How to Grid Search ARIMA Model Hyperparameters with Python

The ARIMA model for time series analysis and forecasting can be tricky to configure.

There are 3 parameters that require estimation by iterative trial and error from reviewing diagnostic plots and using 40-year-old heuristic rules.

We can automate the process of evaluating a large number of hyperparameters for the ARIMA model by using a grid search procedure.

In this tutorial, you will discover how to tune the ARIMA model using a grid search of hyperparameters in Python.

After completing this tutorial, you will know:

  • A general procedure that you can use to tune the ARIMA hyperparameters for a rolling one-step forecast.
  • How to apply ARIMA hyperparameter optimization on a standard univariate time series dataset.
  • Ideas for extending the procedure for more elaborate and robust models.

Let’s get started.

How to Grid Search ARIMA Model Hyperparameters with Python

How to Grid Search ARIMA Model Hyperparameters with Python
Photo by Alpha, some rights reserved.

Grid Searching Method

Diagnostic plots of the time series can be used along with heuristic rules to determine the hyperparameters of the ARIMA model.

These are good in most, but perhaps not all, situations.

We can automate the process of training and evaluating ARIMA models on different combinations of model hyperparameters. In machine learning this is called a grid search or model tuning.

In this tutorial, we will develop a method to grid search ARIMA hyperparameters for a one-step rolling forecast.

The approach is broken down into two parts:

  1. Evaluate an ARIMA model.
  2. Evaluate sets of ARIMA parameters.

The code in this tutorial makes use of the scikit-learn, Pandas, and the statsmodels Python libraries.

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.

Start Your FREE Mini-Course Now!

1. Evaluate ARIMA Model

We can evaluate an ARIMA model by preparing it on a training dataset and evaluating predictions on a test dataset.

This approach involves the following steps:

  1. Split the dataset into training and test sets.
  2. Walk the time steps in the test dataset.
    1. Train an ARIMA model.
    2. Make a one-step prediction.
    3. Store prediction; get and store actual observation.
  3. Calculate error score for predictions compared to expected values.

We can implement this in Python as a new standalone function called evaluate_arima_model() that takes a time series dataset as input as well as a tuple with the p, d, and q parameters for the model to be evaluated.

The dataset is split in two: 66% for the initial training dataset and the remaining 34% for the test dataset.

Each time step of the test set is iterated. Just one iteration provides a model that you could use to make predictions on new data. The iterative approach allows a new ARIMA model to be trained each time step.

A prediction is made each iteration and stored in a list. This is so that at the end of the test set, all predictions can be compared to the list of expected values and an error score calculated. In this case, a mean squared error score is calculated and returned.

The complete function is listed below.

Now that we know how to evaluate one set of ARIMA hyperparameters, let’s see how we can call this function repeatedly for a grid of parameters to evaluate.

2. Iterate ARIMA Parameters

Evaluating a suite of parameters is relatively straightforward.

The user must specify a grid of p, d, and q ARIMA parameters to iterate. A model is created for each parameter and its performance evaluated by calling the evaluate_arima_model() function described in the previous section.

The function must keep track of the lowest error score observed and the configuration that caused it. This can be summarized at the end of the function with a print to standard out.

We can implement this function called evaluate_models() as a series of four loops.

There are two additional considerations. The first is to ensure the input data are floating point values (as opposed to integers or strings), as this can cause the ARIMA procedure to fail.

Second, the statsmodels ARIMA procedure internally uses numerical optimization procedures to find a set of coefficients for the model. These procedures can fail, which in turn can throw an exception. We must catch these exceptions and skip those configurations that cause a problem. This happens more often then you would think.

Additionally, it is recommended that warnings be ignored for this code to avoid a lot of noise from running the procedure. This can be done as follows:

Finally, even with all of these protections, the underlying C and Fortran libraries may still report warnings to standard out, such as:

These have been removed from the results reported in this tutorial for brevity.

The complete procedure for evaluating a grid of ARIMA hyperparameters is listed below.

