SARIMAX fitting the model, problems with summary

[Shcherbakova Anastasiya]

#I am fitting the SARIMAX model without exogenous variables. My data looks like this and do not contain any missing values:

y.head(5)

DateId
2017-09-08      139.0
2017-09-09      232.0
2017-09-10    10129.0
2017-09-11    61154.0
2017-09-12    22093.0
Name: NumberOfViewers, dtype: float32

#checking if any of the values are null
y.isnull().any().any()

False

#Further after some exploration, I consider to use SARIMAX with freq=7 and perform grid search to find the the best order parameters

import warnings
import itertools
import statsmodels.api as sm

# Define the p, d and q parameters to take any value between 0 and 5
p = d = q = range(0, 5)

# Generate all different combinations of p, q and q triplets
pdq = list(itertools.product(p, d, q))

# Generate all different combinations of seasonal p, q and q triplets
seasonal_pdq = [(x[0], x[1], x[2], 7) for x in list(itertools.product(p, d, q))]

warnings.filterwarnings("ignore") # specify to ignore warning messages
param_list = []
param_seasonal_list = []
results_aic_list = []

for param in pdq:
    for param_seasonal in seasonal_pdq:
        try:
            mod = sm.tsa.statespace.SARIMAX(y[: round(0.7 * len(index_))],
                                            order=param,
                                            seasonal_order=param_seasonal,
                                            enforce_stationarity=False,
                                            enforce_invertibility=False)
                                            #simple_differencing=True)

            results = mod.fit()
            param_list.append(param)
            param_seasonal_list.append(param_seasonal)
            results_aic_list.append(results.aic)

            print('ARIMA{}x{} - AIC:{}'.format(param, param_seasonal, results.aic))
        except:
            continue

# then for the minimal AIC i pick up the model parameters, estimate the model

order_best = param_list[results_aic_list.index(min(results_aic_list))]
seasonal_best = param_seasonal_list[results_aic_list.index(min(results_aic_list))]

mod = sm.tsa.statespace.SARIMAX(y[: round(0.7 * len(index_))],
                                order=order_best,
                                seasonal_order=seasonal_best,
                                enforce_stationarity=False,
                                enforce_invertibility=False)
                                #simple_differencing=True)

fit_results = mod.fit(disp=False)
fit_results.params


#and here the problem starts. I can reach the  fit_results parameters, bic, aic coefficients, however I cannot print summary or explore the residuals with the f#it_results.plot_diagnostics(). For the summary() i get the following message "ValueError: On entry to DLASCL parameter number 5 had an illegal value". 
# I tried to play around  with different methods  in fit() as it was offered to try in here but that did not help. Where  is my possible mistake could be or is it a bug? 

#BTW, the range for the grid search i specify in (0, 5), but the actual results i get only for (0, 4) 

[Chad Fulton]

Thanks for letting us know! If you can, please file a bug report on the Github Issue tracker (e.g. by visiting https://github.com/statsmodels/statsmodels/issues/new)

Please also give the full error message - the ValueError should give you a trace that will be required for us to help narrow down the problem.

Additionally, if your dataset isn't confidential, it would be helpful if you could post it so that we can replicate the problem.

Best,

Chad

[josef...@gmail.com]

an additional detail

" On entry to DLASCL parameter number 5 had an illegal value"

usually is caused by either nan or inf in the linear algebra computation. Did you check that you don't have nans in the data?

Given that the estimated params are available, it is more likely that there are some nans produced in one of the results.

Josef

[Shcherbakova Anastasiya]

That was the first thing for which i've checked: there are no  missing values in the data.