The Autoregressive Integrated Moving Average Model, or ARIMA, is a popular linear model for time series analysis and forecasting.

The statsmodels library provides an implementation of ARIMA for use in Python. ARIMA models can be saved to file for later use in making predictions on new data. There is a bug in the current version of the statsmodelsÂ library that prevents saved models from being loaded.

In this tutorial, you will discover how to diagnose and work around this issue.

Let’s get started.

**Updated Apr/2019**: Updated the link to dataset.

## Daily Female Births Dataset

First, let’s look at a standard time series dataset we can use to understand the problem with the statsmodels ARIMA implementation.

This Daily Female Births dataset describes the number of daily female births in California in 1959.

The units are a count and there are 365 observations. The source of the dataset is credited to Newton (1988).

Download the dataset and place it in your current working directory with the filename “*daily-total-female-births.csv*“.

The code snippet below will load and plot the dataset.

1 2 3 4 5 |
from pandas import Series from matplotlib import pyplot series = Series.from_csv('daily-total-female-births.csv', header=0) series.plot() pyplot.show() |

Running the example loads the dataset as a Pandas Series, then shows a line plot of the data.

### Stop learning Time Series Forecasting the *slow way*!

Take my free 7-day email course and discover how to get started (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

## Python Environment

Confirm you are using the latest version of the statsmodels library.

You can do that by running the script below:

1 2 |
import statsmodels print('statsmodels: %s' % statsmodels.__version__) |

Running the script should produce a result showing statsmodels 0.6 or 0.6.1.

1 |
statsmodels: 0.6.1 |

You can use either Python 2 or 3.

**Update**: I can confirm that the fault still exists in statsmodels 0.8 and results in the error message:

1 |
AttributeError: 'ARIMA' object has no attribute 'dates' |

## ARIMA Model Save Bug

We can easily train an ARIMA model on the Daily Female Births dataset.

The code snippet below trains an ARIMA(1,1,1) on the dataset.

The model.fit() function returns an ARIMAResults object on which we can call save() to save the model to file and load() to later load it.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 |
from pandas import Series from statsmodels.tsa.arima_model import ARIMA from statsmodels.tsa.arima_model import ARIMAResults # load data series = Series.from_csv('daily-total-female-births.csv', header=0) # prepare data X = series.values X = X.astype('float32') # fit model model = ARIMA(X, order=(1,1,1)) model_fit = model.fit() # save model model_fit.save('model.pkl') # load model loaded = ARIMAResults.load('model.pkl') |

Running this example will train the model and save it to file without problem.

An error will be reported when you try to load the model from file.

1 2 3 4 5 6 7 8 |
Traceback (most recent call last): File "...", line 16, in <module> loaded = ARIMAResults.load('model.pkl') File ".../site-packages/statsmodels/base/model.py", line 1529, in load return load_pickle(fname) File ".../site-packages/statsmodels/iolib/smpickle.py", line 41, in load_pickle return cPickle.load(fin) TypeError: __new__() takes at least 3 arguments (1 given) |

Specifically, note the line:

1 |
TypeError: __new__() takes at least 3 arguments (1 given) |

So far so good, so how do we fix it?

## ARIMA Model Save Bug Workaround

Zae Myung Kim discovered this bug in September 2016 and reported the fault.

You can read all about it here:

The bug occurs because a function required by pickle (the library used to serialize Python objects) has not been defined in statsmodels.

A function __getnewargs__ must be defined in the ARIMA model prior to saving that defines the arguments needed to construct the object.

We can work around this issue. The fix involves two things:

- Defining an implementation of theÂ
*__getnewargs__*Â function suitable for the ARIMA object. - Adding the new function to ARIMA.

Thankfully, Zae Myung Kim provided an example of the function in his bug report so we can just use that directly:

1 2 |
def __getnewargs__(self): return ((self.endog),(self.k_lags, self.k_diff, self.k_ma)) |

Python allows us to monkey patch an object, even one from a library like statsmodels.

We can define a new function on an existing object using assignment.

We can do this for the *__getnewargs__* function on the ARIMA object as follows:

1 |
ARIMA.__getnewargs__ = __getnewargs__ |

The complete example of training, saving, and loading an ARIMA model in Python with the monkey patch is listed below:

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 |
from pandas import Series from statsmodels.tsa.arima_model import ARIMA from statsmodels.tsa.arima_model import ARIMAResults # monkey patch around bug in ARIMA class def __getnewargs__(self): return ((self.endog),(self.k_lags, self.k_diff, self.k_ma)) ARIMA.__getnewargs__ = __getnewargs__ # load data series = Series.from_csv('daily-total-female-births.csv', header=0) # prepare data X = series.values X = X.astype('float32') # fit model model = ARIMA(X, order=(1,1,1)) model_fit = model.fit() # save model model_fit.save('model.pkl') # load model loaded = ARIMAResults.load('model.pkl') |

Running the example now successfully loads the model without error.