Now that we have a procedure to grid search ARIMA hyperparameters, let’s test the procedure on two univariate time series problems.

We will start with the Shampoo Sales dataset.

Shampoo Sales Case Study

The Shampoo Sales dataset describes the monthly number of sales of shampoo over a 3-year period.

The units are a sales count and there are 36 observations. The original dataset is credited to Makridakis, Wheelwright, and Hyndman (1998).

Learn more about the dataset from here.

Download the dataset and place it into your current working directory with the filename “shampoo-sales.csv“.

The timestamps in the time series do not contain an absolute year component. We can use a custom date-parsing function when loading the data and baseline the year from 1900, as follows:

Once loaded, we can specify a site of p, d, and q values to search and pass them to the evaluate_models() function.

We will try a suite of lag values (p) and just a few difference iterations (d) and residual error lag values (q).

Putting this all together with the generic procedures defined in the previous section, we can grid search ARIMA hyperparameters in the Shampoo Sales dataset.

The complete code example is listed below.

Running the example prints the ARIMA parameters and MSE for each successful evaluation completed.

The best parameters of ARIMA(4, 2, 1) are reported at the end of the run with a mean squared error of 4,694.873.

Daily Female Births Case Study

The 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).

Learn more about the dataset here.

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

This dataset can be easily loaded directly as a Pandas Series.

To keep things simple, we will explore the same grid of ARIMA hyperparameters as in the previous section.

Putting this all together, we can grid search ARIMA parameters on the Daily Female Births dataset. The complete code listing is provided below.

Running the example prints the ARIMA parameters and mean squared error for each configuration successfully evaluated.

The best mean parameters are reported as ARIMA(6, 1, 0) with a mean squared error of 53.187.

Extensions

The grid search method used in this tutorial is simple and can easily be extended.

This section lists some ideas to extend the approach you may wish to explore.

  • Seed Grid. The classical diagnostic tools of ACF and PACF plots can still be used with the results used to seed the grid of ARIMA parameters to search.
  • Alternate Measures. The search seeks to optimize the out-of-sample mean squared error. This could be changed to another out-of-sample statistic, an in-sample statistic, such as AIC or BIC, or some combination of the two. You can choose a metric that is most meaningful on your project.
  • Residual Diagnostics. Statistics can automatically be calculated on the residual forecast errors to provide an additional indication of the quality of the fit. Examples include statistical tests for whether the distribution of residuals is Gaussian and whether there is an autocorrelation in the residuals.
  • Update Model. The ARIMA model is created from scratch for each one-step forecast. With careful inspection of the API, it may be possible to update the internal data of the model with new observations rather than recreating it from scratch.
  • Preconditions. The ARIMA model can make assumptions about the time series dataset, such as normality and stationarity. These could be checked and a warning raised for a given of a dataset prior to a given model being trained.

Summary

In this tutorial, you discovered how to grid search the hyperparameters for the ARIMA model in Python.

Specifically, you learned:

  • A procedure that you can use to grid search ARIMA hyperparameters for a one-step rolling forecast.
  • How to apply ARIMA hyperparameters tuning on standard univariate time series datasets.
  • Ideas on how to further improve grid searching of ARIMA hyperparameters.

Now it’s your turn.

Try this procedure on your favorite time series dataset. What results did you get?
Report your results in the comments below.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.

Want to Develop Time Series Forecasts with Python?

Introduction to Time Series Forecasting With Python

Develop Your Own Forecasts in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Introduction to Time Series Forecasting With Python

It covers self-study tutorials and end-to-end projects on topics like:
Loading data, visualization, modeling, algorithm tuning, and much more...

Finally Bring Time Series Forecasting to
Your Own Projects

Skip the Academics. Just Results.

Click to learn more.

