 417 return np.random.binomial(1, p, [nsamps])
418
419 def simulate_from_prior(self, alpha, beta, nsamps):

mtrand.pyx in numpy.random.mtrand.RandomState.binomial()

_common.pyx in numpy.random._common.check_array_constraint()

ValueError: p < 0, p > 1 or p contains NaNs

Link to the code: https://colab.research.google.com/drive/1_5LHVy1_EHRoohPhk8Dk26xTnVLooCRS?usp=sharing
I could send you the data by email, if it is possible.

Thanks!

Regards,
Nikita

Can you add negative binomial distribution in addition of poisson ?
Here, some

https://minimizeregret.com/post/2018/01/04/understanding-the-negative-binomial-distribution/

https://kth.diva-portal.org/smash/get/diva2:1249681/FULLTEXT01.pdf

thanks