## Summary

In this post, you discovered how to work around a bug in the statsmodels ARIMA implementation that prevented you from saving and loading an ARIMA model to and from file.

You discovered how to write a monkey patch to work around the bug and how to demonstrate that it has indeed been fixed.

Did you use this workaround on your project?

Let me know about it in the comments below.

Hi Jason, thanks for the workaround.

I am using statsmodels 0.8.0 and have been trying to save a dictionary of fitted models using pickle itself i.e. not using the built-in method. I am still getting the error you mention above about the ARIMA object not having a dates attribute. Why it is trying to look for a

`dates`

attribute anyway?Can you think of a work around for this case?

Did you try the workaround?

Hi Dr. Brownlee,

Thank you for sharing. I tried to adjust this workaround so that it can work for seasonal ARIMA modeling but I am still having issues. I would like to save the model into a pickle file but I keep receiving the following error: “Type error: can’t pickle statsmodels.tsa.statespace._statespace.dStatespace objects”

Do you have an idea of how I can go about working around this issue? Here is some of the code I am working with.

mod = sm.tsa.statespace.SARIMAX(

timeseries,

order=(1, 0, 1),

seasonal_order=(0, 1, 1, 28),

enforce_stationarity=False,

enforce_invertibility=False

)

self.trained_model = mod.fit()

self.trained_model.save(‘call model.p’)

Thanks!

Sorry, I have not tried.

Perhaps reach out to the developer that discovered the workaround, he may have some ideas about SARIMAX.

Hi Veeral,

Were you able to find the fix for your issue? I am also facing the same issue with SARIMA model.

Hey everyone,

I am also having the same problem with SARIMAX models; has anyone found a solution?

Hey Jason,

Thanks for the article.

I am using ARIMA to fit values and save it as a pickle file. Post that, the pickle file is used to get out of sample predictions. However, while getting sample predictions I am getting the following error: Cannot cast ufunc subtract output from dtype(‘float64’) to dtype(‘int64’) with casting rule ‘same_kind’. Do you have any idea about the cause?

I have not seen this sorry.

Perhaps try posting to stackoverflow?

Hey Jason, thanks for the awesome post! I followed this and saved a pickle file with a certain ARIMA order. The model was successfully fit and then saved as pickle file. However, when I load the pickle file to get out of sample forecasts, I get null values for any number of steps. Do you have any idea about this issue?

So basically, after fitting the model, model_fit.summary() throws error. I know that its not a code error solving platform, but I think it is a very general case if model summary is throwing error after fitting the model!

That is odd. Perhaps comment out that line?

Sorry, I have not seen this issue. Perhaps ensure that you are using the most recent version of Statsmodels?

Jason, the problem is that , after reading the pickle file, out of sample forecasts give out null values. I think , this is probably a bug, Will try to report it on Git. Thanks.

That is a shape.

I have some examples in the blog of using the coefficients directly, without the wrapper class. Perhaps that would be a good workaround?

hello Jason,

thanks for your article, i use statsmodels 0.9.0, there is not this kind of bug.

Glad to hear it!

Hello, Thanks for your nice post. But I am confused about how to use the loaded model to predict the feature. For example, I am giving here the code(which is also may be taken from one of your tutorials as I am learning all of these stuff with your resources):

Actual = [x for x in train_set]

Predictions = list()

#Function that calls ARIMA model to fit and forecast the data

def StartARIMAForecasting(Actual, P, D, Q):

model = ARIMA(Actual, order=(P, D, Q))

model_fit = model.fit(disp=0)

prediction = model_fit.forecast()[0]

return prediction

for timepoint in range(len(test_set)):

ActualValue = test_set[timepoint]

#forcast value

Prediction = StartARIMAForecasting(Actual, 2,1,2)

print(‘Actual=%f, Predicted=%f’ % (ActualValue, Prediction))

#add it in the list

Predictions.append(Prediction)

Actual.append(ActualValue)

Now, here, Where I can write the line to save the model and after saving and loading what will be the correct line to do the prediction? I am eagerly waiting to get the solution. đź™‚

N:B: In my dataset, I have used datetime as index.

Good question.

The model is saved after it is fit. It can the be loaded to make a prediction.

Perhaps this will help:

https://machinelearningmastery.com/make-sample-forecasts-arima-python/

Thank you. I will check it now. Another question came in my mind. Is it possible to use ARIMA for multivariate and multistep forecasting? Till now I have only found ARIMA for Univariate Time series.

Yes, it is called VAR or VARIMA, see this post:

https://machinelearningmastery.com/time-series-forecasting-methods-in-python-cheat-sheet/