80 Responses to How to Grid Search ARIMA Model Hyperparameters with Python

  1. Gerrit Govaerts January 18, 2017 at 9:01 pm #

    If you are willing to consider an R solution , then I can point you to the function auto.arima in the R package ‘forecast’ : https://cran.r-project.org/web/packages/forecast/forecast.pdf
    This will do all the gridsearch you need without writing a single line of code .
    Now , in general , the use of gridsearch for solving the hyperparameters optimization problem in machine learning models is a poor inefficient choice . It has been proven that random search is faster and Bayesian search is even faster . See this : https://www.youtube.com/watch?v=cWQDeB9WqvU (lecture by Geoff Hinton) . For Python , there is a package called hyperopt that provides this functionality : https://github.com/hyperopt/hyperopt
    An intro to hyperopt is here : https://www.youtube.com/watch?v=Mp1xnPfE4PY

    • Jason Brownlee January 19, 2017 at 7:34 am #

      Thanks for the links Gerrit.

      A noted difference is the optimizaiton of an out of sample statistics, i.e. test performance.

      Re grid vs random search, the ARIMA grid is small enough that it can be enumerated. When working with small grids with low compute times, random search would be much less efficient.

  2. Abdallah January 31, 2017 at 1:49 pm #

    hello I have used the evaluate model function to chose the best configuration, but it skipped those configurations that I expect the best according to the Box-Jenkins Method. what that means? and is there any way to check that configurations?

    • Jason Brownlee February 1, 2017 at 10:44 am #

      Great question Abdallah, I am frustrated by this as well.

      I believe you may be able to tinker with the ARIMA configuration further, such as configuring it use or not use a trend constant.

      The issue is caused by instabilities in the linalg and optimization libraries used under the covers.

      You could try an alternate implementation (R?), try implementing the method from scratch by hand or perhaps try fitting a linear regression model on a version of the dataset transformed using the same ARIMA operations.

      Does that help?

  3. Andres Kull February 8, 2017 at 11:47 pm #

    You are doing here one-step rolling forecast for tuning ARIMA parameters. Will the resulting model behave best for forecasting the next observation only? Let’s assume that I would like to get the best possible prediction for the period of next 30 observations. Should the parameters tuning be changed for 30 steps rolling forecast in this case?

    • Jason Brownlee February 9, 2017 at 7:25 am #

      Yes Andres, spot on. It is critically important to optimize for the outcome you require.

  4. Stuart Farmer May 24, 2017 at 1:04 am #

    Amazing stuff here, man. Love it. Keep up the good work!

  5. Arkojyoti June 6, 2017 at 4:30 am #

    Hi Jason,

    Thanks for the post. However, I encountered problems while trying to parse the date column in the Shampoo Sales example. I had downloaded the data from the following link(csv format) and am using Python 3:

    https://datamarket.com/data/set/22r0/sales-of-shampoo-over-a-three-year-period#!ds=22r0&display=line

    I faced 2 problems during parsing:
    1) The format of Month column was “1-Jan”, indicating that we need to specify “%Y-%b” instead of “%Y-%m”
    2) For values >9, that is , 10-Jan, 11-Jan and so on, the parsed date will be rendered invalid. Since it will be in the format : “19010-Jan” and similar

    Please find the modified function which worked for me:

    def parser(x):
    #the following code chunk will take care of parsing for two conditions:
    #1. for dates 10
    test = int(x.split(‘-‘)[0])
    #print(test)
    if(test < 10):
    return(datetime.strptime("190"+str(x),"%Y-%b"))
    else:
    return(datetime.strptime("19"+str(x),"%Y-%b"))
    series = read_csv('sales-of-shampoo-over-a-three-ye.csv', header=0, parse_dates=[0], index_col=0,
    squeeze=True, date_parser=parser)

    Please correct me if there is a mistake in the approach. Hope this helps. Thanks again for the article. Have a good day 🙂

    • Jason Brownlee June 6, 2017 at 10:08 am #

      I have tested and confirm that the example works in Python3.

      Perhaps confirm that you have the same dataset, that you have removed the footer from the file, and that you have copied the code from the post exactly?

  6. Hans June 20, 2017 at 3:00 am #

    On my computer the first example script breaks with:

    ** On entry to DLASCL, parameter number 4 had an illegal value

    so I get no best settings.

    The second script breaks with “Best ArimaNone MSE=inf”

    I have already removed the footer line. Any hints available?

  7. TaeWoo Kim June 23, 2017 at 3:10 am #

    Hey Jason

    What is the difference (or benefit) of doing the grid search this way vs. using SARIMAX? (reference: https://www.digitalocean.com/community/tutorials/a-guide-to-time-series-forecasting-with-arima-in-python-3)

    • Jason Brownlee June 23, 2017 at 6:48 am #

      I have not read that post, but skimming it suggests that are using a for loop just the same as in my tutorial.

  8. Priya Srinivasan July 4, 2017 at 2:28 am #

    “Each time step of the test set is iterated. Just one iteration provides a model that you could use to make predictions on new data. The iterative approach allows a new ARIMA model to be trained each time step.”

    First of all, thank you for this tutorial ! I am a bit confused about using your iterative approach above. My questions are:

    1. Why are you adding the test example to the training set (in history) and retraining the ARIMA model ? This way each subsequent test prediction is trained on the original training set plus an element added from the prior test example. Is this to improve the test predictions by adding more training data to the model (which now includes original training + test examples )?

    2. Using the predict function, can I just train an ARIMA on the training set and use the in-built predict function on the test example set aside ? What are the pitfalls using this approach ?

    Thank you again !

  9. Sam July 14, 2017 at 6:44 am #

    What does this error mean – Best ARIMANone RMSE=inf?

    • Jason Brownlee July 14, 2017 at 8:37 am #

      No good result found Sam. Did you run the code as-is or adapt it to your problem? Perhaps debug the example?

      • Kailash April 23, 2018 at 10:06 pm #

        First let me thank you for this awsome blog. Coming to my issue, I ran this code with my own dataset. I’m getting same error.After that I tried with same dataset too.Still getting same error.Please help sir.

        • Jason Brownlee April 24, 2018 at 6:34 am #

          Thanks.

          What error are you getting with the exact code in the tutorial?

        • jay November 26, 2018 at 9:38 am #

          Hi

          Were you able to fix the problem?

  10. Marianico July 20, 2017 at 4:32 pm #

    Can the amount of input data affect to the forecast? I mean, maybe the oldest lagged data is not quite correlated with the current one. If so, wouldn’t be better to limit the length of history to 500 rows, for instance? How do I find the optimal amount of training data?

  11. Andrew September 9, 2017 at 3:26 am #

    This model is taking forever to load – is there something I can do to optimize performance?

  12. Udi November 2, 2017 at 8:10 am #

    Hi.
    I’m trying to fit an ARIMA model to a financial dataset and getting really poor results. The predicted trend vary widely from the real one and relative MSE is an order of magnitude higher than required. I’m having difficulties in recognizing the source of my problem.
    I have several questions. Please tell me if my questions are too extensive for this discussion:

    1. How do you choose the range for the grid search? ACF and PACF graphs hinted that the order of p,q is 0,0, but the relevant models still gave poor results (I’ve used aic as my error score, and the most popular model in my runs is ARIMA(0,0,0)…)

    2. Another suspect is my fit-predict procedure. I fitted a model for an X hours period and predicted the next y minutes (repeat ad-infinitum). Surprisingly, using a small x value just gave worse results, not better, although the data is highly volatile.

    3. Can you prove that an ARIMA model would be a poor choice for forecasting for a certain dataset? I mean, a dataset which is already stationary (differentiated), so a Dickey-Fuller test yields good results.

    • Jason Brownlee November 2, 2017 at 3:56 pm #

      Looking at ACF/PACF plots can give you an idea of values to try for q and a:
      https://machinelearningmastery.com/gentle-introduction-autocorrelation-partial-autocorrelation/

      Your testing procedure sounds like walk-forward validation, which is what I would recommend:
      https://machinelearningmastery.com/backtest-machine-learning-models-time-series-forecasting/

      Yes, if the results are bad, move on to other algorithms or more data preparation. I’d recommend both to help sniff out opportunities for your data.

      • Udi November 2, 2017 at 5:36 pm #

        Thank you for your quick answers!

        – I’ve looked at ACF/PACF, and they suggested p=q=0, which is indeed my chosen model according to minimal aic criterion but is still not good enough. Maybe bic/MSE error scores could yield a different result, although I’m not optimistic.

        – I haven’t read the other post yet, and I have to admit I haven’t fully realized all the nitty-gritty details of your code here, but I don’t think my algorithm fits walk-forward validation, because I fit a new model each time for a fixed sized group of data points. It’s just the fixed window that moves forward. This is mainly used for statistics sake, not for improvement of the outcome model.

        – Are there other useful data preparation procedures besides smoothening?

  13. Udi November 2, 2017 at 8:13 am #

    Ah, and the last one (for now…):

    4. Could data smoothening procedures improve the performance of an ARIMA model? It seems to me that if the data is naturally volatile any smoothening is bound to mess up the prediction, but this is intuition, not a mathematical proof

    • Jason Brownlee November 2, 2017 at 3:56 pm #

      It may help, try it and see – the cost is low.

  14. Qian December 6, 2017 at 9:12 am #

    Hi Jason,

    From the code of “calculate out of sample error” as below, is the best fitted model selected by lowest error of testing set? what can I do if I want to use the MSE and also R-square of training set to find the best fitted model?

    (# calculate out of sample error
    error = mean_squared_error(test, predictions)
    return error)

    Thank you in advance.

  15. Pratik January 4, 2018 at 10:22 am #

    Hi Jason,

    I am running hyperparameter search. How to deal situations with like:

    ValueError: The computed initial AR coefficients are not stationary
    You should induce stationarity, choose a different model order, or you can
    pass your own start_params.

    I have multiple datasets (in like 500s) and it won’t be possible for me analyze them individually. Any suggestions?

    Thank you in advance

  16. Siddharth Das January 5, 2018 at 8:14 pm #

    What is the best option to choose the ranges of pdq and then pass it on the functions to predict the best pdq results based on the MSE, how to determine the range of pdq we need to pass?

    Because if we pass blindly range(0,9), the whole model will take lot of time to find out the best result..

  17. Charles February 20, 2018 at 4:49 pm #

    Thank you very much Jason for a great article again.

    This article(
    https://www.linkedin.com/pulse/comparison-between-classical-statistical-model-arima-deep-virmani) compared ARIMA with RNN and LSTM on time series forecasting. The results of LSTM is much better than ARIMA (RMSE: 1.02 vs. 4.74). Given that the selecting p, q, d of Arima is not easy, it seems like that using ARIMA on the forecast is wise.

    Charles

  18. Charles February 22, 2018 at 12:49 pm #

    Thank you Jason, please correct me if I am wrong.

    1). The most of time series issues are nonlinear issues. we have to simplify them to linear issues to apply ARIMA since ARIMA is linear model. This is why the prediction of ARIMA is always quite poor on a time series issue.

    2). RNN and LSTM are able to handle nonlinear issues. Of course, its result is better than ARIMA.

    3). So, people should avoid ANY linear models for the prediction of time series issues.

    I have learned a lot from your articles. Many thanks to you!

    lcy1031

    • Jason Brownlee February 23, 2018 at 11:52 am #

      Mostly. Indeed ARIMA are linear models that capture linear relationships.

      The more pre-processing you can do your data to make the relationship is the data simpler/better exposed, the easier modeling will be regardless of the methods used.

  19. Charles February 24, 2018 at 4:37 am #

    Thank you Jason for reply.

    My concern is that any pre-processing could over simply the issue, and make the prediction not reliable. So far, there is not an approach which can measure whether or not a pre-processing is valid, and could cut the important information out.

    By the way, the trend and seasonality are very good features of a curve which could help the prediction a lot. Using stationary processes to remove them out is wrong.

    Thank you again!

    Charles

    • Jason Brownlee February 24, 2018 at 9:24 am #

      The best measure is model skill with and without the treatment to the data.

  20. Ajay March 7, 2018 at 8:16 pm #

    I m getting following error when I run this line
    model = ARIMA(history, order=(4,1,0))
    Error
    TypeError: descriptor ‘__sub__’ requires a ‘datetime.datetime’ object but received a ‘int’

    • Jason Brownlee March 8, 2018 at 6:22 am #

      Sorry to hear that. Did you copy the code and data from the tutorial exactly?

  21. Germán March 20, 2018 at 2:37 am #

    Hi Jason,

    I was digging into the ‘history’ list update (based on the ‘walk forward validation’), and I see that if you print out the couples ‘predicted VS forecasted’ in the way your code is written, it is actually displaying ‘forecast of the last element added to history’ VS ‘current element test[t]’ just added to ‘history’ but not forcasted yet.

    I propose this change, what do you think:

    • Germán March 20, 2018 at 2:39 am #

      * couples ‘predicted VS expected’ I meant 🙂

    • Jason Brownlee March 20, 2018 at 6:29 am #

      In walk forward validation we want the real ob added to the history (used to fit the model), not the prediction.

      Otherwise we will be using a recursive multi-step prediction model (e.g. different from our goal) and skill will be much worse.

      • Germán March 21, 2018 at 3:21 am #

        Hi Jason, thanks for your answer but I think you misunderstood what I mean, and I actually did not change what you add to the ‘history’ list.

        The change I made (validated running my code with a time series example running fine) is to display the correct pairs ‘predicted VS expected’ (can you checkl my code below in my last entry?)

  22. Germán March 20, 2018 at 2:43 am #

    Sorry for the typos, below the proposed code:

  23. Ajay Verma March 23, 2018 at 2:08 am #

    Hi Jason, thanks for your wonderful session. I have ticket count from Jan2017 till Jan2018.
    I have applied AREMA model , got the predicted value of test data as below this is till Jan2018, how to forecast for Feb2018 and Mar2018 which is not in database
    predicted=2440.666667, expected=2593.000000
    predicted=2642.187500, expected=2289.000000
    predicted=2317.411765, expected=2495.000000
    predicted=2533.277778, expected=3062.000000
    predicted=3128.105263, expected=2719.000000
    predicted=2764.650000, expected=3159.000000
    predicted=3223.428571, expected=3510.000000
    predicted=3587.454545, expected=3155.000000
    predicted=3213.652174, expected=2628.000000

  24. Adam Dziedzic March 29, 2018 at 11:38 am #

    Hi Jason, thank you for the great content. You wrote: “it may be possible to update the internal data of the model with new observations rather than recreating it from scratch.” I inspected the code: statsmodels/tsa/arima_model.py carefully but have not found how to compute the residuals on the new observations (how to update the model for new observations) without revising the parameters (using the model.fit method). Do you know how to do it? It’s possible in R: https://stats.stackexchange.com/a/34191/149565

    • Jason Brownlee March 29, 2018 at 3:17 pm #

      Sorry, I don’t have an example.

      • Adam Dziedzic March 30, 2018 at 8:19 am #

        Jason, thank you for your response. Do you know if it is possible? If so, do you have any hints? For example, I tried to rebuild the model with the new observation added to the history (variable) and reuse as many parameters from the previous model as possible (I created a new class that inherits from ARIMA). The shape of endog (the previous observations) and exog (the error terms) did not agree (and they can be transformed internally in different ways), so I omitted the exog parameter from the previous model and retrained the new model (using fit) with a single iteration. It kind of works but probably there is a better way to do it.

        • Jason Brownlee March 31, 2018 at 6:29 am #

          It might be easier to write a custom ARIMA with the features you need. There’s not a lot to it and I do find the statsmodels implementations not entirely extensible (as I am sure you’re discovering).

          • Adam Dziedzic April 3, 2018 at 7:35 am #

            Okay, thank you for your answer. I appreciate it.

  25. Kate Williams June 8, 2018 at 10:33 pm #

    Hi Jason,
    for my grid search, all MLE values are returned as 0. Indicating to me that possibly MLE failed to converge. However trying a few p, d and q values by hand, some combinations converge perfectly, but returning an MSE of 0.0 nonetheless.
    Do you have any idea on what might cause this?

    • Jason Brownlee June 9, 2018 at 6:53 am #

      Or the model has perfect skill. Perhaps the problem us too simple?

  26. Saurabh Dhokare June 20, 2018 at 10:00 am #

    Hey Jason!!

    An extremely helpful post.

    I just wanted an estimate of runtime considering I have a 6th gen, i5 processor. It’s been 2 days and the program keeps running and is showing no results. I have read most of he comments on the post and tried to debug the code. I have confirmed that the code is running but it isn’t producing any results.

    Your help will be highly appreciated!!
    Thanks 🙂

    • Jason Brownlee June 21, 2018 at 6:03 am #

      Perhaps add some print() statements to help see progress?

  27. Ajay July 2, 2018 at 12:55 am #

    Hi I created some raw code on this here https://github.com/decisionstats/pythonfordatascience/blob/master/Quarterly%2BTime%2BSeries%2Bof%2Bthe%2BNumber%2Bof%2BAustralian%2BResidents.ipynb

    basically the search function was to minimize AIC

    how would this compare to gridsearch for MSE minimization

    warnings.filterwarnings(“ignore”) # specify to ignore warning messages
    c4=[]
    for param in pdq:
    for param_seasonal in seasonal_pdq:
    try:
    mod = sm.tsa.statespace.SARIMAX(y,
    order=param,
    seasonal_order=param_seasonal,
    enforce_stationarity=False,
    enforce_invertibility=False)

    results = mod.fit()

    print(‘ARIMA{}x{}12 – AIC:{}’.format(param, param_seasonal, results.aic))
    c4.append(‘ARIMA{}x{}12 – AIC:{}’.format(param, param_seasonal, results.aic))
    except:
    continue

  28. Galen November 20, 2018 at 12:59 am #

    This is great, but I’m curious:

    Here we are predicting, but what about evaluating the model in itself using the AIC, BIC, or other criteria, without doing prediction.

  29. Hikmet Yavuz November 20, 2018 at 11:54 pm #

    Hi Jason,

    I have a production question for you.

    I have micro datasets (10 rows for each dataset) and it is required to have an ARIMA model for each dataset. I have 21060 datasets and i need to create 21060 ARIMA models and i have to perform grid search for each ARIMA model.

    I genericized your ARIMA grid search code and ran successfully on my computer. However, i observed that creating 1000 ARIMA models (1000 grid searches) takes approximately 1 hour on my computer. Creating 21060 ARIMA models (21060 grid searches) will approximately take 21 hours, if i run the code on my computer.

    My team leader asked me that can we run your code on spark environment or can we run your code on multi thread mode in order to reduce the execution time ( 21 hours).

    What do you suggest me?

  30. jay November 22, 2018 at 7:45 am #

    Hi Jason

    Thanks for the article!
    I ran the same code with the same database. It ran without any error but no result is printed. It prints following:
    Best ARIMANone MSE=inf

    any idea?

      • jay November 26, 2018 at 9:39 am #

        Hi

        I tried everything with the same code and the same data, still I am getting the same result
        i.e Best ARIMANone MSE=inf

        • Jason Brownlee November 26, 2018 at 2:03 pm #

          That is odd, are you able to confirm that statsmodels is up to date?

    • Rima November 24, 2018 at 4:26 am #

      Hello, I had the same problem.
      I think I found the answer, we should reset the index of the test set in the function “Evaluate_arima_model” since we are applying a for loop with the range of 0 to its length.
      This line of code must be added in the function “Evaluate_arima_model” : test.reset_index(drop=True,inplace = True)

      The function will be then:

  31. jay November 26, 2018 at 8:33 am #

    did you run it in terminal or in anaconda?

Leave a Reply